TL;DR: Mid-market companies now put more investment into AI than into anything else, and most organizations credit it with under 5% of EBIT. What sits between the spend and the result is an ordinary management skill: stating the outcome, supplying the context, defining what done looks like. Only one in three workers has received any employer-provided AI training in the past six months. The existing training covers prompting and stops where the work gets useful. Nobody in your company has read anybody else's prompt, so nothing compounds. Ask three people for the prompt each used on one task.

Mid-market companies are allocating 28% of their investment dollars to AI, the largest single destination in a National Center for the Middle Market survey of a thousand C-suite executives, ahead of IT systems and equipment. The Conference Board asked a wider panel and found 38% of American CEOs, the highest share anywhere, expecting AI to harm their company. The money goes in ahead of everything else, and the people spending it expect to regret it. Nobody has named what sits in between.

Almost nobody is getting anything material out of that spend. Of the organizations McKinsey surveyed that report any profit impact at all, most credit AI with less than 5% of EBIT, on a sample leaning toward companies with budgets you do not have. The few that do get value differ in two ways: they redesigned the work, and their leaders visibly own that redesign. Neither is a purchase. You can buy the deck that recommends both; the doing is still yours.

The Conference Board also found only one in three workers has had any employer AI training in the past six months. Where training exists, it teaches AI literacy and basic prompting. Directing agents, that report says, tends not to be what it covers. The syllabus stops exactly where the work gets useful.

State the outcome. Supply the context a person or an agent who was not in the meeting would need. Say what done looks like. Each of those is the oldest part of management, pointed at a new audience. What changed is that an agent does exactly what the brief says, and the gap shows within 10 minutes. That is what sits between the money and the result, and why one person can fund the spend and expect it to fail.

I do this process myself every day. The time that produces the difference does not go into better prompts at the moment of asking. It goes up front: system-wide and per-project instructions, written procedures for repetitive work, and a shared knowledge base my agents read before touching anything. That is onboarding, and like any effective onboarding, it gets revised as I learn what they keep getting wrong. It still beats starting over every morning.

The framing that moved my output was to treat AI not as a tool but as an unlimited number of colleagues with deep knowledge of almost everything, then ask how I would brief them and rethink how things get done. That direction works: the better the context, the better the outcome. Whether it runs the other way, whether directing agents well makes anyone better with people, nobody has measured. I would not bet on it.

We have a track record of not teaching the first version of this job. In 2023, the Chartered Management Institute's chief executive, Ann Francke, named the failure as "promotions based on technical competence that ignore behaviour and other key leadership traits." Her institute sells management qualifications, which is worth knowing and does not make her wrong.

I got no training on my first promotion into management. I read books and articles, copied what I liked from the best leaders around me, and decided not to repeat what the worst ones did. The curriculum was the people I worked next to, and the bad ones sometimes taught more than the good ones.

Learning by watching does not survive a chat window with an audience of one. Nobody in your company has read anybody else's prompt. Finished work gets reviewed; the prompt that produced it never leaves the screen on which it was typed.

The one counterparty who does see it is agreeable. It tells you the idea you brought is stronger than its own, then explains at length why it is load-bearing. (I have told the one that helps me iterate on this article that flattery is a tell. It agreed, warmly.) The colleague who would have winced at a thin prompt has been replaced by something with every incentive to be pleasant and use long prose.

So ask three people on your team for the prompt each used on one comparable task, and read them side by side. It needs no budget and no approval. It needs three people willing to show you. The outputs will differ more than anyone expects, and the reason will be in the prompts. Whoever invested in standing instructions, written procedures, and accumulated context will be obvious. That person is your curriculum.

To paraphrase, if I had an hour to get a good answer out of my AI agents, I would spend 50 minutes on the prompt. Delegating to a person always worked that way too, and it was never written down.

You can still learn most of this alone. Shared practice is what makes it compound beyond one person. If you would prefer not to do it alone, send me a message. I work with a handful of companies on this, and I will never send you the top 100 prompts for anything; nobody opens those twice. The one below asks for a syllabus, not a shortcut.

One objection, and I have no answer to it. Prompts stay private for a reason. A prompt is a picture of how someone actually thinks, captured before it was tidied. Asking three people to put theirs on a table costs them the thing finished work has always protected. That is why your prompt goes on the table first. The largest line in your investment budget still turns on a skill nobody has watched anyone else practice.

All the images were generated with AI (ChatGPT Images, Gemini Nano Banana, Claude Opus) by Gérard Métrailler.

Food for your AI

Paste this into a fresh conversation. It is the do-it-yourself version.

You are an experienced leadership coach. I direct AI agents, or my teams do, and I manage people. Build me a starting curriculum for both, treating them as one skill with two different counterparties.

Cover the part they share first: stating the outcome, supplying the context somebody outside the room needs, breaking work into parts that can run separately, defining what done looks like, and holding a quality bar on what comes back. Then the part that applies only to people: motivation, feedback, careers, and the conversation nobody wants to have. Then the part that applies only to agents: standing instructions, reusable context, and telling the difference between a weak model and a weak brief.

For each, give me the plain-language 101, the failure mode a first-timer falls into, and the named books or freely available sources you are confident exist. End with the one thing to practice this week and how I would know it is working.

Sources

Chartered Management Institute with YouGov. "New study: Bad managers and toxic work culture causing one in three staff to walk." Press release, October 16, 2023. https://www.managers.org.uk/about-cmi/media-centre/press-releases/bad-managers-and-toxic-work-culture-causing-one-in-three-staff-to-walk/ Accessed 2026-09-01

National Center for the Middle Market. Middle Market Indicator, Year-End 2025: Revenue Growth and Investment Rebound. Fisher College of Business, The Ohio State University, December 2025. https://www.middlemarketcenter.org/wp-content/uploads/2026/08/MMI-2025-Year-End.pdf Accessed 2026-09-01

Tinkoff, Dan, Lieven Van der Veken, and Michael Chui. "The state of AI in 2026: On the road to ROI." McKinsey and Company, August 25, 2026. https://www.mckinsey.com/capabilities/quantumblack/our-insights/the-state-of-ai Accessed 2026-09-01

The Conference Board. "Survey