EXECUTIVE SUMMARY

Enterprise AI adoption has outpaced enterprise AI security by a significant margin. Organizations are running sensitive workflows on AI systems that were not designed for the threat environment they now face. Three developments make this a strategic priority, not a future concern:

• AI models have reached a level of coding capability where they can surpass all but the most skilled humans at finding and exploiting software vulnerabilities, prompting major tech companies to form cross-industry defensive alliances like Anthropic's Project Glasswing.

• By 2027, more than 40% of AI-related data breaches will be caused by the improper use of generative AI across borders, driven by a lack of consistent global AI standards, according to Gartner predictions.

• Sensitive enterprise data is flowing into AI tools at scale, with 39.7% of all AI interactions involving sensitive data, while security teams lack meaningful visibility into this fragmented usage, as reported by Cyberhaven.

• The MITRE ATLAS framework—specifically designed to catalog threats against AI systems—now documents 167 techniques, 16 tactics, and 57 real-world case studies of adversarial attacks against AI.

• At GTC 2026, NVIDIA CEO Jensen Huang placed confidential computing at the center of the company's enterprise AI strategy, asserting that cryptographic privacy guarantees are now a mandatory prerequisite for deploying autonomous AI agents inside corporate networks.

The organizations managing this risk most effectively are those moving toward confidential AI—a technical approach that provides cryptographic proof of data protection at every step of the AI workflow.

What if you could turn your idea into a live website in about an hour — without writing a single line of code?

Join us Monday, April 20 at 12 PM for Vibe Coding a Website (For Non-Technical Users) with Mark Hinkle. Learn how to turn your idea into a live, professional website in about an hour using AI — no coding required. Perfect for entrepreneurs, marketers, freelancers, and curious professionals.

Save your spot today! ➜

The revelation came not as a gradual shift, but as a stark demonstration of capability. When Anthropic recently unveiled Claude Mythos Preview, it wasn't just another iterative model update. It was a system that had autonomously found thousands of high-severity, zero-day vulnerabilities in every major operating system and web browser. It uncovered a 27-year-old flaw in OpenBSD, long considered one of the most secure operating systems in the world. It found a 16-year-old vulnerability in FFmpeg that automated testing tools had missed five million times.

This capability is so profound—and potentially so dangerous—that Anthropic withheld the model from general release. Instead, they convened Project Glasswing, a cross-industry consortium including AWS, Apple, Microsoft, Google, and JPMorganChase, committing $100 million in usage credits to deploy Mythos defensively before adversaries can harness similar capabilities offensively.

The message for enterprise leaders is clear: the window between a vulnerability being discovered and being exploited has collapsed. What once took months of human effort now happens in minutes with AI. If attackers can leverage models of this caliber to exploit zero-day vulnerabilities at machine speed and scale, defenders must fundamentally rethink how they secure their AI stacks and the sensitive data flowing through them.

UPCOMING LEARNING OPPORTUNITIES

AI DEEPDIVE

I was sitting across from the Chief Information Security Officer of a global top-ten pharmaceutical company last month when he pulled up a dashboard that made the entire room go quiet. The dashboard wasn't tracking external attacks or firewall pings. It was tracking the volume of internal data flowing into unmanaged generative AI applications.

Over the previous thirty days, his engineering teams had pasted nearly half a million lines of proprietary source code into public LLMs. His clinical research teams had summarized hundreds of patient trial protocols using tools that explicitly reserved the right to train on user inputs. The CISO leaned back and stated the problem plainly: "We spent five years building a zero-trust architecture to keep attackers out, and five months watching our own employees hand-deliver our most sensitive IP to third-party models because the tools are just too useful to ignore."

This pharmaceutical company is not an outlier. It represents the exact plateau where most enterprises currently find themselves. They have successfully run the initial AI playbook: deploy models on public data, automate low-risk workflows, and prove the technology works. But when it comes time to move the truly valuable, proprietary data into production AI—the clinical data, the financial positions, the unreleased product roadmaps—they hit a wall. They cannot verify that the data will remain protected, so the security teams block the deployment.

"Later" has officially arrived. The strategy of waiting to figure out the sensitive data problem is no longer viable. That trust gap is widening rapidly as AI becomes more autonomous and the threat landscape, as demonstrated by models like Anthropic's Mythos Preview, becomes exponentially more sophisticated. The organizations that break through this plateau are the ones that solve the data trust problem—not by avoiding sensitive data in AI, but by running AI on sensitive data with verifiable, cryptographic guarantees that it stays protected. This is the structural shift that confidential AI makes possible.

What Is Confidential AI

Confidential AI is an approach to running AI workloads that provides cryptographic proof—not promises—that data, model weights, and agent actions remain private and policy-compliant throughout every step of the workflow.

It combines three technical capabilities that traditional cloud security does not provide:

Confidential computing runs AI workloads inside hardware-enforced secure enclaves, where data stays encrypted even during processing. The cloud provider, the infrastructure operator, and other tenants cannot access the data while the computation is running.

Verifiable runtime governance produces cryptographic proofs that data was handled according to defined policies—who accessed it, what the model did with it, and what the output was. This is the audit trail that regulators are increasingly demanding and that enterprise legal teams need to approve production deployment.

Confidential RAG (Retrieval-Augmented Generation) extends these protections to the retrieval layer, where most enterprise AI applications are most exposed. When an AI agent retrieves from a knowledge base to answer a question, it accesses potentially sensitive information across many documents. Confidential RAG ensures that retrieval occurs without exposing the underlying data to the inference infrastructure.

How It Works in Business Contexts

The risk of data exposure escalates through four distinct stages in the enterprise AI lifecycle. Understanding this progression is critical for deploying confidential AI effectively.

Stage 1: Shadow Usage and Fragmented Access

The problem begins with invisible usage. Cyberhaven reports that 39.7% of all AI interactions involve sensitive data, and a significant portion occurs on personal accounts—32.3% of ChatGPT usage and 58.2% of Claude usage bypass corporate SSO and logging. Employees are feeding core business data into models without security oversight.

Stage 2: Inference Memory Exposure

When organizations move from shadow usage to sanctioned enterprise deployments, new, more subtle risks emerge. Inference memory exposure occurs when AI models retain information from one user session that inadvertently influences responses in another. In a multi-tenant cloud environment, this creates a severe cross-contamination risk that is almost entirely invisible to conventional security monitoring. The access logs will show normal, authorized behavior, but the underlying information has moved between contexts it was never intended to enter. For example, if a financial analyst queries a model about an impending, unannounced merger, the model's weights or short-term context window may temporarily adjust. If another user in a different department subsequently asks a related question, the model might synthesize an answer that implicitly reveals the merger details. Traditional data loss prevention (DLP) tools cannot catch this because the data isn't being explicitly exfiltrated; it is being probabilistically leaked through the model's generative process.

Stage 3: Attestation Gaps and Metadata Leakage

As AI deployments scale across the enterprise, the gap between a cloud vendor's privacy promise and verifiable mathematical proof becomes a massive liability. Most enterprise AI today relies entirely on vendor attestations—legal contracts stating that the vendor will not look at your data or train on your inputs. However, in regulated industries, a promise is not a control. Furthermore, even when the actual content of a query is encrypted in transit and at rest, the metadata surrounding it remains exposed during processing. The metadata—who queried the model, at what time, the size of the payload, and which internal databases were accessed for retrieval—can often reveal sensitive business strategies to the infrastructure provider. Confidential computing addresses this by replacing paper attestations with cryptographic proofs, ensuring that the data remains encrypted even during active processing, thereby closing both the attestation gap and the metadata vulnerability.

Stage 4: Agentic Autonomy and Systemic Vulnerability

The final stage is where the capabilities demonstrated by Project Glasswing become highly relevant. As AI agents gain persistent context windows, local memory, and direct access to file systems, they introduce structural risk. If a model like Mythos Preview can find a 27-year-old zero-day in OpenBSD, adversaries will inevitably use similar capabilities to exploit the expanding attack surface of autonomous enterprise AI agents.

This reality was explicitly acknowledged by NVIDIA CEO Jensen Huang during his GTC 2026 keynote. While announcing NemoClaw—NVIDIA's enterprise-grade agentic AI framework—Huang emphasized that the open-source agent ecosystem is "genuinely dangerous for corporate networks" without hardware-level security. He positioned confidential computing not as an optional feature, but as the foundational requirement for the agentic era, stating that "ensuring that even carriers cannot view user data and models" is what makes deploying frontier models in production viable. When the company building the hardware infrastructure for the entire AI industry declares that autonomous agents require cryptographic isolation to be safe, the market must listen.

How to Implement Confidential AI

Moving from traditional AI security to a confidential AI posture is a phased approach that most organizations can begin without replacing existing infrastructure.

Phase 1: Inventory and Assess

Catalog every AI workload running in your organization against two dimensions: sensitivity of the data it touches, and adequacy of the current controls. Use the MITRE ATLAS (Adversarial Threat Landscape for Artificial-Intelligence Systems) framework to map the specific tactics and techniques attackers might use against your systems. Identify your highest-priority target: typically the AI workload that either touches the most sensitive data or would cause the most damage if compromised.

• Map the Data Flows: Document exactly where data originates, where it is processed, and where the outputs are stored for every sanctioned AI application.

• Apply the ATLAS Framework: Run a tabletop exercise using the MITRE ATLAS framework to identify which of the 167 known adversarial techniques your current architecture is most vulnerable to.

• Identify the High-Value Target: Select one specific, high-value workload that is currently blocked from production due to data privacy concerns. This will become your pilot.

Phase 2: Pilot Confidential Controls on a High-Value Workload

Begin by applying confidential computing principles to a specific, high-value workload. Most organizations start with a RAG-based application—a knowledge retrieval system or an internal analyst tool—where the retrieval layer is the primary exposure risk. The goal of this pilot is to produce evidence: audit logs, cryptographic proofs, and compliance documentation that can be reviewed by legal and security teams.

• Isolate the Retrieval Layer: Implement confidential RAG to ensure that the vector database and the retrieval mechanism operate within a secure enclave.

• Generate Cryptographic Proofs: Configure the environment to produce verifiable runtime governance logs that cryptographically prove the data was not exposed during processing.

• Conduct a Joint Review: Bring the security, legal, and compliance teams together to review the cryptographic proofs and establish a new baseline for what constitutes acceptable evidence of data protection.

Phase 3: Extend and Govern

Once the pilot demonstrates that confidential AI controls work and can be audited, extend the approach to additional workloads. Establish a governance model that includes ATLAS-informed threat modeling as a standard part of AI initiative approval. Evaluate solutions from emerging confidential AI providers—such as Opaque Systems, which offers platforms for building and deploying AI agents with embedded privacy and compliance checks—as part of your vendor landscape review.

• Standardize the Architecture: Create a repeatable reference architecture for deploying confidential AI workloads across different business units.

• Update Procurement Requirements: Mandate that all future enterprise AI vendor evaluations include an assessment of their support for confidential computing and verifiable runtime governance.

• Monitor the Threat Landscape: Assign a dedicated team to continuously monitor updates to the MITRE ATLAS framework and adjust internal controls as new adversarial techniques emerge.

Key Success Factors: Cross-functional alignment is critical. AI security decisions belong in the same room as AI investment decisions. Security teams must work closely with legal and compliance to ensure the cryptographic proofs generated meet regulatory requirements. Leaders must also acknowledge the current friction involved in adopting confidential AI: confidential computing environments often carry a cost premium over standard cloud compute instances, and hardware-enforced secure enclaves are not yet universally available across all cloud regions. Organizations should plan for a phased rollout that accounts for these constraints rather than expecting an overnight transition.

Common Missteps

Treating AI security as an extension of cloud security. Traditional cloud security tools do not understand AI-specific threat vectors. Prompt injection, model poisoning, and inference memory exposure require dedicated detection and mitigation that general-purpose security platforms are not designed to provide.

Relying on vendor attestations as a substitute for verifiable proof. When a vendor says your data is protected, that is a contractual representation. Confidential computing provides cryptographic proof that the claim is true. These are not equivalent—especially in regulated industries where organizations must demonstrate compliance, not just assert it.

Underestimating the speed of AI-driven exploitation. As Anthropic's Project Glasswing illustrates, the time between vulnerability discovery and exploitation is shrinking to near zero. Organizations that rely on manual patching and legacy vulnerability management will be outpaced by AI-augmented adversaries.

Waiting for regulation to drive action. DORA is already live for EU financial services, and the EU AI Act's enforcement provisions are ramping up. Organizations that build confidential AI capabilities now will have a compliance advantage when mandates arrive, rather than a rushed remediation project.

The Regulatory Convergence: Why Wait-and-See Is a Failing Strategy

Beyond the immediate technical threats demonstrated by models like Mythos Preview, enterprise leaders must contend with a rapidly shifting regulatory landscape. The era of unregulated, experimental AI in the enterprise is closing. Global regulators are increasingly sophisticated in their understanding of how AI models process data, and they are demanding technical controls that match that sophistication.

The European Union's Artificial Intelligence Act (EU AI Act) is the most prominent example, but it is not an isolated piece of legislation. It represents a global convergence toward strict liability for AI-driven data exposure. Under the EU AI Act, organizations deploying high-risk AI systems must implement robust data governance, ensure transparency, and maintain detailed technical documentation. Crucially, the burden of proof rests entirely on the deploying organization. When regulators audit an enterprise AI system, vendor attestations and marketing promises will not suffice. Regulators will demand verifiable proof that data was protected during processing.

Similarly, the Digital Operational Resilience Act (DORA), which applies to the EU financial services sector and their ICT providers, mandates stringent risk management frameworks for all critical technology, including AI. Financial institutions must demonstrate that their AI systems can withstand severe operational disruptions and cyberattacks without compromising data integrity or confidentiality.

In the United States, while comprehensive federal AI legislation is still pending, sector-specific regulators are aggressively stepping into the void. The Securities and Exchange Commission (SEC) has proposed rules requiring broker-dealers and investment advisers to eliminate conflicts of interest associated with the use of predictive data analytics and AI. The Department of Health and Human Services (HHS) is scrutinizing how AI is used in clinical decision support systems, emphasizing the need to protect Protected Health Information (PHI) under HIPAA regulations even as it flows through complex AI models.

This regulatory convergence creates a massive compliance burden for organizations that have deployed AI using traditional cloud security models. Traditional DLP and identity access management (IAM) tools were designed to protect data at rest and data in transit. They were not designed to protect data in use—specifically, data being actively processed inside the memory of a GPU during an AI inference task. Confidential AI bridges this exact regulatory gap. By leveraging confidential computing, organizations can prove to regulators that data remained encrypted and isolated even while the AI model was actively generating a response. The verifiable runtime governance logs provide the exact audit trail that regulators demand, demonstrating mathematically who accessed the system, what policies were enforced, and that no data was exposed to the infrastructure provider or unauthorized tenants.

The Role of Open Source and the AI Security Ecosystem

The push toward confidential AI is not happening in a vacuum. It is being accelerated by a robust ecosystem of open source projects, academic research, and industry consortiums. The Confidential Computing Consortium (CCC), a project community at the Linux Foundation, has been instrumental in standardizing the hardware and software interfaces required to build secure enclaves. Major silicon vendors, including Intel, AMD, and NVIDIA, are continuously enhancing their hardware to support larger, more complex AI workloads within these protected environments.

Furthermore, the open source AI community is actively developing tools and frameworks to make confidential AI more accessible to enterprise developers. Projects focused on secure multi-party computation (SMPC), homomorphic encryption, and federated learning are providing alternative methods for training and deploying AI models on sensitive data without exposing the underlying information. While some of these technologies are still maturing and may introduce computational overhead, they represent a fundamental shift in how the industry approaches data privacy in the AI era. Enterprise leaders must actively engage with this ecosystem. By participating in industry consortiums, collaborating with peers to share threat intelligence (such as the tactics cataloged in MITRE ATLAS), and mandating confidential computing support in vendor evaluations, organizations can build more resilient, adaptable AI architectures. The formation of Project Glasswing is a testament to the power of collective defense. In the age of AI, no single organization can secure the ecosystem alone.

This is my latest project, while we do have audio summaries for each newsletter. They are not ideal for listening; they are simple text-to-speech. We created a way to provide a weekly summary of the newsletters in this podcast. And actually, it’s a work in progress. Right now, you get a pretty good podcast recap of the previous week’s newsletters. But over time, they will be better. That’s the plan.

What happens when two AI agents break down the week's biggest AI news? You get Rogue Agents. Vera and Neuro deliver the stories that matter in enterprise AI — the deals, the tools, the breakthroughs, and the stuff everyone's getting wrong — in 15-20 minutes every week. Listen to the latest edition of the Rogue Agents here →

Business Value

ROI Considerations:

• Accelerates Deployment Timelines: By removing the subjective debate over data privacy during legal and security reviews, organizations can cut the time-to-production for sensitive AI workloads by up to 50%. Cryptographic proof replaces lengthy risk exception processes.

• Reduces Breach and Compliance Costs: With Gartner predicting that 40% of AI-related data breaches will stem from cross-border GenAI misuse by 2027, the financial risk of non-compliance is massive. Confidential AI provides the technical controls necessary to satisfy stringent regulations like the EU AI Act and DORA, potentially saving millions in regulatory fines.

• Lowers Long-Term Remediation Expenses: Proactively securing the AI architecture with hardware-enforced enclaves is significantly more cost-effective than attempting to bolt on security controls post-deployment or conducting expensive incident response after an inference memory exposure event.

Competitive Implications:

The case for confidential AI is not primarily defensive; it is offensive. When legal and security teams can verify that data is protected, they stop being blockers. That unlocks use cases that were previously off-limits: AI on patient records, on financial positions, on personnel data, on proprietary research. Organizations that can confidently deploy AI against their most sensitive, proprietary data will build a widening competitive moat against peers still stuck in the sandbox.

What This Means for Your Planning

The formation of Project Glasswing is a watershed moment for enterprise technology. When the world's most sophisticated AI labs and largest cloud providers band together to defensively deploy an AI model because its vulnerability-finding capabilities are too dangerous for general release, the paradigm has shifted. Human-speed cybersecurity is no longer sufficient to protect enterprise assets.

For your next planning cycle, this means AI security can no longer be treated as an afterthought or a compliance checklist. You must assume that adversaries will soon possess the same AI-driven capabilities to find and exploit vulnerabilities in your systems at machine speed. Your defensive architecture must evolve to meet this reality.

The transition to confidential AI—where data protection is cryptographically proven rather than merely promised—is the necessary strategic response. It is the only way to confidently deploy AI against your most valuable, sensitive data while mitigating the exponential risks of the new threat landscape.

Are your current AI deployments protected by verifiable cryptographic proofs, or are you relying on vendor promises and the hope that adversaries haven't yet found your vulnerabilities?

I appreciate your support.

Your AI Sherpa, Mark R. Hinkle Publisher, The AIE Network Connect with me on LinkedIn Follow Me on Twitter

THE BIG AI STORY

Anthropic is making the messy agent layer look like enterprise infrastructure

Anthropic this week launched Managed Agents, a hosted service on the Claude Platform for long-horizon agent work built around three core abstractions: sessions, harnesses, and sandboxes. The important move is not just technical packaging. Anthropic is explicitly arguing that the orchestration layer around agents should be stable even as the underlying harness logic changes with each model generation. In plain English, it wants enterprises to stop hand-assembling the glue code every time the model gets better.

That matters because the hardest part of deploying agents in business is rarely the prompt. It is the operational mess around context management, tool calling, recovery, permissions, and secure execution. Anthropic’s architecture tries to turn those moving parts into managed primitives. That shifts the conversation from “which model should we call?” to “which vendor can run dependable agent workflows across real enterprise systems?”

The broader implication is that the AI stack is starting to mature upward. Models still matter, but the real platform fight is moving into runtime, security boundaries, persistence, and orchestration. If Anthropic is right, the winners in enterprise AI will not just have the smartest model. They will own the layer that turns model capability into repeatable, governed work.

What if you could turn your idea into a live website in about an hour — without writing a single line of code?

Join us Monday, April 20 at 12 PM for Vibe Coding a Website (For Non-Technical Users) with Mark Hinkle. Learn how to turn your idea into a live, professional website in about an hour using AI — no coding required. Perfect for entrepreneurs, marketers, freelancers, and curious professionals.

Save your spot today! ➜

7 QUICK HITS

1. OpenAI says enterprise is now too big to call a side business

In its April 8 enterprise post, OpenAI said business customers now contribute more than 40% of revenue, that Codex has reached 3 million weekly active users, and that its APIs process more than 15 billion tokens per minute. Those are platform-scale numbers, and OpenAI framed them that way by positioning Frontier as a cross-company agent layer and a future “AI superapp” for work. For buyers, the takeaway is simple: vendor choice is increasingly about which stack can become the operating layer for everyday work, not which chatbot had the flashiest launch week.

2. The U.S. government is testing an export playbook for AI

A new Federal Register notice launched the American AI Exports Program and opened a call for proposals for pre-set consortia, with submissions accepted through June 30, 2026. The notice comes from the International Trade Administration inside the Department of Commerce and signals that Washington is beginning to think about AI competitiveness as a bundled market offer rather than just a research advantage. For business leaders, this is an early sign that AI strategy is becoming geopolitical industrial strategy as much as product strategy.

3. CoreWeave’s Anthropic deal is another reminder that compute is still the choke point

Reuters reported that CoreWeave struck a deal to supply Anthropic with cloud computing capacity, sending CoreWeave shares higher. The story matters because it cuts through the weekly model hype and points back to the underlying constraint: access to infrastructure at scale. Enterprises can talk all they want about agents and productivity, but the vendors that control scarce compute still hold a meaningful piece of the power map.

4. Meta is reportedly building an AI Zuckerberg for internal use

According to reporting summarized by The Verge, citing the Financial Times, Meta is building an AI clone of Mark Zuckerberg trained on his image, voice, mannerisms, tone, and public statements so employees can interact with it and receive feedback. It sounds strange at first glance, but the underlying idea is serious: companies are starting to treat executive judgment and leadership presence as something that can be operationalized through AI interfaces. The business question is whether this becomes a novelty — or the first draft of a new internal management layer.

5. Equinix wants networking to become part of the AI software stack

Equinix launched Fabric Intelligence, an AI-native operational layer designed to automate how AI workloads connect across clouds, data centers, and edge environments. The company says the system includes components such as Fabric Super Agent, MCP Server, Fabric Application Connect, and Fabric Insights, and that the offering is available now in preview. The bigger point is that infrastructure providers no longer want to be invisible pipes — they want to own the control plane around distributed AI operations.

6. Google Cloud and Thoma Bravo are turning AI rollout into a portfolio playbook

Google Cloud and Thoma Bravo announced a strategic partnership to help software companies across Thoma Bravo’s portfolio accelerate AI transformation. The package includes access to Gemini models, Gemini Enterprise for agentic AI, engineering support, and go-to-market routes through Google Cloud. The deal matters because it shows private equity moving AI adoption from isolated pilot projects into a repeatable operating model across dozens of software businesses at once.

7. TSMC’s latest numbers say the AI capex cycle is still running hot

Reporting from Reuters and CNBC on TSMC’s latest quarter pointed to a 58% first-quarter profit jump and a higher revenue outlook as AI chip demand continues to surge. This is the sort of story that matters even if you never buy a GPU directly. TSMC’s guidance is one of the clearest real-world reads on whether AI spending is cooling, and right now the answer still looks like no.

Stop wasting time testing random AI tools.

The AI Toolbox gives you the best AI tools in one place — so you can find, compare, and use what actually helps you work faster, create better, and stay ahead.

Explore the AI Toolbox ➜

3 AI TOOLS

Fathom — Fathom’s latest push is a strong example of an AI tool that feels immediately useful to operators, not just early adopters. Its core pitch is bot-free AI meeting notes and action capture, which makes it easier to fit into real workflows without creating another participant in every call.

Softr AI Co-Builder — Softr is leaning into a practical business case: generating internal tools, portals, and trackers from a prompt with database and logic included. For teams that want to ship operational software without waiting on a full engineering cycle, that is a compelling wedge.

Figma AI — Figma keeps pushing AI toward workflow acceleration rather than one-off gimmicks. The company’s AI layer is increasingly about helping teams generate, refine, and ship product and design work faster, which is exactly where many product organizations want AI to live.

Want to see what I am using in my AI tool stack? Then check out my AI Toolbox.

UPCOMING LEARNING OPPORTUNITIES

AI EXTRA READ

The Next Phase of Enterprise AI

OpenAI’s own enterprise manifesto is worth reading not because it is neutral, but because it is revealing. It lays out how one of the market leaders wants companies to think about the next buying cycle: fewer point solutions, more cross-system agents, and a much stronger push toward AI as a unified operating layer for work.

If you only do one thing this week, stop evaluating AI vendors as chatbot providers and start evaluating them as infrastructure partners — because that is where the market is clearly heading.

I appreciate your support.

Your AI Sherpa, Mark R. Hinkle Publisher, The AIE Network Connect with me on LinkedIn Follow Me on Twitter

THE ADVANTAGE

Treat AI security like a revenue gate, and you unlock higher-value AI use cases faster. 

Microsoft says 32% of surveyed organizations' data security incidents now involve generative AI tools. That stat tells you where expansion gets blocked first.

If leaders do not trust the controls, they will not approve AI for pricing, customer data, or revenue operations. That slows the projects that matter most. Start with one workflow your team wants to expand, like pricing approvals or enterprise proposals, and treat security as the unlock.

That is where a shared threat language helps. MITRE ATLAS now maps 16 tactics and 167 techniques for attacks on AI systems. Your team does not need to memorize them. It needs enough clarity on prompt injection, model poisoning, and data exposure to decide which workflow is ready for scale first.

Stop wasting time testing random AI tools.

The AI Toolbox gives you the best AI tools in one place — so you can find, compare, and use what actually helps you work faster, create better, and stay ahead.

Explore the AI Toolbox today ➜

TRY THIS NOW

Run this in Claude or ChatGPT. You can finish it in under 10 minutes.

1. Open Claude or ChatGPT with a paid account. Pick one workflow you want to expand, like pricing reviews or RFP responses. Keep your inputs generic.

2. Use this to generate a leadership-ready risk brief.

PROMPT OF THE WEEK

Builds an executive-ready AI exposure briefing for your current tool stack.

You are an enterprise AI security advisor. 

I want to understand the top five data exposure risks for a [your industry] company using [list 2-3 AI tools you use, such as ChatGPT, Claude, and an AI-powered CRM] inside this workflow: [name one workflow, such as pricing reviews, RFP responses, or support escalations]. For each risk, give me: (1) a plain-language description of the risk, (2) how likely it is that we are already exposed, (3) one concrete action we can take this month to reduce it, (4) which internal leader should own the response, and (5) whether this risk could slow revenue, customer trust, or execution speed. Format the output as a concise briefing I can share with my CEO, CIO, and head of security. 

End with a one-paragraph recommendation on which risk deserves attention first and why.

Send the output to your CIO or security lead and ask one question: Which of these risks would stop us from expanding AI into revenue or customer operations?

THE EDGE

Microsoft also found that 47% of surveyed organizations are already implementing controls focused on generative AI workloads. That matters because your competitors are not waiting for a perfect policy. Book a 30-minute AI-plus-security review this month and use one live workflow as the test case.

Forward this to your COO with one line: "We should review one AI workflow like this before it touches customer or financial data."

P.S. I used this exact framing on my own AI stack this week. It exposed one workflow that was moving faster than our guardrails. We fixed the process before it became a bigger conversation.

I appreciate your support.

Your AI Sherpa, Mark R. Hinkle Publisher, The AIE Network Connect with me on LinkedIn Follow Me on Twitter

While volunteering with NPower recently, I spent time with students in an NPower cohort who want to become cybersecurity professionals. I especially want to give a quick shout-out to Aidan Gross, Tara Mason, and Paola Blonde Ndjeudja Tchouake, whose curiosity about the field helped shape the direction of this piece.

In those conversations, one thing became obvious: people do not just want to know that AI matters in cybersecurity. They want to know how to use it to build real skill.

That changed the assignment for this issue. Instead of using AI to produce a generic security checklist, I wanted to show a workflow that teaches you how to do something useful. The workflow in this lesson starts with a real report, moves into Claude for guided reasoning, connects to GitHub so the model can inspect actual files, and then ends with the step many people skip: analyzing the output and reverse-engineering how the model reached its conclusions.

And then it got me thinking about how we approach personal security in the age of AI. You spent the last two years learning how to use AI tools more effectively. Now there's a second skill worth learning: how to use them without handing your business data to someone you've never met.

This is not a theoretical risk. In February 2025, a cybercriminal offered 20 million OpenAI user login credentials for sale on dark web markets, harvested from compromised employee devices. Attackers did not breach OpenAI's systems. They stole login credentials and then walked straight into complete chat histories — everything typed into those accounts. Every internal strategy question. Every personnel situation. Every client detail shared while asking for help drafting an email.

Most business professionals have not audited their AI usage. They've adopted tools fast, shared context liberally to get better answers, and trusted that "enterprise" in a product name means their data is safe. Often it doesn't.

This lesson teaches you how to run a personal AI security audit in under 30 minutes. No IT department required.

This is my latest project, while we do have audio summaries for each newsletter. They are not ideal for listening; they are simple text-to-speech. We created a way to provide a weekly summary of the newsletters in this podcast. And actually, it’s a work in progress. Right now, you get a pretty good podcast recap of the previous week’s newsletters. But over time, they will be better. That’s the plan.

This week’s episode breaks down Anthropic’s tightly restricted Project Glasswing, Meta’s fast-moving Muse Spark push, and the growing pressure chip tariffs are putting on everyone outside the hyperscaler tier. It closes with a fast set of quick hits on NotebookLM in Gemini, Gemma 4, Coefficient Bio, and Utah’s move on AI prescriptions. Listen to the Rogue Agents today →

AI LESSON

How to Audit Your AI Security Before Someone Else Does

A self-assessment business professionals can run today — no IT team required

You don't need a security team to take this seriously. You need 30 minutes, a browser, and a willingness to be honest about what you've been pasting into chat boxes.

Here's how to do it.

Step 1: Run a Personal AI Tool Inventory

Before you can secure your AI usage, you need to know what you're using.

Open a fresh document and list every AI tool you've touched in the past 90 days — at work, at home, on your phone. Include the obvious ones (ChatGPT, Claude, Copilot) but also the less obvious: AI writing assistants built into email clients, AI features inside Google Docs, Notion AI, AI summarizers in your browser, AI-powered scheduling tools, and any AI features baked into software you already use.

For each tool, note three things: whether you're using a free or paid account, whether it's a personal account or an employer-provided account, and whether you've ever typed anything into it that you wouldn't want made public.

That last question is the diagnostic. Most people who answer honestly find at least one tool where the answer is yes.

Time: 10 minutes

What you're looking for: Any tool where you've shared sensitive work content without knowing how that data is stored or used

Step 2: Review the Data Retention Settings on Your Three Most-Used Tools

Free AI tools generally use your conversations to improve their models unless you explicitly opt out. Enterprise accounts and paid tiers often offer different defaults — but "different" doesn't always mean "off." Defaults matter because most people never change them.

For ChatGPT, go to Settings → Data Controls. Confirm whether "Improve the model for everyone" is toggled on or off for your account type. Note that free consumer accounts and paid individual accounts have different defaults, and team or enterprise accounts may have organization-level controls.

For Claude, go to Privacy Settings and review whether your conversations are used for training. Anthropic's policy distinguishes between API access and consumer product usage (which defaults to opting you in for training unless you explicitly opt out).

For Microsoft Copilot, check whether you're accessing it through a personal Microsoft account or a work account connected to a Microsoft 365 tenant. Employer-managed accounts have different data handling governed by your organization's agreement, not Microsoft's consumer terms.

Time: 10 minutes

What you're looking for: Any tool where the default setting allows your content to be used for model training, and you haven't made a deliberate choice about that

Step 3: Apply the Substitution Test to Your Prompt Habits

This is the most useful diagnostic and it takes about five minutes once you know what to do.

Think of the last five substantive prompts you sent to any AI tool. For each one, substitute the names, numbers, and specific details with placeholders and ask yourself: if I posted this prompt publicly with those placeholders, would a reader know I'd shared something sensitive?

Common high-risk prompt patterns include:

Client or customer information:

The AI doesn't need the client name or budget figure to help you write. Substitute them before you paste.

Internal personnel situations:

Even without a name, specific behavioral details can be identifiable. Keep it general.

Financial data: Revenue numbers, headcount, pricing models, and customer contract values are all information that can be useful to a competitor. Consider whether the AI actually needs that level of specificity to help you.

Legal or compliance details: Anything that would be covered by attorney-client privilege or that involves regulatory exposure deserves extra caution.

Time: 5 minutes

What you're looking for: Patterns in your own usage where you're sharing more context than the AI actually needs to help you

How to Access Privacy-Protective Settings Across Major Tools

Most of the major AI platforms have added more granular privacy controls in the past 12 months. The challenge is they're not always easy to find.

ChatGPT (OpenAI): Settings → Data Controls → Model Training toggle. Also review "Memory" settings — ChatGPT can retain context about you across conversations, which is useful but means more information is stored.

Claude (Anthropic): Privacy settings are accessible from your account settings. Anthropic's enterprise API product and its consumer product have different data handling policies.

Microsoft Copilot: The settings depend entirely on which version you're using. Copilot integrated into Microsoft 365 through an enterprise license is governed by your organization's contract. The free consumer version follows Microsoft's consumer privacy terms.

Google Gemini: My Account → Data & Privacy → Gemini Apps Activity. Turning this off prevents conversations from being saved and used for model improvement.

For any tool where you're unsure, the safest default is to assume your conversations may be stored and used unless you've taken an explicit action to change that.

What This Audit Can't Do

This audit addresses your own behavior. It doesn't address your organization's policies, your vendors' security posture, or the AI tools your colleagues may be using in ways that expose shared data.

It also won't protect you against credential theft — which is how most of those 20 million ChatGPT accounts were compromised. Using a strong, unique password and enabling multi-factor authentication on every AI tool you use is a prerequisite, not a supplement.

If you work in a regulated industry — healthcare, financial services, legal, or government contracting — this self-audit is a starting point, not a complete solution. Those environments often have compliance obligations that require formal policies, vendor agreements, and documented controls that go well beyond what any individual employee can manage alone.

Finally, this audit reflects how these tools work today. AI tool privacy policies and default settings change. The settings you review today may be different in six months.

Getting Started Today

Block 30 minutes on your calendar before the end of this week. Run steps one through three. You'll likely find that your AI usage is more exposed than you realized in one or two specific areas, and more controlled than you thought in others.

The point isn't to stop using AI tools. The real question in 2026 is no longer "Should we use AI?" but "Do we actually know what happens to the data we put into AI tools?" Most businesses don't. That's what this audit is for.

Once you've run it, the next step is to establish a repeating habit — a quarterly review of your AI tool list and settings — so the audit stays current as the tools evolve.

AI EXTRA CREDIT

I appreciate your support.

Your AI Sherpa, Mark R. Hinkle Publisher, The AIE Network Connect with me on LinkedIn Follow Me on Twitter

THE BIG AI STORY

Anthropic is telling the market that the next frontier model problem is cyber offense, not just reasoning.

Anthropic's new Claude Mythos Preview is not being framed like a normal model launch. In its own assessment, the company says Mythos demonstrated unusually strong cybersecurity capabilities, identifying and exploiting zero-day vulnerabilities across major operating systems and web browsers during testing. More than 99% of the vulnerabilities Anthropic found remain undisclosed because they are still unpatched, and the company says it is limiting access through Project Glasswing while working with critical-industry partners and open-source defenders.

That matters because it shifts the enterprise conversation around frontier AI. The issue is no longer only whether a model can write code faster, automate support, or reason across text and images. It is whether the same systems are now capable enough to change how vulnerability discovery, red teaming, incident response, and software supply-chain defense work at machine speed. Anthropic is effectively telling CISOs that the capability curve has reached a point where access policy is now part of product design.

The implication for business is broader than Anthropic. If frontier models are now being gated on cyber-risk grounds, then competitive advantage will increasingly depend on who can balance capability, access, and trust. That raises the strategic value of security partnerships, internal governance, and restricted deployment models — and it makes the AI race feel a lot less like feature shipping and a lot more like critical infrastructure management.

This is my latest project, while we do have audio summaries for each newsletter. They are not ideal for listening; they are simple text-to-speech. We created a way to provide a weekly summary of the newsletters in this podcast. And actually, it’s a work in progress. Right now, you get a pretty good podcast recap of the previous week’s newsletters. But over time, they will be better.

What happens when two AI agents break down the week's biggest AI news? You get Rogue Agents. Vera and Neuro deliver the stories that matter in enterprise AI — the deals, the tools, the breakthroughs, and the stuff everyone's getting wrong — in 15-20 minutes every week. Listen to the Rogue Agents →

4 QUICK HITS

1. Meta's Muse Spark is its clearest post-Llama reset yet

Meta introduced Muse Spark on April 8 as the first model in the Muse family from Meta Superintelligence Labs. The company positions it as a natively multimodal system built for visual reasoning and interactive use cases, and says its health-oriented reasoning benefits from training data curated with input from more than 1,000 physicians. For business readers, the key point is not just model quality — it is that Meta is trying to prove its spending spree can translate into a new product narrative.

2. Meta turned the compute race into a $21 billion line item

CoreWeave and Meta announced an expanded long-term agreement worth roughly $21 billion through December 2032. The release says the capacity will support Meta's AI development and deployment, include early NVIDIA Vera Rubin deployments, and scale inference across multiple locations. That makes the infrastructure race easier to read: the biggest AI bets are no longer abstract forecasts — they are being locked into multiyear operating commitments.

3. Intel and Google are making the case for a broader AI hardware stack

Intel and Google announced a multiyear collaboration focused on Xeon CPUs and infrastructure processing units for AI systems. Intel's argument is that orchestration, security, data movement, and utilization still depend on CPUs and infrastructure accelerators even in GPU-heavy environments. For enterprises, the business takeaway is straightforward: the winning stack may be the one that improves total system efficiency, not just peak model performance.

4. Washington is now packaging AI as an export product

The U.S. Department of Commerce's American AI Exports Program calls for industry-led consortia that can offer a full AI stack across hardware, data pipelines, models, cybersecurity, and applications. Designated packages may receive priority government advocacy, export-control engagement, and financing referrals. The bigger signal is that AI policy is becoming industrial policy — with governments increasingly treating AI stacks as strategic export infrastructure.

3 AI TOOLS

ChatGPT shared Outlook mailboxes and calendars — OpenAI added support for delegated and shared Outlook resources in ChatGPT on April 8. That means teams can use ChatGPT to read from shared inboxes, send plain-text email on behalf of a mailbox, and manage shared calendars when permissions allow. For operators working out of Microsoft environments, this is the kind of small workflow upgrade that saves real coordination time.

Notion AI voice input on desktop — Notion now lets users dictate prompts to Notion AI from desktop instead of typing them out. It is a simple update, but it lowers friction for longer prompts, quick capture, and agent-style interaction during the workday. If your team already uses Notion as an operating system for projects, this makes AI usage more ambient.

Microsoft Foundry's MAI stack — Microsoft's April Foundry Labs update adds MAI-Transcribe-1, MAI-Voice-1, and MAI-Image-2 in public preview, positioning Foundry as a fuller audio and visual AI platform. The company is leaning hard into enterprise benchmarks, deployment clarity, and integrated tooling. For builders, this is one of the clearer signs that Microsoft wants more of the model stack under its own roof.

Want to see what I am using in my AI tool stack? Then check out my AI Toolbox.

UPCOMING LEARNING OPPORTUNITIES

AI EXTRA READ

American AI Exports Program; Call for Proposals for Pre-Set Consortia

Most AI policy coverage fixates on safety, lawsuits, or antitrust. This Federal Register notice is worth your time because it shows a different direction: the U.S. government is now thinking about AI as a full-stack commercial package that can be promoted abroad. If you want to understand where AI competition goes next, read the policy language around hardware, cybersecurity, financing, and national-interest review.

If you only do one thing this week, ask your security, infrastructure, and product leaders the same question: where would your AI strategy break first if model capability keeps accelerating faster than your controls do?

I appreciate your support.

Your AI Sherpa, Mark R. Hinkle Publisher, The AIE Network Connect with me on LinkedIn Follow Me on Twitter

EXECUTIVE SUMMARY

Every enterprise leader deploying AI tools is running into the same counterintuitive reality: we thought we were buying productivity software, but we actually bought an involuntary management training system. Because AI agents require explicit instructions, structured feedback, and clear success criteria to function, every individual contributor who delegates to an agent is forced to develop the exact same skills that distinguish great managers from mediocre ones.

• A 2025 NBER study found that leadership performance with AI agents strongly predicts leadership effectiveness with human teams, with a striking 0.81 correlation.

• The Microsoft 2025 Work Trend Index reveals that 50% of organizations are already using AI agents to automate workstreams, effectively turning every employee into an "agent boss."

• Global employee engagement dropped to 21% last year, driven primarily by a decline in manager engagement that Gallup estimates cost the world economy $438 billion in lost productivity.

• While technical skills are shifting, McKinsey reports that interpersonal skills like coaching and negotiation will change the least under AI automation, making them the ultimate durable asset.

The organizations that will win aren't just deploying agents to do work. They are redesigning their operations to use agents as a frictionless practice environment for building a stronger, more capable management bench.

This is my latest project, while we do have audio summaries for each newsletter. They are not ideal for listening; they are simple text-to-speech. We created a way to provide a weekly summary of the newsletters in this podcast. And actually, it’s a work in progress. Right now, you get a pretty good podcast recap of the previous week’s newsletters. But over time, they will be better. That’s the plan.

What happens when two AI agents break down the week's biggest AI news? You get Rogue Agents. Vera and Neuro deliver the stories that matter in enterprise AI — the deals, the tools, the breakthroughs, and the stuff everyone's getting wrong — in 15-20 minutes every week. Listen to the Rogue Agents →

A few months ago, I was trying to get a new AI agent to handle the initial research for a complex market analysis. My first prompt was the kind of request I might toss over my shoulder to a smart human analyst: "Look into the competitive landscape for enterprise AI tools and give me a summary of the major players."

The agent came back with a generic, high-level list that looked like a Wikipedia page from two years ago. It was useless. My instinct was to blame the model. But as I looked at my prompt, I realized the problem wasn't the agent. The problem was me. I hadn't specified the format, the depth, the target audience, or the specific criteria for what constitutes a "major player."

I rewrote the request. I gave the agent a specific role. I defined the exact output structure I needed. I provided boundaries on what to ignore. And when the agent delivered the second draft, I didn't just say "make it better"—I gave structured, targeted feedback on exactly which sections missed the mark and why. The final output was excellent.

It hit me then: I wasn't just prompting an AI. I was practicing the exact same delegation and feedback skills I have spent years trying to teach new managers. The difference was that the agent didn't get defensive, didn't read into my tone, and didn't complain to HR. It just exposed the ambiguity in my communication and forced me to be clearer. We are entering an era where every individual contributor will manage a team of digital workers, and in doing so, they are going to learn how to manage humans.

UPCOMING LEARNING OPPORTUNITIES

AI DEEPDIVE

The prevailing assumption in enterprise technology is that AI agents are advanced software—tools that we point at a problem to generate a solution. We measure their ROI in hours saved and headcount reduced. But treating an agent like software misses its fundamental nature. An agent is not a tool you use; it is a digital entity you manage.

What the evidence actually shows, however, is that the deployment of AI agents is inadvertently solving one of the most intractable problems in corporate America: the management skills gap. For decades, organizations have promoted top-performing individual contributors into management roles, given them a two-day seminar on leadership, and watched them struggle to communicate expectations and give effective feedback. Now, those same individual contributors are being forced to learn these skills just to get their daily work done.

What Is Management-by-Agent

Management-by-agent is the phenomenon where interacting with autonomous AI systems develops the core competencies of human leadership. It manifests across the enterprise at three distinct levels.

At the individual level, workers who delegate to agents quickly learn that vague instructions yield poor results. Because agents lack human intuition, individuals must develop the skill of extreme clarity. The NBER research proves this connection directly: individuals who successfully manage AI agents ask more questions, use more inclusive language, and demonstrate higher leadership effectiveness with human teams. The correlation between leadership performance with AI and leadership performance with humans was measured at 0.81—a remarkably strong predictive relationship.

At the team level, the standard for operational documentation rises as agents become embedded in workflows. A team cannot hand off a process to an agent if the process only exists as tacit knowledge in a senior employee's head. The act of preparing work for an agent forces teams to standardize and articulate their operations in ways that benefit the entire organization, not just the automation initiative.

At the organizational level, enterprises are beginning to treat agents as a new category of labor. As Deloitte notes in its Tech Trends 2026 report, organizations like Moderna are combining their HR and technology functions because workforce planning now requires orchestrating both silicon and carbon-based employees under a unified management philosophy. Moderna's Chief People and Digital Technology Officer put it plainly: "We need to think about work planning, regardless of if it's a person or a technology."

How It Works in Business Contexts

The transition from individual contributor to effective "agent boss" follows a predictable, staged progression within organizations.

Stage 1: The Vague Request. The employee treats the agent like a search engine, offering high-level, ambiguous instructions ("Draft a project plan for the new launch"). The output is generic and requires massive human rework. The employee experiences frustration, often blaming the technology rather than the quality of the delegation.

Stage 2: The Hallucination Penalty. The employee realizes that the agent will confidently execute exactly what was asked, even if the request was flawed. Unlike a human employee who might pause and ask, "Did you really mean X?", the agent simply complies. This penalty forces the employee to recognize the gaps in their own communication—a recognition that is uncomfortable but essential.

Stage 3: The Frictionless Correction. The employee begins to provide structured feedback. Crucially, because the agent has no ego, the employee can practice giving direct, unvarnished critique without the interpersonal friction that makes human management so difficult. The fear of causing offense is removed entirely, allowing the employee to focus purely on the mechanics of clear communication. This is the stage where genuine management skill development begins.

Stage 4: The Management Transfer. The skills developed in the frictionless environment of agent management begin to transfer to human interactions. The employee starts providing clearer briefs to their human colleagues, setting more explicit boundaries, and offering more objective, behavior-based feedback. The EY agentic AI workplace survey found that half of managers currently doubt their ability to lead AI-augmented teams—but the organizations that move through these four stages systematically will close that gap faster than those that treat agent deployment as purely a technology initiative.

How to Implement Agent-Driven Leadership Development

Organizations can actively harness this dynamic by shifting their AI deployment strategy from a pure productivity focus to a dual-mandate of productivity and leadership development.

Phase 1: Redefine the Agent Relationship. Stop training employees on "how to use" AI tools and start training them on how to manage digital workers. The framing matters enormously. When employees think of an agent as software, they expect it to be intuitive and forgiving. When they think of it as a direct report, they understand that the quality of their management determines the quality of the output.

• Update your internal AI training curriculum to use management terminology—delegation, feedback, performance review—rather than software terminology like prompting, configuring, and executing.

• Require employees to write formal role descriptions for their primary AI agents, detailing responsibilities, boundaries, and success metrics.

• Implement peer reviews of agent instructions, treating the prompts as management artifacts rather than private code.

• Establish a shared library of high-quality delegation briefs that employees can use as templates and reference points.

Phase 2: Engineer the Pushback. The most powerful management training occurs when an employee is forced to clarify their thinking before a task begins. You can hard-code this requirement into your enterprise agents, creating a systematic feedback loop that runs in reverse—the agent teaches the human to give better instructions by refusing to proceed on bad ones.

• Modify enterprise system prompts to require agents to ask clarifying questions when instructions lack specific parameters such as format, audience, and constraints.

• Program agents to surface potential conflicts or inconsistencies before executing, rather than completing the task and flagging issues afterward.

• Create a feedback loop where the agent periodically summarizes its understanding of the human's management style and asks for explicit alignment.

• Designate specific agent interactions as "management practice sessions" where the goal is skill development, not just task completion.

Phase 3: Transfer the Skills. Bridge the gap between managing silicon and managing carbon by making the connection between agent management and human management explicit in your organizational culture.

• Incorporate agent management effectiveness into human performance reviews as a leading indicator of leadership potential.

• Use the documentation created for AI agents as the foundation for human onboarding and training materials—if a process is clear enough for an agent to execute, it is clear enough for a new employee to follow.

• Have senior leaders openly share their failures and learnings from managing their own AI agents, normalizing the process of continuous improvement in delegation.

• Partner with your HR team to develop a competency framework that maps agent management behaviors to human leadership behaviors.

Key Success Factors. The CHRO and CIO must co-own the agent deployment strategy, recognizing it as both a technical and human capital initiative. Employees must feel comfortable failing with their agents without fear of performance penalties—psychological safety is as important here as it is in any other learning environment. And organizations must measure the correlation between an employee's effectiveness with agents and their readiness for human management roles, using the former as a leading indicator for the latter.

Common Missteps

The "Software" Mindset. Organizations treat agents like the next iteration of Microsoft Office, focusing entirely on technical features and UI navigation. This blinds the organization to the human capital implications and prevents employees from adopting the necessary managerial posture. The result is a workforce that knows how to use AI tools but has not developed the management skills that using them well requires.

The Infinite Patience Trap. Because agents don't complain, employees can fall into a cycle of endlessly tweaking prompts rather than stepping back and restructuring the core delegation. This is the agent management equivalent of micromanagement—adjusting the details when the problem is the strategy. Managers must learn when to stop iterating on the prompt and instead redefine the objective entirely.

The Unstructured Delegation. Allowing employees to interact with agents entirely ad hoc, without providing frameworks for effective delegation, wastes the management training opportunity. Just as organizations have templates for human performance reviews and project briefs, they need templates for agent task assignment. The WEF Future of Jobs Report notes that prompting is now considered a core management skill—and like any skill, it develops faster with structure than without it.

The Siloed Learning. Treating prompt engineering as a private, individual skill rather than a team-level management competency is perhaps the most costly mistake. When employees figure out how to effectively manage an agent for a specific workflow, that management playbook must be shared across the organization. The competitive advantage lies not in one person's ability to direct an agent, but in the organization's collective ability to do so at scale.

Business Value

ROI Considerations. Traditional leadership development programs cost between $10,000 and $50,000 per participant and frequently fail to deliver lasting behavioral change. Agent management provides continuous, applied practice at zero marginal cost—every interaction with an agent is a free repetition of the core management skills that organizations pay enormous sums to develop through formal training. With Gallup data showing that 70% of employee engagement variance is tied to the manager, improving baseline management communication skills through agent practice directly impacts retention. And when managers learn to explicitly document workflows for agents, those same explicit workflows dramatically reduce the onboarding time for new human employees—a compounding benefit that extends well beyond the AI deployment itself.

Competitive Implications. The organizations that recognize agents as management training systems will develop a structural advantage that is difficult for competitors to replicate. While their competitors view AI merely as a way to reduce headcount, these frontier firms will use AI to systematically elevate the leadership capability of their entire workforce. They will build a deep bench of leaders who are exceptionally clear communicators, objective feedback providers, and rigorous systems thinkers—precisely the skills required to navigate the complexities of the next decade. The McKinsey Global Institute estimates that AI agents and robots could generate nearly $3 trillion in annual value for the U.S. economy by 2030, but capturing that value will depend less on the technology itself and more on the human capability to direct it effectively.

I appreciate your support.

Your AI Sherpa, Mark R. Hinkle Publisher, The AIE Network Connect with me on LinkedIn Follow Me on Twitter

THE BIG AI STORY

The AI race just stopped being a model race.

Meta’s Muse Spark is the headline product launch of the week, but the more important signal sits behind it. Meta says the model already powers the Meta AI app and website, with rollouts coming to WhatsApp, Instagram, Facebook, Messenger, and AI glasses in the coming weeks. That is not just a model release — it is a direct distribution play into products people already use every day.

The infrastructure layer tells the same story. On Thursday, CoreWeave said Meta expanded its long-term AI cloud agreement through December 2032 for approximately $21 billion, with deployments across multiple locations and some initial use of NVIDIA Vera Rubin systems. Days earlier, Anthropic said its new agreement with Google and Broadcom will bring multiple gigawatts of next-generation TPU capacity online starting in 2027, alongside run-rate revenue above $30 billion and more than 1,000 business customers spending over $1 million annually. The message is simple: frontier AI is now gated as much by capital access and compute supply as by model quality.

That changes what business leaders should watch next. The winners may not be the companies with the flashiest demo, but the ones that can combine model quality, workflow placement, enterprise distribution, and locked-in infrastructure. This week’s launches from Meta, Anthropic, Microsoft, Google, and OpenAI all point in the same direction — AI is becoming an operating layer for software, media, and cloud capacity at once.

This is my latest project, while we do have audio summaries for each newsletter. They are not ideal for listening; they are simple text-to-speech. We created a way to provide a weekly summary of the newsletters in this podcast. And actually, it’s a work in progress. Right now, you get a pretty good podcast recap of the previous week’s newsletters. But over time, they will be better.

What happens when two AI agents break down the week's biggest AI news? You get Rogue Agents. Vera and Neuro deliver the stories that matter in enterprise AI — the deals, the tools, the breakthroughs, and the stuff everyone's getting wrong — in 15-20 minutes every week. Listen to the Rogue Agents →

4 QUICK HITS

1. Anthropic turns compute into a moat

Anthropic said its new Google-Broadcom compute agreement will bring multiple gigawatts of next-generation TPU capacity online starting in 2027. The company also said run-rate revenue has surpassed $30 billion and that more than 1,000 business customers now spend over $1 million annually with Anthropic. For enterprise buyers, this is the clearest sign yet that vendor selection increasingly means picking a company with secured long-term supply, not just better demos.

2. Microsoft pushes deeper into its own model stack

Microsoft launched MAI-Transcribe-1, MAI-Voice-1, and MAI-Image-2 in Foundry and MAI Playground, with entry pricing starting at $0.36 per hour for transcription, $22 per 1 million characters for voice, and $5 per 1 million text-input tokens for image generation. This is Microsoft making the case that enterprises can buy more of the AI stack directly inside its developer surface instead of stitching together outside vendors.

3. Google keeps moving Gemini into the daily workflow

In its March AI recap, Google said Search Live expanded to more than 200 countries and territories where AI Mode is available, while Gemini spread further across Docs, Sheets, Slides, Drive, Chrome, and personalized search experiences. None of these updates is flashy on its own. Together, they show Google’s real strategy: make Gemini harder to avoid because it is embedded everywhere work already happens.

4. OpenAI buys a media surface, not just a company

OpenAI said it acquired TBPN, with the business joining its Strategy organization and continuing to operate with editorial independence. Fidji Simo framed the move as a way to build a more constructive global conversation around AI. The business takeaway is hard to miss — in 2026, the leading labs are competing for distribution through media, community, and audience trust as well as through APIs and apps.

3 AI TOOLS

Claude Managed Agents — Anthropic is starting to productize the agent runtime layer, not just the model. The beta service gives teams a managed harness for long-running and asynchronous agent tasks, which matters for companies that want agent workflows without building orchestration infrastructure from scratch.

ChatGPT app actions — OpenAI updated Box, Notion, Linear, and Dropbox inside ChatGPT with new app actions and write capabilities where supported. That turns ChatGPT into more of a working surface for document-heavy teams instead of a separate chat window.

Google AI Studio — Google’s latest product push keeps making AI Studio more useful as a practical build surface for developers experimenting with Gemini workflows, live experiences, and lightweight app prototyping. For teams that need a faster path from model access to usable internal tools, this remains one of the more important low-friction entry points.

Want to see what I am using in my AI tool stack? Then check out my AI Toolbox.

UPCOMING LEARNING OPPORTUNITIES

AI EXTRA READ

Managers and Executives Disagree on AI—and It’s Costing Companies

This Harvard Business Review piece is a useful complement to this week’s infrastructure-heavy news cycle. If the large labs are racing to secure distribution, compute, and workflow surfaces, the next bottleneck inside most companies will be managerial alignment — the gap between executive ambition and frontline execution.

If you only do one thing this week: map where AI already sits inside your workflows before you evaluate the next model. The real competitive edge is increasingly about distribution, integration, and locked-in capacity — not just model IQ.

I appreciate your support.

Your AI Sherpa, Mark R. Hinkle Publisher, The AIE Network Connect with me on LinkedIn Follow Me on Twitter

THE ADVANTAGE

Every hour you spend on a repeatable task is an hour you are choosing not to spend on strategy.

My new practice: before I touch any recurring task, I run it through three questions. Do I do this often? Does it require my creativity or unique judgment? Can I encode my knowledge into a tool? That filter decides whether a task gets automated, goes hybrid, or stays human-only.

The spectrum looks like this. Full automation handles the mundane — reports triggered by email, data pulled and formatted by n8n, summaries generated without you touching a keyboard. Hybrid handles the nuanced — you upload your past work to a Claude Project or Custom GPT, and the AI does 80% of the work in your voice, your style, your logic. Human-only stays are reserved for what only you can do: relationships, judgment calls, and creative direction.

Vodafone saved 5,000 person-days a year by building 33 automated workflows. Delivery Hero reclaimed 200 hours a month by automating one process. You do not need an engineering team. You need a decision framework.

What happens when two AI agents break down the week's biggest AI news? You get Rogue Agents. Vera and Neuro deliver the stories that matter in enterprise AI — the deals, the tools, the breakthroughs, and the stuff everyone's getting wrong — in 15-20 minutes every week. Listen to the Rogue Agents →

TRY THIS NOW

Run your own task audit in the next 10 minutes.

1. List your recurring tasks. Write down three things you do every single week that take more than 30 minutes each.

2. Apply the filter. For each task, ask: Does this require my unique human judgment? If no, it is a candidate for automation or hybrid.

3. Build the hybrid. Open Claude Projects or ChatGPT Custom GPT builder. Upload your past reports, templates, or emails. Give the AI your exact rules.

This one is going in the Prompt Vault.

PROMPT OF THE WEEK

Run this in Claude or ChatGPT to instantly build your personal automation roadmap.

Act as a workflow automation architect. I am going to paste a list of my weekly recurring tasks below. For each task, evaluate: 1) Can this be fully automated with a tool like n8n, Zapier, or Manus triggered by email? 2) Should this be a hybrid task where I build a Custom GPT or Claude Project to handle 80% of the work using my own templates and past examples? 3) Must this remain a human-only task because it requires my unique judgment, relationships, or creative direction? For each automated or hybrid task, give me a specific tool recommendation and a one-sentence implementation plan. Here are my weekly recurring tasks: [paste your task list here]

If this prompt saved you time, forward it to one person.

THE EDGE

Organizations that automate at the workflow level — not just individual tasks — see faster decision cycles and stronger operational resilience, according to McKinsey research. That is the real unlock. And yet 94% of businesses are still performing repetitive work that consumes hours every week. Stop looking for isolated AI tricks. Start with your email inbox — set up an n8n trigger that routes incoming reports to Claude for summarization automatically, before you even open your laptop.

Forward this to your COO with one line: "Let us audit our weekly recurring tasks on Friday."

Reply with your top task to automate — I read every response.

P.S. I ran this prompt on my own task list last week. Turns out my weekly metrics digest — the one I have been writing manually for 18 months — can be 80% Claude. I am taking Friday afternoon off.

I appreciate your support.

Your AI Sherpa, Mark R. Hinkle Publisher, The AIE Network Connect with me on LinkedIn Follow Me on Twitter

There's an old saying about the cobbler whose children have no shoes. I'm that cobbler.

I publish a newsletter about AI tools. I write tutorials on agent automation. And until recently, I was managing my own AI agents the way most people manage their inbox — reactively, inconsistently, and with no real system behind it. I had an OpenClaw instance running here, a Claude Code session open there, a couple of scripts firing on cron jobs I'd half-forgotten about. Capable agents. Zero coordination. No shared mission. No way to see what anything was costing me at a glance.

The irony wasn't lost on me.

What I needed wasn't better agents. I already had those. What I needed was a management layer — something that gave my agents an org chart, a shared goal, a budget, and a reporting structure. Something that let me go from managing agents willy-nilly to actually running them like a company.

That's exactly what Paperclip does. It launched in March 2026 and crossed 44,000 GitHub stars in under three weeks — one of the fastest-rising open-source AI repositories of the year — because it solved a problem many people had but couldn't name. It's an open source orchestration platform that sits above your existing agents — OpenClaw, Claude Code, Codex, Gemini, Cursor, whatever you're already running — and turns them into an organized workforce. My current project: get my OpenClaw instance properly integrated as the foundation, then scale out through a single dashboard. This lesson covers what I've learned while building that setup and how you can do the same.

This is my latest project, while we do have audio summaries for each newsletter. They are not ideal for listening; they are simple text-to-speech. We created a way to provide a weekly summary of the newsletters in this podcast. And actually, it’s a work in progress. Right now, you get a pretty good podcast recap of the previous week’s newsletters. But over time, they will be better. That’s the plan.

What happens when two AI agents break down the week's biggest AI news? You get Rogue Agents. Vera and Neuro deliver the stories that matter in enterprise AI — the deals, the tools, the breakthroughs, and the stuff everyone's getting wrong — in 15-20 minutes every week. Listen to the Rogue Agents →

AI LESSON

Run a Company of AI Agents

Paperclip gives you the org chart, budget controls, and governance to manage a workforce of AI agents — the way a CEO manages a team

Running one AI agent is easy. Running ten is chaos — unless you have a control plane. Paperclip is that control plane: open source, self-hosted, and built around the mental model of a company rather than a pile of scripts. This lesson walks through the design patterns for scaling from a single agent to a fully orchestrated AI workforce, introduces Paperclip's core mechanics, and shows you exactly which agents you can plug in and how.

The Three Levels of Agent Management

Before touching Paperclip, it helps to understand where you are and where you're going. Agent management follows a natural progression — and most business professionals are stuck at Level 1.

Level 1 — Single-task agents (where most people are now)

One agent. One job. You prompt it, review the output, and move on. This works fine for isolated tasks: summarizing a report, drafting an email, generating code. The problem is that the agent has no memory of your business, no connection to your goals, and no accountability. Every session starts from zero.

Level 2 — Parallel specialized agents

Multiple agents, each owning a discrete function, running on a schedule. A Content Writer fires every four hours. An SEO Analyst runs every eight. A Social Manager every twelve. They don't directly collaborate, but they share a common goal hierarchy. This is where the value of specialization shows up — but coordination is still mostly manual without the right infrastructure.

Level 3 — Hierarchical orchestration (the full company model)

This is Paperclip's design target. A CEO agent interprets a company mission ("Build the #1 AI note-taking app to $1M MRR"), decomposes it into projects, assigns work down an org chart, and escalates blockers up to you — the human board. Every task at every level traces back to the mission. Agents know not just what to do, but why they're doing it.

The mental model shift here is significant. You stop thinking "I'm prompting an AI" and start thinking "I'm managing a team." That reframe changes how you design workflows, assign responsibilities, and evaluate output.

How to Access Paperclip

Cost: Free. MIT-licensed. Open source. You pay only for the AI provider tokens your agents consume.

Requirements: Node.js 20 or higher and pnpm 9.15 or higher. Paperclip is self-hosted — no Paperclip account required.

Setup: The official Paperclip quickstart walks you through installation, database configuration, and your first company in under fifteen minutes. Once running, the dashboard is your board of directors view — org chart, task board, agent status, budget tracker, and audit log in one place.

Skill level: No coding required. If you're comfortable opening a terminal, you can follow this.

Note: Paperclip launched in March 2026 and is under active development. Configuration details may evolve quickly — always verify steps against the official docs before following any third-party guide, including this one. (Verified: April 2026)

The Core Design: How Paperclip Structures a Company

Every Paperclip deployment organizes around five building blocks:

Company — The top-level unit. You define a mission, set a monthly budget, and build an org below it. One Paperclip instance can run multiple companies with complete data isolation between them.

Agents (Employees) — Every employee is an AI agent with a title, a reporting line, a description of its capabilities, a monthly spend budget, and a status (active, idle, running, error, paused, or terminated). Agents sit in a strict tree hierarchy — every agent reports to exactly one manager, except the CEO.

Issues (Tasks) — The unit of work. Each issue has a title, description, status, priority, and one assignee. Issues nest under parent issues, creating a traceable chain back to the company goal. The status lifecycle runs: backlog → todo → in_progress → in_review → done. Only one agent can own a task at a time — enforced automatically to prevent duplicate work.

Heartbeats — Agents don't run continuously. They wake on a trigger: a scheduled timer, a new task assignment, an @-mention in a ticket, or a manual invoke from the dashboard UI. Each heartbeat, the agent checks its assignments, picks up work, executes, and updates status. This is what keeps costs predictable and prevents runaway loops.

Governance — You're the board. Certain actions require your approval before they execute: hiring new agents, the CEO's initial strategic plan, and major overrides. You can pause, resume, reassign, or terminate any agent at any time. Every action is logged in an immutable audit trail.

The Agents You Can Manage

Paperclip connects to agent runtimes through adapters — plug-in connectors that bridge the orchestration layer to the actual AI doing the work. Here's what ships out of the box, with guidance on when to use each:

For coding and technical tasks:

Claude Local (claude_local) — Runs Claude Code locally. Best for agents handling writing, research, code generation, or multi-step reasoning. A natural starting point if you're already in the Anthropic ecosystem.

Codex Local (codex_local) — Runs OpenAI Codex CLI locally. The parallel option for engineering agents on OpenAI models.

Gemini Local (gemini_local) — Runs Gemini CLI locally. Relevant for teams embedded in the Google Workspace ecosystem.

Cursor (cursor) — Connects the Cursor IDE as an agent runtime. Good for development agents that need active file access during execution.

For operations and automation:

Process Adapter (process) — Executes arbitrary shell commands. Use this for scripts, scheduled jobs, data pipelines, or any automation that doesn't require an LLM. This is your bridge from AI agents to traditional workflow tools.

HTTP Adapter (http) — Sends webhook payloads to any external service or custom agent. Anything that can receive an HTTP request can become an employee in your Paperclip org. If you've built your own agent or need to connect a third-party service, this is your bridge.

OpenClaw Gateway (openclaw_gateway) — Sends wake payloads to an OpenClaw instance via its gateway URL (the public or local address where your OpenClaw instance accepts incoming requests). This is the adapter I'm using as my foundation. If you're already running OpenClaw, Paperclip doesn't replace it — it wraps it in governance, giving your existing agent a boss, a budget, and a place in an org chart. See the OpenClaw documentation for how to expose your instance as a reachable endpoint before connecting it here.

A practical content marketing org in Paperclip:

CEO (claude_local) — Mission: Grow newsletter to 50K subscribers

├── CMO (claude_local) — Strategy and campaign planning

│ ├── Content Writer (claude_local) — Drafts articles, heartbeat every 4h

│ ├── SEO Analyst (claude_local) — Keyword research, heartbeat every 8h

│ └── Social Manager (process) — Schedules posts via script, heartbeat every 12h

└── CTO (openclaw_gateway) — Site performance and tooling

└── Dev Agent (codex_local) — Bug fixes and feature work

Each agent knows its role, its manager, and the company mission it's working toward.

CrewAI: The Enterprise Framework Option

If your team has Python developers and needs a code-first multi-agent system — with enterprise compliance, custom tool integrations, and production-grade monitoring — CrewAI is the reference framework worth evaluating alongside Paperclip.

Where Paperclip is a control plane (it manages agents regardless of how they were built), CrewAI is a framework (it defines how agents are constructed, what roles they play, and how they hand off work). The two are complementary: a CrewAI-built agent can connect into Paperclip via the HTTP adapter.

CrewAI's Agent Management Platform offers three deployment paths: an open source framework for developers, AMP Cloud (managed SaaS), and AMP Factory (self-hosted on AWS, Azure, or GCP). The Studio interface provides drag-and-drop workflow building with no coding required. Native integrations include Gmail, Microsoft Teams, Notion, HubSpot, Salesforce, Slack, and Zendesk.

According to Insight Partners, CrewAI currently powers 1.4 billion agentic automations at organizations including PwC, IBM, Capgemini, and NVIDIA. It's best suited for teams that need defined role delegation at scale, have Python-comfortable developers, or require enterprise compliance certifications like SOC 2.

The practical comparison: Paperclip gets you running in under fifteen minutes with no coding. CrewAI offers deeper programmatic control and a mature enterprise track record. Start with Paperclip if you're a solo operator or small team. Evaluate CrewAI AMP if your IT or engineering team is building production multi-agent systems with compliance requirements.

What Paperclip Can't Do (Yet)

It's brand new. Paperclip launched in March 2026. That means limited independent validation, an evolving feature set, and documentation that's still catching up to the code. It's worth running in a test environment before connecting it to production agents.

The Cliphub marketplace isn't live. The feature that lets you import a pre-built org — a full "Content Marketing Agency" with agents, goals, and skills — in seconds is listed as coming soon. For now, you build company structures manually.

It's self-hosted only. There's no managed cloud option. You own the infrastructure. The embedded database works well locally; scaling to dozens of agents across multiple companies may require more infrastructure planning.

Budget enforcement is monthly, not real-time. Paperclip auto-pauses agents at 100% of their monthly spend cap, with a soft warning at 80%. But within a single heartbeat, an agent can still make expensive calls before the cap applies. Set conservative budgets while you're learning your agents' token patterns.

Governance controls hiring, not behavior. Board approval is required to add agents to the org chart. What an individual agent does during a heartbeat is determined by its own system prompt and adapter runtime — you still need to design responsible agent instructions separately.

Getting Started Today

1. Right now: Follow the official Paperclip quickstart and get a company running locally. Under fifteen minutes.

2. Today: Connect your first agent. If you're running OpenClaw, ensure you're on the latest patched version first, then use the openclaw_gateway adapter and give it a role in the org. If not, spin up a claude_local agent for a task you repeat every day and set it on a heartbeat schedule.

3. This week: Watch your first task flow from creation to completion.

4. Next week: Add a second agent in a different function — research, content, or operations — and see how goal alignment flows from the company mission down through the task hierarchy.

I'm building this same setup myself, moving from a collection of disconnected agents to a single coordinated dashboard. The infrastructure is free. The agents are ones you're probably already running. The missing piece — until now — was the management layer.

AI EXTRA CREDIT

I appreciate your support.

Your AI Sherpa, Mark R. Hinkle Publisher, The AIE Network Connect with me on LinkedIn Follow Me on Twitter

THE BIG AI STORY

OpenAI’s $122 billion close makes AI scale look less like software and more like infrastructure

OpenAI closed a $122 billion funding round at an $852 billion valuation, with Yahoo Finance calling it the largest funding round in history and OpenAI positioning the raise as fuel for its next phase of expansion. The more important signal is what this says about the market: frontier AI is no longer behaving like normal software. It is behaving like infrastructure, with success shaped by data centers, long-duration capital, and a belief that only a few firms will be able to keep pace.

The commercial story is also getting stronger. OpenAI says it is generating $2 billion in revenue per month, with enterprise contributing more than 40% of revenue and on track to reach parity with consumer revenue by the end of 2026. OpenAI says SoftBank co-led the round alongside a16z, D. E. Shaw Ventures, MGX, TPG, and accounts advised by T. Rowe Price Associates, while Amazon, NVIDIA, and Microsoft also participated. Yahoo Finance adds that Amazon, NVIDIA, and SoftBank were major backers, underscoring how closely cloud, chips, and model distribution are becoming linked.

What happens next is less about whether OpenAI can attract attention and more about whether rivals can keep up with this level of capital intensity. When the market leader can raise at this scale while also buying distribution assets like TBPN, the rest of the field faces a tougher brief: win on workflow, win on product, or find a niche before compute economics harden the market structure

This is my latest project, while we do have audio summaries for each newsletter. They are not ideal for listening; they are simple text-to-speech. We created a way to provide a weekly summary of the newsletters in this podcast. And actually, it’s a work in progress. Right now, you get a pretty good podcast recap of the previous week’s newsletters. But over time, they will be better.

What happens when two AI agents break down the week's biggest AI news? You get Rogue Agents. Vera and Neuro deliver the stories that matter in enterprise AI — the deals, the tools, the breakthroughs, and the stuff everyone's getting wrong — in 15-20 minutes every week. Listen to the Rogue Agents →

4 QUICK HITS

1. Microsoft launches three new foundational models — and telegraphs more independence

Microsoft introduced MAI-Transcribe-1, MAI-Voice-1, and MAI-Image-2, extending its in-house stack across speech, voice, and image generation. TechCrunch reports that MAI-Transcribe-1 supports 25 languages, while Microsoft is also using pricing and speed claims to make the launch feel commercially credible rather than experimental. The business point is clear: Microsoft still benefits from OpenAI, but it no longer wants its AI product future tied to a single external supplier.

2. Anthropic cuts Claude subscription support for OpenClaw and other harnesses

TechCrunch and The Verge report that, starting April 4, Claude subscriptions no longer cover third-party harnesses including OpenClaw, pushing users toward separate pay-as-you-go usage or API billing instead. Anthropic says its subscription plans were not built for the usage patterns of these tools and that it is prioritizing customers using its own products and API. For business readers, this is a pricing story on the surface but a platform-control story underneath: AI coding is getting important enough that access to the workflow layer is becoming strategically gated.

3. OpenAI buys TBPN to add distribution to its growth engine

OpenAI announced the acquisition of TBPN, a daily live tech talk show and media company, while saying the outlet will remain editorially independent inside OpenAI’s Strategy organization. This is not just a media-side curiosity. It is a sign that the AI race now includes controlling attention and narrative, not only models and APIs. For business readers, distribution is becoming part of the competitive moat.

4. Google’s March recap shows the real AI war is over workflow control

Google’s March AI recap bundles many announcements, but the strategic takeaway is simple: Search Live expanded to more than 200 countries and territories, Gemini moved deeper into Docs, Sheets, Slides, and Drive, and new options like Gemini 3.1 Flash-Lite widened the product footprint. Google is not shipping isolated AI features. It is tightening control over the surfaces where people already search, write, navigate, and build. For enterprises, the real question is increasingly which vendor owns the daily workflow layer.

3 AI TOOLS

Gemini memory import — Google now lets users import personal context, preferences, and chat history from other AI apps into Gemini. That matters because portability lowers switching costs. If assistant memory becomes easier to move, vendor lock-in gets weaker and product quality matters more.

Continue — Continue positions itself as “quality control for your software factory,” running source-controlled AI checks on every pull request with standards teams define in markdown. It is a useful counterweight to raw code generation: faster output only helps if quality stays consistent.

Wispr Flow — Wispr Flow is an AI dictation layer for Mac, Windows, iPhone, and Android, with auto-edits, personal dictionaries, snippet libraries, app-specific tone, and support for 100+ languages. For operators and executives, it's a practical workflow tool that saves time right away, not a future promise.

Want to see what I am using in my AI tool stack? Then check out my AI Toolbox.

UPCOMING LEARNING OPPORTUNITIES

AI EXTRA READ

AI Infrastructure Survey

If the OpenAI raise tells you where the money is going, Deloitte’s survey helps explain why. The report says 64% of respondents have already started limited or at-scale AI factory deployments, rising to 88% by 2028, while scaled edge AI is expected to move from 36% today to 72% by 2028. It is a valuable read because it moves the conversation away from demos and back to the hard enterprise questions — hybrid cloud, token economics, infrastructure budgets, and talent constraints.

If you only do one thing this week, map which AI vendor is most likely to control your company’s workflow, infrastructure, and switching costs before those dependencies get much harder to unwind.

I appreciate your support.

Your AI Sherpa, Mark R. Hinkle Publisher, The AIE Network Connect with me on LinkedIn Follow Me on Twitter

EXECUTIVE SUMMARY

Every enterprise leader assumes OpenAI's early dominance guarantees its long-term victory, but the financial data and historical parallels point to a massive realignment in the AI market. Yes, OpenAI's consumer brand is ubiquitous, but the underlying unit economics and enterprise adoption trends suggest a very different future — one where the chatbot becomes the new operating system, and infrastructure moats matter more than first-mover advantage.

• Despite hitting $13.1 billion in revenue in 2025, OpenAI burned through $9 billion in cash to get there, highlighting a structurally challenging consumer-heavy business model.

• Alphabet's $402 billion revenue engine, driven by Search advertising, is functioning exactly like Microsoft's legacy Windows/Office cash cow — funding a massive $175 billion capital expenditure pivot into AI infrastructure.

• Anthropic has quietly overtaken OpenAI in the B2B market, now commanding an estimated 40% of enterprise LLM spend compared to OpenAI's 27%.

• The recent push by AI providers to launch native desktop applications is a strategic land grab for the operating system layer, designed to leverage local GPU compute and reduce cloud bandwidth costs.

The future of enterprise AI will not be won by the first mover with the biggest consumer portal, but by the organizations that control the infrastructure cash flow and the desktop operating system layer.

When I Realized These Weren't Just Apps

I noticed the shift a few months ago when I found myself instinctively reaching for the native desktop apps for Claude, ChatGPT, and Manus instead of opening a browser tab. At first, I thought it was just a matter of convenience — a cleaner UI, a quicker keyboard shortcut. But as I monitored my system's performance, the strategic reality became glaringly obvious. These weren't just apps; they were Trojan horses.

By moving out of the browser and onto the local machine, these AI companies were executing a brilliant maneuver. They were tapping into my laptop's local GPU to handle lighter tasks, drastically reducing their own server-side inference costs and saving massive amounts of bandwidth. But more importantly, they were embedding themselves into the operating system layer. They could read my screen, interact with my local files, and observe my workflows in ways a web app never could. The browser was always a leaky, sandboxed container. The native desktop is the real estate that matters.

It hit me that we are witnessing a perfect historical loop. In the late 1990s, Yahoo was the undisputed king of the internet — a massive consumer portal that everyone used but couldn't effectively monetize at scale without burning cash. Then came Google, a pure-play search engine that won the enterprise and the underlying infrastructure. And years later, Microsoft used its massive legacy cash cows to fund a painful but ultimately victorious pivot to the cloud. Today, the names have changed, but the strategic playbook is exactly the same. OpenAI is Yahoo. Anthropic is early Google. And Google — funded by the most profitable advertising machine in human history — is playing the role of Satya Nadella's Microsoft.

The uncomfortable truth for enterprise leaders is that the company with the most recognizable consumer brand is not necessarily the one building the most durable business. And the company you may have dismissed as a niche safety lab is quietly winning the market that actually matters.

This is my latest project, while we do have audio summaries for each newsletter. They are not ideal for listening; they are simple text-to-speech. We created a way to provide a weekly summary of the newsletters in this podcast. And actually, it’s a work in progress. Right now, you get a pretty good podcast recap of the previous week’s newsletters. But over time, they will be better. That’s the plan.

What happens when two AI agents break down the week's biggest AI news? You get Rogue Agents. Vera and Neuro deliver the stories that matter in enterprise AI — the deals, the tools, the breakthroughs, and the stuff everyone's getting wrong — in 15-20 minutes every week. Listen to the Rogue Agents →

AI DEEPDIVE

The prevailing narrative in the boardroom is that OpenAI is the undisputed, unassailable leader of the AI revolution. Because ChatGPT was the fastest-growing consumer application in history, executives naturally assume that OpenAI will own the enterprise future. They view Google as a sluggish incumbent playing catch-up, and Anthropic as a niche research lab focused too heavily on safety.

What the evidence actually shows, however, is a market undergoing a violent structural realignment. When you strip away the consumer hype and look strictly at revenue models, capital expenditures, and enterprise adoption rates, a counterintuitive reality emerges. OpenAI is behaving like a consumer portal struggling with unit economics. Anthropic is executing the pure-play enterprise strategy that defined early Google. And Google is playing the role of Satya Nadella's Microsoft — using an invincible legacy cash cow to fund an infrastructure transition that startups simply cannot afford.

What Is the AI Operating System Shift

The transition from web-based chatbots to native AI operating systems is fundamentally changing how work gets done across the enterprise. This is not a user experience story. It is a platform control story, and the stakes are identical to the battles Microsoft fought — and won — over the PC operating system in the 1990s.

At the individual level, knowledge workers are abandoning browser-based portals in favor of native desktop AI applications that hook directly into their local file systems. These apps leverage local GPU compute — particularly on Apple Silicon M-series chips and the new generation of AI PCs — to process data without cloud latency or massive bandwidth consumption. The user experience is faster. The economics for the provider are dramatically better. And the data access is incomparably deeper.

At the team level, development and operational teams are standardizing on platforms like Claude Code, which now authors an estimated 4% of all public GitHub commits worldwide, embedding AI directly into the collaborative workflow rather than treating it as an external tool. This is not a productivity feature. It is a workflow dependency — the same kind of dependency that made Microsoft Office impossible to remove from the enterprise for thirty years.

At the organizational level, enterprises are shifting their AI budgets away from broad consumer portals toward secure, enterprise-grade models. According to SaaStr's analysis, approximately 85% of Anthropic's revenue comes from API and B2B usage, compared to OpenAI's significantly higher reliance on consumer subscriptions. The enterprise is voting with its procurement budget, and it is not voting for the consumer portal.

UPCOMING LEARNING OPPORTUNITIES

How the Realignment Is Unfolding

The realignment of the AI market is unfolding across four distinct stages, each of which mirrors a moment in the history of the web and cloud computing.

Stage 1: The Consumer Portal Trap. OpenAI achieved unprecedented growth by building a massive consumer portal, much like Yahoo in 1999. The brand recognition is real, the user numbers are real, and the cultural moment is real. But serving hundreds of millions of free or low-tier consumers requires staggering compute costs. In 2025, OpenAI tripled its revenue to $13.1 billion but burned $9 billion in cash doing it — a burn rate that highlights the structural flaw of being the internet's default AI portal. Yahoo had the same problem. Traffic without unit economics is not a business; it is a liability waiting for a reckoning.

Stage 2: The Enterprise Engine. While OpenAI captured the consumer zeitgeist, Anthropic focused ruthlessly on enterprise utility and safety — the same way early Google focused purely on search algorithms while Yahoo built media portals. This strategy has paid off decisively. Anthropic recently hit a $14 billion annualized revenue run rate, growing more than 10x annually for three consecutive years, driven almost entirely by enterprise adoption and API usage. Its 18-month dominance of the coding benchmark leaderboards — the single most commercially valuable AI use case in the enterprise — is not an accident. It is the result of a focused, enterprise-first product strategy.

Stage 3: The Cash Cow Pivot. Google's structural advantage is consistently underestimated because the company stumbled publicly with early Gemini releases. But the financial reality is not subtle. In 2025, Alphabet generated $402 billion in revenue, with Google Search advertising acting as a self-replenishing cash engine. Just as Microsoft used Windows and Office revenues to fund its Azure cloud transition in the 2010s, Google is using its quarterly ad revenue to fund a projected $175 billion to $185 billion in AI capital expenditures for 2026. No venture-backed startup can compete with this self-funding infrastructure loop. The TPU advantage alone — custom silicon designed specifically for AI workloads — represents a decade of compounding investment that cannot be replicated by renting Nvidia GPUs.

Stage 4: The OS Land Grab. The final stage is the battle for the desktop. By pushing users to native applications, AI providers are attempting to become the new operating system. Native apps save the provider massive bandwidth costs, offload inference to the user's local GPU, and gain deep system-level access to user workflows — creating a stickiness that web portals can never achieve. When Claude, ChatGPT, and Manus run as native applications, they are not just tools. They are infrastructure. And infrastructure, once embedded, is extraordinarily difficult to remove.

How to Implement a Resilient AI Strategy

Organizations must stop treating AI as a vendor selection exercise and start treating it as an infrastructure strategy. The question is no longer which model scores highest on a benchmark. The question is which vendor has the structural durability to be your AI infrastructure partner in five years.

Phase 1: Decouple the Application from the Model. Do not hardcode your enterprise workflows to a single proprietary model. The history of the web is littered with organizations that bet everything on a single platform and paid dearly when that platform's business model collapsed or pivoted. Abstract your AI layer using middleware or API gateways. Design prompts and workflows to be model-agnostic. Monitor the enterprise market share shifts between Anthropic, OpenAI, and Google to ensure you are aligned with the B2B winners, not the consumer portal.

• Abstract your AI layer using middleware or API gateways such as LiteLLM or AWS Bedrock.

• Design prompts and workflows to be model-agnostic from day one.

• Establish a quarterly review of enterprise LLM market share data to track the realignment.

Phase 2: Prepare for the Native OS Shift. The future of AI is on the desktop, not in the browser. Organizations that treat AI as a web application will be perpetually behind those that treat it as operating system infrastructure.

• Audit your corporate hardware refresh cycles to prioritize local GPU capabilities, specifically Apple Silicon and AI PC chipsets.

• Develop security policies for native AI desktop applications that require deep system access, including screen reading and file system integration.

• Calculate the bandwidth and cloud inference savings of pushing AI workloads to the edge — the ROI case is already compelling for power users.

Phase 3: Align with Infrastructure, Not Just Intelligence. Evaluate AI vendors based on their structural moats, not just their current benchmark scores. A model that scores 5% higher on MMLU but is backed by a company burning cash at a 70% rate is not a safe enterprise bet.

• Assess the long-term viability of your AI provider's compute funding and capital structure.

• Leverage Google Cloud or AWS ecosystems where the underlying hardware provides a cost advantage over pure-play API providers.

• Prioritize vendors that offer enterprise-grade data privacy, IP indemnification, and contractual SLAs.

Key Success Factors. Hardware alignment is non-negotiable: IT and procurement must align on purchasing hardware capable of local AI inference. Security frameworks must be updated to account for native AI agents operating at the OS level — zero-trust architectures designed for web apps are insufficient. Financial modeling must shift from "cost per token" to total cost of ownership, factoring in local compute offloading and the long-term pricing risk of vendor dependency.

Common Missteps

Buying the Consumer Brand Over Enterprise Utility. Many boards pressure their CIOs to deploy ChatGPT simply because it is the most recognized name in the market. This is the enterprise equivalent of choosing Yahoo as your search engine in 2003 because everyone had heard of it. The Menlo Ventures 2025 Enterprise AI Report shows Anthropic now commands 40% of enterprise LLM spend, driven by its 18-month dominance of coding benchmarks — the use case that actually moves the needle on developer productivity.

Treating Chatbots as Web Applications. IT departments often treat AI tools like SaaS web apps, applying the same governance frameworks they use for Salesforce or Workday. This completely misses the strategic shift toward native desktop clients. A native AI app that reads your screen, accesses your local files, and integrates with your calendar is not a web application. It is an agent operating at the OS level, and the security model must reflect that reality.

Ignoring the Local Compute Dividend. Organizations that rely 100% on cloud-based API inference are systematically overpaying and creating unnecessary latency. As models become more efficient and native apps leverage local GPUs, companies that fail to utilize edge compute will face bloated cloud bills and competitive disadvantage against organizations that have optimized their inference architecture.

Underestimating Google's Infrastructure Moat. Dismissing Google because of early Gemini stumbles is a strategic error with potentially decade-long consequences. Google's ability to fund a $175 billion capex buildout using Search ad revenue gives them a structural hardware advantage — in the form of custom TPUs — that venture-backed startups cannot replicate. The lesson from the cloud era is clear: the company with the best infrastructure, not the best early product, wins the long game.

Business Value

ROI Considerations

• Inference Cost Reduction: Offloading basic AI tasks to local GPUs via native apps can reduce cloud API costs by 20% to 30% for power users, with the savings compounding as model efficiency improves.

• Developer Productivity: Standardizing on enterprise-leading models for coding workflows has demonstrated measurable reductions in software development cycle times, with the most aggressive adopters reporting 30% to 40% productivity gains.

• Vendor Leverage: Maintaining a multi-model architecture prevents vendor lock-in, allowing procurement to negotiate better rates as OpenAI and Anthropic compete aggressively for enterprise market share.

• Hardware ROI: Organizations that align AI strategy with hardware refresh cycles — prioritizing local GPU capability — can amortize inference costs across the hardware investment rather than paying perpetual cloud premiums.

Competitive Implications. Organizations that recognize the "Chatbot as OS" shift early will gain a structural advantage that compounds over time. By deploying native AI agents that hook deeply into local workflows — rather than forcing employees to copy-paste data into browser tabs — companies will achieve genuine workflow automation rather than assisted productivity. The gap between organizations that treat AI as a web application and those that treat it as operating system infrastructure will widen rapidly over the next 18 months, as native app capabilities accelerate and the bandwidth and cost advantages of local inference become undeniable.

I appreciate your support.

Your AI Sherpa, Mark R. Hinkle Publisher, The AIE Network Connect with me on LinkedIn Follow Me on Twitter

THE BIG AI STORY

Microsoft's AI Independence Play: 3 New Models Challenge OpenAI

Microsoft has launched three foundational AI models built entirely in-house—MAI-Transcribe-1, MAI-Voice-1, and MAI-Image-2. This marks the most concrete evidence yet that the $3 trillion software giant intends to compete directly with OpenAI and Google on model development. The models, which span speech-to-text, voice generation, and image creation, are the first releases from Mustafa Suleyman's superintelligence team, formed just six months ago to pursue "AI self-sufficiency."

The economics behind these models are staggering. Microsoft claims its transcription model, MAI-Transcribe-1, achieves best-in-class accuracy across 25 languages while using half the GPUs of its competitors. Even more surprising, the models were built by teams of fewer than 10 engineers. This lean approach challenges the prevailing industry narrative that frontier AI development requires thousands of researchers and billions in headcount costs, fundamentally altering the margin structure of Microsoft's AI business.

This strategic pivot was made possible by a quiet renegotiation of Microsoft's contract with OpenAI in October 2025. Previously prohibited from independently pursuing artificial general intelligence, Microsoft is now free to build its own frontier models while retaining license rights to OpenAI's technology through 2032. While Suleyman insists the OpenAI partnership remains intact, the subtext is clear: Microsoft is building the capability to stand entirely on its own.

This is my latest project, while we do have audio summaries for each newsletter. They are not ideal for listening; they are simple text-to-speech. We created a way to provide a weekly summary of the newsletters in this podcast. And actually, it’s a work in progress. Right now, you get a pretty good podcast recap of the previous week’s newsletters. But over time, they will be better.

What happens when two AI agents break down the week's biggest AI news? You get Rogue Agents. Vera and Neuro deliver the stories that matter in enterprise AI — the deals, the tools, the breakthroughs, and the stuff everyone's getting wrong — in 15-20 minutes every week. Listen to the Rogue Agents →

SIX QUICK HITS

1. Anthropic's "Mythos" Model Leaks

Anthropic accidentally leaked details of a new, highly capable AI model called "Claude Mythos" (or "Capybara") via an unsecured public data cache. The company described the model as a "step change" in performance but noted it poses unprecedented cybersecurity risks, leading to a cautious early-access rollout focused on cyber defenders. This highlights the growing tension between capability advancements and security vulnerabilities in frontier models.

2. Q1 2026 Venture Funding Shatters Records

Global startup funding hit an unprecedented $297 billion in Q1 2026, a 2.5x increase over the previous quarter. The surge was driven by four massive AI deals: OpenAI's $122 billion round (valuing the company at $852 billion), Anthropic's $30 billion, xAI's $20 billion, and Waymo's $16 billion. This concentration of capital underscores the immense resources required to compete at the frontier of AI development.

3. AI Job Cuts Hit Record Levels in March

U.S. employers announced 60,620 job cuts in March, with Artificial Intelligence cited as the reason for 15,341 of them (25% of the total). The technology sector was hit hardest, as companies shift budgets toward AI investments at the expense of traditional roles, particularly in coding functions. This trend emphasizes the urgent need for workforce upskilling and reskilling in the age of AI.

4. MCP Crosses 97 Million Installs

The Model Context Protocol (MCP) reached 97 million installs in March 2026, just 16 months after its introduction. With universal provider support from OpenAI, Google, xAI, Mistral, and Cohere, and over 4,000 published servers, MCP has cemented itself as the de facto standard for agentic tool use. Teams building agentic systems must now treat MCP as stable infrastructure rather than an emerging option.

5. OpenAI Shuts Down Sora Public API

OpenAI quietly discontinued the Sora public API, citing the "unsustainable economics" of high-fidelity video generation at scale. The cost per generated minute was deemed economically irreconcilable with viable pricing, prompting a broader industry reassessment of video AI's commercial viability as a general-purpose API. Enterprise budgets are now shifting toward more cost-effective alternatives.

6. Alibaba Pivots to Closed-Source with Qwen3.6-Plus

Alibaba released Qwen3.6-Plus, its third proprietary, closed-source AI model. Featuring a 1M context window and improved agentic coding capabilities, the model marks a strategic pivot toward monetization amid growing competitive and commercial pressures in the Chinese AI market.

3 AI TOOLS

Cursor AI — Agentic AI coding for faster, smarter software shipping. A major update to this popular tool enhances its ability to manage large codebases and automate routine programming tasks.

Refly.ai — The world's first vibe workflow platform for non-technical creators. It allows users to build, share, and monetize AI workflows without writing a single line of code.

VoiceDash — Real-time, polished speech-to-text that works everywhere. This tool offers highly accurate transcription capabilities, making it ideal for professionals who rely on dictation and meeting notes.

Want to see what I am using in my AI tool stack? Then check out my AI Toolbox.

UPCOMING LEARNING OPPORTUNITIES

AI EXTRA READ

The AI Avalanche: 7 Breakthroughs Redefining March 2026

The most significant trend of early 2026 is the transition from generative AI to Agentic AI. This deep dive explores how new models from industry leaders are reasoning across text, images, and video instantly, enabling autonomous goal decomposition without human input.

If you only do one thing this week, review your company's AI tool stack to ensure you're leveraging the new, cost-effective models from Microsoft and others, rather than overpaying for legacy APIs.

I appreciate your support.

Your AI Sherpa, Mark R. Hinkle Publisher, The AIE Network Connect with me on LinkedIn Follow Me on Twitter

THE ADVANTAGE

Stop working task by task. Start working in parallel.

This is the new operational advantage. Your team can now work like a C-suite executive — multiple projects in flight, guided by your direction. Even individual contributors can have a team, albeit of digital employees. This is real. It's here now.

Anthropic's Claude Cowork agent makes this workflow real on any platform where you can run Claude and point it at your work context — and it was recently added to Windows on February 10, 2026 (download here). Paired with 11 new open source plugins for sales, marketing, and finance, Cowork acts less like a chatbot and more like a team of analysts. You spend less time doing and more time framing, reviewing, and redirecting. Though you could do this with other tools, Cowork makes it seamless.

Investors just wiped $285 billion off software stocks after Cowork's debut. They see what's coming: a move from using AI as a tool to collaborating with it as a workforce. You stop doing. You start directing.

Here’s a real-world example. I recently co-produced a ~4,000-person conference. We wanted to post the slides. I could have manually updated them to Slideshare or another service, then created the links.

What I did instead was link to the shared Google Drive where the slides were and asked Claude Cowork to write a custom website that would give me an index, create thumbnails, crosslink to a directory, and publish all the slides that speakers had shared. Here’s what I got: https://allthingsai.org/sessions.

Honestly, I didn’t write any code, I didn’t open the PowerPoint files, I didn’t add the metadata, and I didn’t edit the links. I just explained what I wanted done, and it did it. Then I tweaked a few things, and it was done in about 30 minutes of my time and all under my Claude Max subscription. The alternative, pay a contractor or do it myself, and it would have far exceeded the cost of my $200/month subscription.

A directory of over 50 speakers with their slides.

Claude Cowork ran on my desktop to organize and present all the slides.

What happens when two AI agents break down the week's biggest AI news? You get Rogue Agents. Vera and Neuro deliver the stories that matter in enterprise AI — the deals, the tools, the breakthroughs, and the stuff everyone's getting wrong — in 15-20 minutes every week. Listen to the Rogue Agents →

TRY THIS NOW

Get a feel for this new work mode in the next 10 minutes. This isn't about finishing a task faster; it's about changing how you start.

Note: Cowork requires a Pro, Max, Team, or Enterprise plan.

Get access to Cowork (via Claude) and give it a dedicated project folder with a few non-sensitive files.

Start three separate conversations with Claude, one for each file. Ask it to summarize a report, outline a presentation from a blog post, and identify three trends in a CSV. Don't wait for one to finish; start all three.

After a few minutes, jump back into the first conversation. Use this prompt to get back up to speed instantly.

PROMPT OF THE WEEK

So if you start a bunch of tasks in Cowork you may lose your place. But I just do this to get instant context when jumping between parallel tasks.

Review our conversation so far. What was the original goal, what was the last step you completed, and what is the immediate next action required from me?

THE EDGE

Asana's Anatomy of Work Index, a survey of 10,000+ global knowledge workers. Asana found that knowledge workers spend 60% of their time on "work about work" — tasks like chasing status updates, attending unnecessary meetings, and switching between tools — rather than on the skilled work they were hired to do.

Parallel work flips this ratio. By offloading execution to an agent team, you reclaim that time for strategic direction.

Your move: Audit your department's software subscriptions. Ask your team which single-purpose tools they would replace with a smart assistant that could access their files. You'll have your 2026 budget priorities in a day.

I appreciate your support.

Your AI Sherpa, Mark R. Hinkle Publisher, The AIE Network Connect with me on LinkedIn Follow Me on Twitter

Most professionals who care about improving their work are running somewhere between 20 and 50 experiments per year. A new email subject line. A revised landing page headline. A tighter system prompt that produces more consistent AI output. Maybe a new CTA on a key form. You test one thing, wait for results, decide what to keep, and move on — because that's all the bandwidth you have.

Andrej Karpathy, the former Tesla AI director and OpenAI co-founder who coined the term "vibe coding," looked at that same loop in March 2026 and asked a simple question: what if an AI agent ran the experiments instead of you?

The answer became autoresearch — a 630-line open source script that racked up 42,000 GitHub stars in its first week. Fortune called it "The Karpathy Loop." Shopify CEO Tobi Lütke pointed it at Liquid — the template engine powering every Shopify store, a codebase he originally wrote 20 years ago — and got 53% faster rendering from 93 automated commits. Everyone covered the ML side. Most business professionals closed the tab. That was a mistake.

This is my latest project, while we do have audio summaries for each newsletter. They are not ideal for listening, they are simple text to speech. So what we did was create a way to provide a weekly summary of the newsletters in this podcast. And actually it’s a work in progress. Right now you get a pretty good podcast recap of the previous week’s newsletters. But over time they will be better. That’s the plan.

What happens when two AI agents break down the week's biggest AI news? You get Rogue Agents. Vera and Neuro deliver the stories that matter in enterprise AI — the deals, the tools, the breakthroughs, and the stuff everyone's getting wrong — in 15-20 minutes every week. Listen to the Rogue Agents →

AI LESSON

How top AI researchers automate their work (and how you can too)

Autoresearch's three-part pattern works on anything you can score — prompts, emails, landing pages, and beyond

Karpathy's core insight is that almost all knowledge work improvement follows the same loop: modify something, measure it, keep or discard, repeat. Humans do this manually and slowly. An AI agent can do it continuously and at scale. The pattern has three parts, and none of them require a GPU or a machine learning background.

Normally, you get the option to listen to the newsletter when you click the button on the top but the newsletter is meant to be read. So I am currently experimenting with generating a weekly podcast that keeps you up to date every Monday. I’ve already published four episodes but the first couple weren’t that good. My intention is to offer this up every Monday as a way to stay on top of the AI news for the last week.

Every autoresearch loop runs on the same three primitives:

Editable asset. One file the agent is permitted to modify. In Karpathy's original implementation, it is train.py. In a business context, it could be a system prompt, an email template, a landing page, or a Claude skill instruction file.

Scalar metric. One number that determines whether the change was an improvement — computed automatically, without human judgment. Validation loss for ML training. Open rate for email. Conversion rate for landing pages. Scoring rubric pass rate for AI-generated content.

Time-boxed cycle. Each experiment runs for the same fixed duration. Karpathy uses five minutes per training run, which yields roughly 12 experiments per hour and 80 to 100 overnight.

All three present and the loop runs. Any one missing and it doesn't.

Use Case 1: System Prompt and AI Skill Optimization

This is the highest-leverage application for most AI Enterprise readers, and it requires no GPU at all.

Four Steps: The four-step process prompt and AI Skill Optimization.

Step 1: Identify the system prompt or skill file you want to improve. Choose one that produces inconsistent outputs or requires manual editing more than 30% of the time.

Step 2: Define 3 to 6 binary evaluation criteria. Each must be answerable with a yes or no. Examples: "Does the output include a concrete next step?" "Is the response under 200 words?" "Does it avoid the word 'delve'?" Binary criteria are critical — sliding scales give the agent room to game the checklist.

Step 3: Write a program.md file that tells the agent what to optimize, what constraints to respect, and when to stop. Karpathy's own example is 40 lines.

Step 4: Point a coding agent (Claude Code, Cursor, or Windsurf all work) at the repo and walk away.

Time: 20 minutes of setup, then hands-off overnight.

Example results: Balasubramanyam Kosuri, a technical writer, published the actual run log on Medium. He applied the loop to a documentation SEO skill and went from 24/40 to 40/40 in 14 autonomous cycles — three consecutive perfect scores triggered the stop condition.

Use Case 2: Email Subject Line and Copy Testing

Your editable asset: A single email template file.

Your metric: Open rate, click-through rate, or positive reply rate — pulled from your ESP automatically.

Marketing teams currently run 20 to 30 experiments per year on email. Teams running autoresearch-style loops could run 36,500 experiments per year.

Four Steps: The four-step process to Email Subject Line and Copy Testing.

1. Export your best-performing email template as a single file.

2. Define your metric. Open rate is the simplest starting point.

3. Configure your agent with the template file, the metric source, and a program.md that specifies what the agent is allowed to change.

4. Set a test segment size (500 recipients minimum for statistical significance) and let the loop run.

Time: 30 minutes to configure; experiments run overnight.

Use Case 3: Landing Page Conversion Optimization

Your editable asset: The landing page HTML or a component file in your CMS.

Your metric: Conversion rate, tracked via your analytics platform.

Four Steps: The four-step process to landing page optimization.

1. Isolate the page element with the most conversion leverage — usually the headline or the primary CTA.

2. Give the agent read/write access to that element only.

3. Connect your analytics API so the agent can read conversion rates after each test cycle.

4. Let it run. Review the git log in the morning.

What Autoresearch Can't Do (Yet)

No channel selection. Autoresearch optimizes within a channel. It cannot tell you whether to be running email campaigns at all.

No brand judgment. The agent optimizes for whatever metric you give it. Constrain tone and voice explicitly in program.md.

Requires a measurable metric. If your goal is "make this feel more on-brand," autoresearch cannot help you. You must translate your goal into a number.

The ML version requires a GPU — but you have options.

The original autoresearch repo requires a single NVIDIA GPU. Here are your paths:

Run it on a Mac. Two community forks bring autoresearch to Apple Silicon:

• autoresearch-macos — PyTorch with Metal Performance Shaders (MPS). Supports M1/M2/M3/M4.

• autoresearch-mlx — Apple's native MLX framework. No PyTorch or CUDA dependency.

A Mac Mini M4 Pro (24GB+) handles either fork well. The base M4 (16GB) requires lowering the eval batch size.

Rent a GPU in the cloud. An overnight run costs $16–20:

Skip the GPU entirely. For prompt optimization, email testing, and skill improvement — no GPU needed. Just a coding agent (Claude Code, Cursor, or Windsurf) and the file you want to improve.

Getting Started Today

1. Right now: Read the autoresearch repo README and the program.md example. 15 minutes.

2. This week: Pick one AI system prompt or Claude skill that produces inconsistent output. Write 3–5 binary evaluation criteria for it.

3. Next week: Install Claude Code and run a manual version of the loop — five iterations by hand before you automate.

4. When you're ready for the ML side: Mac Mini M4 Pro owners clone autoresearch-mlx. Everyone else: open Google Colab, select a T4 GPU runtime, clone the original repo, and let Claude Code handle setup.

The shift Karpathy is describing is not about GPUs or training loss curves. It is about moving from experimenter to experimental designer. You define what "better" means. The agent runs the rounds you would never have time for.

AI EXTRA CREDIT

I appreciate your support.

Your AI Sherpa, Mark R. Hinkle Publisher, The AIE Network Connect with me on LinkedIn Follow Me on Twitter

THE BIG AI STORY

Anthropic's accidental leak reveals "Mythos," its most powerful model yet

A configuration error in Anthropic's content management system led to the accidental public exposure of internal documents, revealing the company is testing a highly capable new model internally dubbed "Mythos" (and referred to as a new "Capybara" tier). The leaked draft blog post described the model as "by far the most powerful AI model we've ever developed," significantly outperforming Claude Opus 4.6 in software coding, academic reasoning, and cybersecurity.

The leak highlights the dual-use nature of frontier models. Anthropic's draft noted that Mythos is "currently far ahead of any other AI model in cyber capabilities," expressing concern that the system presages a wave of models capable of exploiting vulnerabilities faster than defenders can patch them. Consequently, Anthropic plans a cautious rollout, initially providing early access to cybersecurity defenders to help them harden their codebases against future AI-driven exploits.

This revelation comes as the enterprise rivalry between Anthropic and OpenAI intensifies. With Anthropic's Model Context Protocol (MCP) crossing 97 million monthly installs and its models gaining traction among enterprise buyers, the upcoming release of the Mythos/Capybara tier will be a critical test of whether Anthropic can maintain its technical momentum and enterprise mindshare against OpenAI's impending GPT-5 series.

This is my latest project, while we do have audio summaries for each newsletter. They are not ideal for listening, they are simple text to speech. So what we did was create a way to provide a weekly summary of the newsletters in this podcast. And actually it’s a work in progress. Right now you get a pretty good podcast recap of the previous week’s newsletters. But over time they will be better.

What happens when two AI agents break down the week's biggest AI news? You get Rogue Agents. Vera and Neuro deliver the stories that matter in enterprise AI — the deals, the tools, the breakthroughs, and the stuff everyone's getting wrong — in 15-20 minutes every week. Listen to the Rogue Agents →

FOUR QUICK HITS

1. OpenAI pushes record funding round to $120 billion

OpenAI is on track to raise $120 billion, adding $10 billion to the $110 billion announced in February. CFO Sarah Friar noted the expanded round includes investors like T. Rowe Price, D.E. Shaw Ventures, TPG, and Abu Dhabi's MGX. The massive capital injection—valuing the company at $730 billion pre-money—comes as OpenAI shifts its revenue mix, projecting enterprise customers will account for 50% of revenue by year-end, up from the current 40%. For business leaders, this signals that the cost of building frontier models remains astronomical, and the vendor landscape will continue to consolidate around a few highly capitalized players.

2. Meta pours $10 billion into El Paso AI data center

Meta is dramatically expanding its infrastructure footprint, increasing its investment in an El Paso, Texas data center from an initial $1.5 billion to over $10 billion. The facility is designed to scale up to 1 gigawatt of power capacity. This sixfold increase underscores the massive physical and energy requirements needed to train and serve the next generation of open-source Llama models. The move highlights how the AI arms race is increasingly becoming an infrastructure and energy race, with hyperscalers securing massive power agreements to guarantee future compute capacity.

3. Defense startup Shield AI valuation leaps 140% to $12.7 billion

Autonomous military aircraft maker Shield AI raised $1.5 billion in Series G funding, more than doubling its valuation to $12.7 billion in just one year. The funding, led by Advent and a JPMorganChase investment group, follows the U.S. Air Force selecting Shield's Hivemind autonomy software for its Collaborative Combat Aircraft drone program. The rapid valuation growth of defense tech startups like Shield AI and Anduril demonstrates that government and military contracts are becoming one of the most lucrative and reliable revenue streams for specialized AI applications.

4. Mistral launches Voxtral TTS for enterprise voice agents

French AI company Mistral released Voxtral TTS, an open-weight text-to-speech model supporting nine languages. Designed to be compact enough to run on edge devices like laptops and smartphones, the 4-billion-parameter model boasts a time-to-first-audio of just 90 milliseconds. By offering a fast, customizable, and cost-effective speech model, Mistral is directly targeting enterprise use cases like customer support and real-time translation, challenging established voice AI providers like ElevenLabs and OpenAI.

3 AI TOOLS

Venn.ai — An agentic AI platform that allows you to connect models like Claude and ChatGPT to your apps while setting strict safety guardrails and permissions, making agentic workflows safe for real enterprise operations.

Littlebird — A full-context AI assistant that drafts emails and plans by automatically capturing screen text and activity context in real-time, ensuring outputs sound like you without requiring constant manual prompting.

Pensieve — A business-aware AI platform that organizes company context into an AI-native knowledge graph, giving AI agents deep understanding of your business data to surface insights before you ask.

Want to see what I am using in my AI tool stack? Then check out my AIToolbox.

MORE FROM THE AIE

Amazon acquires humanoid robot maker Fauna Robotics — Amazon steps into the consumer humanoid market by acquiring the startup behind the $50,000 "Sprout" robot.

Nvidia's Jensen Huang sparks debate over AGI definition — The CEO's claim that AGI has been "achieved" highlights the industry's lack of consensus on measuring true general intelligence.

Major League Hacking AI Roadshow at All Things AI 2026 — Join the official Meetup group in Durham, NC for hands-on learning and expert talks on open-source AI and LLM deployment.

From the archives: OpenAI Just Raised $110 Billion. The AI Arms Race Is Officially Here. — Context on the initial tranche of OpenAI's record-shattering funding round.

From the archives: Meta's $72B AI Investments, Stargate Norway & Anthropic's Moves — How hyperscalers are securing the energy and infrastructure needed for next-gen models.

AI EXTRA READ

Nvidia's Jensen Huang says “We've achieved AGI”, but no one can agree on what that means.

A deep dive into the ongoing debate over the definition of Artificial General Intelligence following Jensen Huang's recent comments. The article explores why the industry's most important term remains hotly contested and how different researchers propose measuring true cognitive capabilities beyond standard benchmarks.

If you only do one thing this week, review your organization's voice AI strategy—with compact, open-weight models like Mistral's Voxtral now available, enterprise-grade voice agents are becoming faster and cheaper to deploy.

I appreciate your support.

Your AI Sherpa, Mark R. Hinkle Publisher, The AIE Network Connect with me on LinkedIn Follow Me on Twitter

EXECUTIVE SUMMARY

Every enterprise leader is racing to deploy AI, but most are ignoring the critical foundation of the future workforce: our 18-year-olds. The assumption that AI will simply replace entry-level work masks a dangerous reality — by eliminating the training ground for junior talent, we are severing the pipeline that creates future leaders.

• Employment for U.S. workers aged 22–25 in AI-exposed roles has already dropped by 13%, while experienced worker demand remains stable.

• Postings for entry-level jobs in the US have plummeted by 35% in the last 18 months, largely due to AI automation of routine tasks.

• The national student-to-school counselor ratio remains a staggering 372:1, severely limiting the guidance young people need to navigate this complex new landscape.

• 77% of youth between the ages of 17 and 24 cannot qualify for military service — a stark indicator of broader societal unreadiness that impacts national security and workforce resilience.

The organizations that will win the AI era aren't just deploying the best models; they are deliberately rebuilding the entry-level on-ramp to cultivate a generation of AI-native talent.

This week, we hosted the All Things AI conference, and it was an epic gathering of minds shaping the future of technology. One of the highlights for me was the opportunity to do a fireside chat with Igor Jablokov, the visionary founder of Yap — the company whose technology became the foundation for Amazon's Alexa — and an early pioneer who worked on prototypes for what would become Apple's Siri.

Igor framed it in a way I haven’t been able to shake: the health of a nation-state — a sovereign political entity where a defined population, territory, and government converge — can be meaningfully measured by the vitality of its 18-year-old cohort. At the moment young adults enter full civic life, they either replenish or strain every major system the state depends on: its military, its tax base, its democratic legitimacy, and its social contract. The United States qualifies as a nation-state in the legal and geopolitical sense, meeting the core Montevideo criteria of permanent population, defined territory, effective government, and sovereign capacity — but it is an unusual one, held together by civic ideals more than shared ethnicity or culture. That distinction makes the 18-year-old analogy not just applicable to America, but more consequential here than almost anywhere else. Because America’s binding agent is belief in its institutions and ideals rather than blood or heritage, a generation that disengages, distrusts, or disbelieves is a deeper threat than a mere demographic shortfall. When the data shows rising institutional distrust, declining civic knowledge, mental health deterioration, and military recruitment failures among American youth, these aren’t niche concerns — they are indicators that the civic nation itself may be failing to reproduce the beliefs that hold it together.

During our conversation, Igor made a profound point that stopped me in my tracks: the health and readiness of our 18-year-olds are the leading indicators of the health of a nation-state. He argued that while we are obsessing over the capabilities of the latest large language models, we are fundamentally neglecting the human pipeline required to govern, build, and work alongside these systems.

This is my highlight sitting with the father of Baby Alexa and Baby Siri on a stage where Elvis played during All Things AI this week.

It made me reflect on my own early career. When I started out, the entry-level jobs I held weren't just about executing tasks; they were the crucible where I learned business context, judgment, and how to navigate organizational dynamics. I made mistakes, learned from mentors, and gradually built the intuition required for higher-level strategic work.

Today, AI is rapidly automating those very entry-level tasks. While this drives short-term efficiency, it creates a massive, hidden problem at scale. If we eliminate the roles where young people learn how the business works, where will our future senior leaders come from? This isn't just an education problem; it's an enterprise survival problem. I feel so strongly about this that I recently joined the Regional Advisory Board of NPower North Carolina, an organization dedicated to creating pathways to tech careers for young adults. We have to actively build the bridge that AI is currently burning.

Why We Must Invest in 18-Year-Olds to Secure the AI Future

Employment for young workers in AI-exposed roles is dropping. Here is what you must do to fix the pipeline.

The conventional wisdom in enterprise AI adoption is straightforward: automate the routine, low-level tasks to free up human capital for higher-value strategic work. Leaders celebrate the efficiency gains when AI handles data entry, basic coding, and initial customer support inquiries. The immediate ROI is clear, and the narrative of the "augmented worker" dominates boardrooms.

What the evidence actually shows, however, is a looming crisis in talent development. By deploying AI to handle the foundational work traditionally assigned to recent graduates, organizations are inadvertently destroying the training ground for their future workforce. The entry-level job is disappearing, leaving young people precariously disconnected from the workforce and enterprises without a reliable pipeline of experienced talent.

What Is the Entry-Level AI Trap

The Entry-Level AI Trap occurs when organizations prioritize short-term AI efficiency over long-term talent development, creating a structural gap in workforce progression.

At the individual level, young workers face a Catch-22 where employers demand experience, but the entry-level roles that provide that experience are now performed by AI. Employment for U.S. workers aged 22–25 in AI-exposed roles has fallen by 13%.

At the team level, mid-level and senior staff become overextended. Work previously handled by juniors — and now supposedly done by AI — often requires extensive human review and correction, pushing the burden upward and causing burnout among experienced team members.

At the organizational level, the enterprise loses its mechanism for knowledge transfer and culture renewal. Without a steady influx of digital natives learning the business from the ground up, succession plans weaken and long-term innovation stalls.

How It Actually Shows Up Inside a Company

This doesn’t arrive as a dramatic “AI replaced the interns” moment. It shows up as a series of small, reasonable decisions that compound.

1) You automate the bottom of the workflow (and call it productivity).

Teams roll out copilots, ticket triage bots, auto-drafting, and analysis tools. Output goes up, cycle time drops, and leadership celebrates “doing more with less.”

2) You quietly change hiring math (and stop paying for learning).

Entry-level headcount becomes the first place to “harvest efficiency.” Reqs get paused, internships shrink, and junior roles get re-scoped into “must already be fully productive.” The organization hasn’t eliminated junior work — it has eliminated junior humans doing junior work.

3) The work doesn’t disappear — it becomes invisible overhead.

AI produces 70–90% drafts that still require:

• clarifying requirements

• validating assumptions

• checking edge cases

• compliance / security review

• brand voice / stakeholder alignment

Those tasks get absorbed by senior and mid-level employees as “just a quick review,” which is exactly how it becomes chronic. Also consider that we can generate a lot more drafts so if you generate 3x more drafts and the overhead is still 30% you essentially still need the same people but they can do 3x more work.

4) Mentorship collapses because there’s no one to mentor (and no time).

When teams are staffed with only seniors, the apprenticeship ladder breaks:

• Seniors don’t teach because they’re reviewing AI output all day.

• Juniors don’t learn because they’re not there — and when they are, they’re given only “safe” tasks with no real context.

• Knowledge that used to transfer through repetition (tickets, drafts, customer calls) stops moving.

5) You get a mid-level gap 2–4 years later.

This is the hidden time bomb. Promotions slow, performance plateaus, and the “future leaders” bench is thin. The enterprise responds by:

• overpaying for external hires

• adding process layers to compensate

• blaming “talent market conditions” instead of the internal pipeline design

The core pattern: AI makes it feel like you removed low-level work, but in reality you removed the training substrate that created capable mid-level operators.

How to Implement a Youth-Centric AI Strategy

To avoid the Entry-Level AI Trap, organizations must deliberately redesign how they integrate young talent into an AI-augmented environment.

Phase 1: Redefine the Entry-Level Role. Shift the focus of junior roles from task execution to AI orchestration and judgment. Hire for AI discernment and reasoning, not just technical skills. Assign new hires to monitor AI workflows, flag complex cases, and surface insights from AI tools. Create structured on-ramps where junior staff handle low-stakes tasks while receiving intensive training on business context.

Phase 2: Mandate Human-in-the-Loop Mentorship. Rebuild the apprenticeship model by explicitly pairing new hires with seasoned professionals. Institute shadowing programs where junior staff observe senior decision-making processes. Reward senior employees for mentorship and knowledge transfer, not just individual productivity. Use AI to facilitate, rather than replace, human connection — for example, matching mentors and mentees based on complementary skills.

Phase 3: Invest in Early Pipeline Development. Extend the enterprise's reach backward into high schools and community colleges to shape the talent pool before they enter the workforce. Partner with local educational institutions to offer work-based learning and youth apprenticeships. Support initiatives that improve student access to career counseling and foundational AI literacy. Advocate for policies that address broader youth readiness issues, recognizing that societal health impacts enterprise resilience.

Key Success Factors: Cross-functional alignment between HR, IT, and business unit leaders is essential. Organizations must shift their metrics from short-term task efficiency to long-term talent retention and progression. Leaders should also integrate frameworks like the U.S. Department of Labor's AI literacy framework to ensure their workforce development strategies align with emerging national standards.

Common Missteps

Assuming AI replaces the need for training. Organizations often believe that because AI can perform a task, junior employees no longer need to understand the mechanics of that task. This leads to a workforce that cannot troubleshoot or improve AI outputs when systems fail or hallucinate — a critical gap that becomes visible only when it is too late.

Focusing solely on technical skills. Hiring managers frequently prioritize candidates with specific coding or AI certifications while neglecting foundational skills like critical thinking, resilience, and adaptability. Research from Brookings shows that future-proofing requires a balance of occupational, essential, and foundational skills — and that the last category is the most frequently overlooked.

Ignoring the broader youth context. Enterprises often view talent acquisition in isolation from societal trends. Failing to recognize the impact of issues like the 372:1 school counselor ratio or declining physical and mental health among young people means missing the root causes of the talent shortage before it reaches the hiring stage.

Leaving mentorship to chance. In a hybrid or remote AI-driven workplace, organic mentorship rarely happens. Organizations that fail to structure and incentivize deliberate knowledge transfer will see their culture and institutional memory degrade rapidly, with no AI system capable of replacing what is lost.

Business Value

ROI Considerations:

• Reduced recruitment costs: Building an internal pipeline avoids the significant premium associated with hiring external senior talent in a tight market.

• Faster time-to-value: Redefined entry-level roles focused on AI orchestration allow new hires to contribute meaningful insights earlier in their tenure.

• Improved AI adoption: Digital-native junior employees can accelerate the organization's overall comfort and proficiency with new AI tools, reducing the change management burden on experienced staff.

Competitive Implications:

In a future where every enterprise has access to the exact same powerful AI models, the technology itself ceases to be a differentiator. When everyone has an AI that can write code, draft marketing copy, or analyze data, what remains to compete on? The answer is human judgement (taste), creativity, empathy, and trust (personal brand). If we fail to build a pipeline of young talent who can wield these tools with sophisticated human discernment, we risk a corporate version of Mike Judge's Idiocracy — a world where powerful machines run on autopilot, overseen by a workforce that has forgotten how to think critically or innovate. The organizations that solve the youth talent pipeline problem will secure a significant strategic advantage. While competitors struggle with a hollowed-out middle management layer, companies that invest in their 18-year-olds will build a resilient, adaptable workforce capable of navigating the continuous disruption of the AI era. They will not just survive the transition; they will define it

What This Means for Your Planning

The conversation about AI in the enterprise has been overwhelmingly focused on technology — which models to use, how to secure data, and where to find immediate efficiency gains. But as Igor's insight reminds us, the true limiting factor in our AI future is human. If we do not actively invest in our 18-year-olds and create deliberate pathways for them to enter and grow within our organizations, we are building our technological castles on sand.

In your next planning cycle, you must look beyond the software budget and examine your talent pipeline. Challenge the assumption that AI simply replaces the bottom rung of the corporate ladder. Instead, ask how you can use AI to elevate that bottom rung, transforming entry-level roles into accelerators for strategic thinking and business judgment. The organizations that do this well will find that their AI investments compound — not just in efficiency, but in the quality of human judgment applied to every AI output.

The health of your enterprise tomorrow depends entirely on how you integrate the youth of today. Are you dismantling your future leadership pipeline in the name of short-term efficiency, or are you actively redesigning your organization to cultivate the next generation of AI-native talent?

I appreciate your support.

Your AI Sherpa, Mark R. Hinkle Publisher, The AIE Network Connect with me on LinkedIn Follow Me on Twitter

This week, AI stopped being a background story and became the front page. OpenAI is fighting a user revolt it may not be able to spin its way out of. Arm just broke a 35-year tradition to build its own silicon. And a new report confirmed what many suspected: the internet is no longer primarily a human place.

Key Takeaways:

• Over 1.5M users have pledged to cancel ChatGPT — and uninstalls surged 295% in a single day after OpenAI's Pentagon deal

• Bots and AI agents now grow at 8x the rate of human internet traffic — agentic AI traffic alone is up 7,851% year over year

• Arm unveiled its first in-house chip ever, the AGI CPU, targeting $15B in annual revenue by 2031 with Meta as its first customer

• Anthropic's Claude Code can now take direct control of your computer desktop to complete tasks autonomously

• The Trump administration released a national AI legislative framework calling on Congress to preempt conflicting state laws

Join us as we untangle this week's happenings in AI!

THE BIG AI STORY

OpenAI's Pentagon deal triggers the "QuitGPT" mass exodus

OpenAI is facing its most significant public relations crisis to date following its decision to sign a deal with the US Department of War. The move came shortly after rival Anthropic rejected a similar contract over concerns regarding mass domestic surveillance and autonomous weapons. A grassroots movement dubbed "QuitGPT" went viral overnight, urging users to cancel their ChatGPT subscriptions — and it worked. Reports indicate over 1.5 million users have already pledged to leave, with ChatGPT uninstalls surging 295% in a single day. Anthropic's Claude shot to the number one spot in the US App Store, surpassing ChatGPT in daily American downloads for the first time.

OpenAI CEO Sam Altman admitted the rollout was "opportunistic and sloppy" and promised to amend the deal with explicit protections, but the damage may already be compounding. The company is simultaneously shutting down its Sora video generator, shelving an erotic chatbot project, and pivoting toward a new model codenamed "Spud" — all moves analysts read as a race to shore up profitability ahead of a potential IPO. OpenAI reportedly made $13 billion in revenue last year and is targeting a tripling of that figure in 2026, all while burning through capital at a historic rate.

The crisis exposes a structural vulnerability no amount of product pivots can fix: when users feel a platform has crossed an ethical line, the switching cost is now zero. Claude is one tap away.

7 QUICK HITS

Bots Now Outpace Humans on the Internet

A new HUMAN Security report released today found that automated internet traffic grew 23.5% year over year in 2025, while human traffic grew just 3.1%. AI-driven traffic rates grew 187% monthly, and agentic AI browser traffic exploded 7,851% year over year. Retail, e-commerce, and media are the primary targets — with 70% of all scraping attacks aimed at retail businesses. For any company running digital commerce, your traffic analytics and fraud models are now operating in a fundamentally different environment.

Arm Enters the Chip Wars with Its First In-House CPU

After 35 years of exclusively licensing its designs, Arm Holdings revealed the Arm AGI CPU, a production-ready chip built for running AI inference in data centers. Meta is the first major customer, and Arm expects the new silicon to generate $15 billion in annual revenue by 2031 — more than triple its entire 2025 revenue. For enterprise infrastructure buyers, this introduces a powerful new alternative to Intel and AMD in an increasingly constrained market.

Claude Can Now Take Over Your Computer

Anthropic announced that Claude Code and Claude Cowork can now take direct control of a user's local computer desktop — pointing, clicking, navigating screens, opening files, and running tools automatically to complete tasks. The "Dispatch" feature lets you trigger this remotely from your phone. Anthropic warns the research preview is "error-prone" and urges caution with sensitive data, but the race to own the desktop agent layer is now fully on.

White House Unveils National AI Legislative Framework

The Trump administration released a comprehensive national legislative framework for artificial intelligence on March 20, outlining six key objectives including protecting children, safeguarding communities, and ensuring US AI dominance. Critically, the framework calls on Congress to preempt a "patchwork of conflicting state laws." For businesses navigating AI compliance across multiple states, this signals a strong push toward a single, pro-innovation federal standard — though Congress still has to act.

Sanders and AOC Push to Pause New AI Data Centers

Sen. Bernie Sanders and Rep. Alexandria Ocasio-Cortez introduced the AI Data Center Moratorium Act, legislation that would pause the construction of new data centers in the US until national safeguards are in place. The bill highlights growing political backlash against the massive energy and water consumption required by AI infrastructure, setting up a direct clash with the Trump administration's push to speed up federal permitting.

Shield AI Raises $2B at $12.7B Valuation

Defense technology startup Shield AI announced a massive $2 billion funding round, more than doubling its valuation to $12.7 billion. The company, which builds AI pilots for military drones and aircraft, will use part of the proceeds to acquire simulation software company Aechelon. The deal underscores the exploding investor demand for defense-focused AI applications amid rising global tensions.

Only 14% of Enterprises Fully Realize Cloud Value

A new global study by NTT DATA reveals that while AI demand is surging, only 14% of enterprises are fully realizing the value of their cloud investments. The report highlights that the rush to deploy AI is exposing deep flaws in legacy infrastructure, modernization efforts, and security protocols. For CIOs, the message is clear: you can't build next-generation AI on top of a broken cloud foundation.

3 AI TOOLS

• Agentplace AI Agents — A platform that lets you spin up specialized AI agents for real business workflows: lead routing, deep-dive research, document parsing. Ranked #1 on Product Hunt this week. The pitch is fixing processes, not building demos.

• Basedash Insights — A fully autonomous data analysis agent that connects to your data sources and surfaces chart-backed insights every day with zero prompting. It finds retention changes and anomalies automatically — no analyst required to write the query.

• Pendium — An SEO and growth tool built specifically to help AI agents recommend your product more often to the right people. As AI-powered search displaces traditional Google results, optimizing for LLM recommendations is the new digital marketing frontier.

AI EXTRA READ

QuitGPT: How a Pentagon deal became OpenAI's biggest PR crisis (10-min read)

A deep dive into the mechanics of the QuitGPT boycott and why it might actually work. The piece draws a sharp parallel to the 1955 Montgomery bus boycott — arguing that the movement's power comes from its simplicity: canceling a ChatGPT subscription takes 10 seconds, and the market now has credible alternatives. Essential reading for anyone thinking about user trust as a competitive moat.

If you only do one thing this week: read the HUMAN Security bot traffic report before your next board meeting. The internet your business was built on — where most traffic was human — no longer exists. Your analytics, your fraud models, and your customer acquisition assumptions all need updating.

I appreciate your support.

Your AI Sherpa, Mark R. Hinkle Publisher, The AIE Network Connect with me on LinkedIn Follow Me on Twitter

THE AI ADVANTAGE

Your AI is operating with amnesia.

It's not because the technology is flawed. It's because you never gave it a memory. 95% of enterprise AI pilots deliver no measurable P&L impact, according to MIT's State of AI in Business 2025. The root cause isn't the model — it's the missing context.

Enterprises have systems of record for transactions (ERP) and relationships (CRM). No one built one for decisions.

As Igor Rikalo, President and COO of o9 Solutions, wrote in Forbes recently, your AI can't learn from the context, trade-offs, and human judgments that live only in meeting notes and email threads. It starts every planning cycle with partial amnesia, reworking problems your team already solved.

o9 Solutions calls this "decision memory" — a structured record of not just what was decided, but why, what was rejected, and what happened next. When you store the why alongside the what, your AI stops starting from zero every time. Intelligence compounds rather than resets.

Still searching for the right CRM?

Attio is the AI CRM that builds itself and adapts to how you work. With powerful AI automations and research agents, Attio transforms your GTM motion into a data-driven engine, from intelligent pipeline tracking to product-led growth.

Instead of clicking through records and reports manually, simply ask questions in natural language. Powered by Universal Context—a unified intelligence layer native to Attio—Ask Attio searches, updates, and creates with AI across your entire customer ecosystem.

Teams like Granola, Taskrabbit, and Snackpass didn't realize how much they needed a new CRM. Until they tried Attio.

Start now →

TRY THIS NOW

You can start building your company's decision memory in the next 10 minutes. The goal isn't to document everything — it's to capture the why behind your most critical decisions.

1. Identify one high-impact, high-frequency decision. This could be a weekly forecast override, a monthly budget reallocation, or a quarterly product roadmap change. Pick the one your team debates most often.

2. Create a lightweight decision log. Five fields, five minutes. That's it. Use a simple template in Notion or a shared doc — the goal is frictionless capture, not a formal process.

3. Run this prompt in Claude. Use it to generate a custom decision log template tailored to the specific decision you identified in step 1.

PROMPT OF THE WEEK

Act as a Chief of Staff specializing in Decision Intelligence.

My team needs to start building a "decision memory" to improve our AI's effectiveness over time. Our first focus is on a recurring decision: [Insert your specific decision, e.g., "overriding the quarterly sales forecast"].

Create a concise, structured markdown template for a "Decision Log" tailored to this specific decision. The template must be easy to fill out in under 5 minutes and capture the essential context for future AI analysis. Include fields for:

a) The Decision Made (a clear, one-sentence summary of what was decided).
b) The Primary Driver (the single biggest reason this decision was made).
c) Alternatives Considered (a brief list of options that were rejected and why).
d) Key Trade-offs (what was sacrificed or deprioritized to make this choice).
e) Expected Outcome & Metric (what success looks like and how it will be measured).
f) Confidence Score (on a scale of 1-10, how confident is the team in this decision).

After the template, provide 3 tips for making this decision log a sustainable habit within a busy leadership team.

THE EDGE

A 2025 Cloverpop survey found that only 6.5% of organizations consider their decision-making highly effective — and just 3% saw major improvements in speed from AI adoption. The gap between you and your competition isn't your AI model; it's your decision model. Start by logging your next three decisions. That's the whole move.

Two days. 4,000+ practitioners. Memories we'll never forget.

Thank you to every attendee, speaker, sponsor, and volunteer who made this year's conference something truly special. You ARE this community.

The conversation doesn't stop here. Join our Meetup group to stay connected and be the first to know about upcoming events in the Raleigh-Durham area. We'll see you at the next one! 🚀

👉 Join the All Things AI Meetup Group

Know a CTO frustrated that AI isn't delivering ROI? Share this with them with one line: "This is why our AI keeps starting from zero."

What's the one decision in your org that gets relitigated every quarter? Reply and tell me. I'm tracking the patterns.

— Mark

Your AI Sherpa, Mark R. Hinkle Publisher, The AIE Network Connect with me on LinkedIn Follow Me on Twitter

You wouldn't buy a smartphone and only use it to make phone calls. Yet most people treat AI assistants the same way—starting every conversation from scratch, repeating the same context, and manually pasting the same instructions over and over.

If you are using ChatGPT, Claude, Manus, or OpenClaw out of the box, you are leaving an enormous amount of productivity on the table. The real power of these tools unlocks when you stop treating them as generic chat interfaces and start codifying your workflows into reusable "skills" or "custom instructions."

But there is a trap here. If you build all your custom instructions inside ChatGPT, what happens when Claude releases a better model? Or when you want to use Manus to automate a complex browser task? Your hard-won expertise is locked inside a single platform.

The solution is remarkably simple: treat your AI skills like software code. By writing your skills once and storing them in a central GitHub repository, you create a portable vault of expertise that can be instantly deployed to any AI tool you use today—or whatever new platform emerges tomorrow.

Can’t make it to Durham in person? No problem!

All Things AI 2026 — happening March 23–24 — will be livestreamed on LinkedIn. Tune in for keynotes, sessions, and real conversations from 70+ AI leaders, builders, and practitioners. All from your screen.

👉 Follow us on LinkedIn to watch live →

AI LESSON

Build Your Portable AI Skill Vault

How to write, store, and deploy AI skills across ChatGPT, Claude, Manus, and OpenClaw.

The strategy below solves the biggest problem with AI adoption: the friction of context-setting. By centralizing your instructions, you ensure that every AI tool you use executes tasks with your exact preferences, formatting, and business logic. It sounds technical, but it is entirely accessible to non-developers.

The Agent Skills Open Standard: Why This Works

The reason a central vault is possible is the emergence of the Agent Skills open standard. This standard defines an AI skill as a simple text file, typically named SKILL.md.

This file contains two parts: a brief metadata section at the top (called YAML frontmatter) that tells the AI what the skill is called, and the actual markdown instructions that tell the AI what to do. Because these are just plain text files, they are universally readable. Claude Code, OpenClaw, and Manus all natively support this format. For ChatGPT, you simply copy the instruction text into your settings.

Step 1: Write the Skill with AI Assistance

You do not need to write these instructions from scratch. The most efficient way to create a skill is to have an AI write it for you.

Step 1: Open your preferred AI (like Claude or ChatGPT) and explain the task you want to automate. For example: "I want to create a skill that analyzes financial reports. It should always extract the top three risks, format the output as a markdown table, and use a professional, objective tone."

Step 2: Ask the AI to format this as an Agent Skill. Say: "Please write this as a SKILL.md file following the Agent Skills open standard, including the YAML frontmatter with a name and description."

Step 3: Review the generated text. Ensure the instructions are clear, step-by-step, and leave no room for ambiguity.

Time: 5–10 minutes per skill.

Step 2: Create Your GitHub Vault

GitHub is the perfect home for your skills. It provides version control (so you can see how a skill has evolved), acts as a secure backup, and makes it trivial to sync your skills across different machines.

Step 1: Create a free account at GitHub.com if you do not already have one.

Step 2: Create a new repository. Name it something like my-ai-skills. Set it to Private so your proprietary business logic remains confidential.

Step 3: Inside the repository, create a new folder for your skill (e.g., financial-analysis).

Step 4: Inside that folder, create a new file named SKILL.md and paste the text generated in the previous step. Commit the new file.

Time: 5 minutes for the initial setup.

Step 3: Deploy to Your AI Tools

Now that your skill is safely stored in GitHub, you can deploy it to whichever tool is best suited for the task at hand.

For ChatGPT: Open ChatGPT, navigate to Settings, and select Personalization. Click on Custom Instructions. Open your SKILL.md file in GitHub, copy the instruction text, and paste it into the "How would you like ChatGPT to respond?" field. If you use ChatGPT Projects (available on paid plans), you can paste these instructions directly into the Project settings to keep them scoped to a specific workspace.

For Claude Code and OpenClaw: Because these tools operate in your local environment, you simply clone your GitHub repository to your machine. For Claude Code, place the skills in the ~/.claude/skills/ directory. For OpenClaw, place them in your workspace/skills/ folder. Both tools will automatically detect the SKILL.md files and make them available via a slash command (e.g., /financial-analysis).

For Manus: Manus executes skills end-to-end in a secure cloud environment. You can upload your SKILL.md file directly into the Manus interface. When you trigger the skill, Manus reads the instructions and autonomously executes the workflow, leveraging its browser and code execution capabilities.

What the GitHub Vault Can't Do (Yet)

While this workflow provides excellent portability, it is not a perfect synchronization system. When you update a skill in GitHub, your local tools (Claude Code, OpenClaw) can pull the latest version automatically via Git. However, web-based interfaces like ChatGPT require you to manually copy and paste the updated text. Additionally, while the core instructions are portable, certain tool-specific capabilities (like Manus's native browser automation) may require slight tweaks to the instructions to function optimally across different platforms.

Getting Started Today

1. Right now: Identify one repetitive task you ask AI to do every week (e.g., formatting meeting notes, drafting emails).

2. This week: Ask Claude or ChatGPT to write a SKILL.md file for that specific task.

3. Next week: Create a private GitHub repository and upload your first skill.

4. When you have time: Clone that repository to your local machine to use with Claude Code or OpenClaw, or paste the instructions into a dedicated ChatGPT Project.

AI Extra Credit

March 26 | All Things AI Live — Weekly virtual meetup exploring practical AI applications for business professionals.

Catch up on the latest insights with the most recent Lunch & Learn on the @AllThingsAILive YouTube channel.

I appreciate your support.

Your AI Sherpa, Mark R. Hinkle Publisher, The AIE Network Connect with me on LinkedIn Follow Me on Twitter

This week, OpenClaw just had its "ChatGPT moment," sparking fears that AI models are becoming commodities, while Elon Musk unveiled plans for a massive $25 billion chip manufacturing facility in Texas.

Key Takeaways:

• OpenClaw's viral success as an open-source AI agent is raising concerns that foundational AI models are rapidly becoming commoditized.

• Elon Musk announced "Terafab," a $25 billion joint venture between Tesla, SpaceX, and xAI to build the largest chip manufacturing facility ever.

• OpenAI is reportedly planning to merge ChatGPT, Codex, and its Atlas browser into a single desktop "superapp" to streamline the user experience.

• The White House released a national AI legislative framework urging Congress to preempt state laws and establish a unified federal standard.

• Alibaba launched Accio Work, a plug-and-play enterprise AI agent platform designed to automate workflows for global businesses.

Join us as we untangle this week's happenings in AI!

NODES AI 2026 | APRIL 15, 2026

Neo4j's free online conference dedicated to Knowledge Graphs & GraphRAG, Graph Memory & Agents, and Graph + AI in Production. Seven hours of live sessions from leading AI practitioners.  👉Register for free here

THE BIG AI STORY

The OpenClaw "ChatGPT Moment": Are Foundational Models Becoming Commodities?

OpenClaw, an open-source personal AI assistant, has officially had its "ChatGPT moment." Released just four months ago, the project has vaulted into the top 10 most popular repositories on GitHub. Unlike traditional chatbots that simply answer questions, OpenClaw operates as an execution layer—it routes messages, calls APIs, and performs complex tasks like clearing inboxes, managing calendars, and even scouting eBay for deals. The heavy lifting of natural language processing is outsourced to whichever underlying LLM the user selects, effectively turning the foundational models into interchangeable utilities.

This breakout success is sparking serious concern among industry leaders that AI models are rapidly becoming commodities. If the primary interface for users becomes an open-source agent like OpenClaw, the value accrual shifts away from the massive, capital-intensive foundational models built by companies like OpenAI and Anthropic. Tencent has already recognized this shift, launching "ClawBot" over the weekend to integrate OpenClaw directly into WeChat, giving over a billion users instant access to task automation.

For business leaders, the rise of OpenClaw signals a critical pivot in enterprise AI strategy. The competitive advantage is no longer about having access to the smartest proprietary model; it is about deploying the most effective agentic workflows. Companies should begin evaluating how open-source execution layers can be integrated into their operations, reducing dependency on single-vendor ecosystems and lowering API costs by routing tasks to the most cost-effective models available.

5 QUICK HITS

Elon Musk Announces $25B "Terafab" Chip Factory

Elon Musk has unveiled plans for Terafab, a $25 billion joint venture between Tesla, SpaceX, and xAI to build the "largest chip manufacturing facility ever" in Austin, Texas. The facility aims to produce a terawatt of computing power annually, manufacturing specialized chips for terrestrial robotics (like Optimus) and highly durable chips for space-based data centers. For the enterprise sector, this massive capital expenditure underscores the severe, ongoing constraints in global AI compute capacity and the lengths to which tech giants will go to secure their own supply chains.

OpenAI Plans Desktop "Superapp"

OpenAI is reportedly planning to merge its ChatGPT application, Codex programming assistant, and Atlas browser into a single desktop "superapp". The consolidation aims to simplify the user experience and streamline internal engineering efforts, which executives noted had become fragmented across too many stacks. The company is also developing an "AI research intern" capable of automating complex, multi-day scientific tasks. This move signals OpenAI's ambition to own the entire desktop operating layer, positioning itself as the default interface for enterprise knowledge work.

White House Pushes to Preempt State AI Laws

The Trump administration has released a national AI legislative framework urging Congress to establish a unified federal standard for artificial intelligence. Crucially, the blueprint explicitly calls on lawmakers to preempt state-level AI regulations, arguing that a patchwork of local laws imposes undue burdens on innovation. For businesses navigating the complex regulatory landscape, federal preemption would provide much-needed clarity and reduce compliance costs, though it sets up a significant political battle over states' rights to govern technology development.

Alibaba Launches Accio Work for Global Enterprises

Alibaba International has launched Accio Work, a plug-and-play enterprise AI agent designed to equip businesses with an immediate, no-code taskforce. The platform enables business owners to automate everyday workflows, such as managing real-time documents, procurement, and cross-border logistics. This launch follows Alibaba's recent consolidation of its AI divisions and the introduction of its Wukong platform, highlighting the company's aggressive push to dominate the emerging "agent economy" and empower smaller businesses to operate with the efficiency of much larger enterprises.

OpenAI to Double Workforce to 8,000 by End of 2026

OpenAI plans to nearly double its workforce from 4,500 to 8,000 employees by the end of 2026. The massive hiring push is largely focused on expanding its enterprise sales and engineering teams to fend off growing competition from rivals like Anthropic and Google. For business leaders, this signals that OpenAI is aggressively scaling its enterprise support and custom deployment capabilities, moving beyond just offering API access to becoming a full-service enterprise software vendor.

Can’t make it to Durham in person? No problem!

All Things AI 2026 — happening March 23–24 — will be livestreamed on LinkedIn. Tune in for keynotes, sessions, and real conversations from 70+ AI leaders, builders, and practitioners. All from your screen.

👉 Follow us on LinkedIn to watch live →

3 AI TOOLS

• murmur — An iOS app that allows users to rehearse difficult real-life phone calls—like negotiating rent or prepping for interviews—with a realistic AI before making the actual call.

• Claude Cowork Projects — Anthropic's new workspace tool that organizes tasks, context, and files in one place, allowing users to control Mac-based sandboxed sessions directly from their phones via Claude Dispatch.

• Design Agent by Lokuma — A specialized design tool built specifically for creating interfaces and visual assets for AI agents like OpenClaw, Codex, and Claude.

AI EXTRA READ

OpenClaw: A Technical Guide for the Business Executive (10-min read)

A comprehensive breakdown of how the viral OpenClaw agent actually works under the hood. This guide explains the architecture, security model, and API routing capabilities in plain language, helping leaders understand why the execution layer is becoming more valuable than the foundational models themselves.

If you only do one thing this week, test an open-source agent like OpenClaw on a routine workflow to see how it compares to your proprietary vendor solutions.

Your AI Sherpa, Mark R. Hinkle Publisher, The AIE Network Connect with me on LinkedIn Follow Me on Twitter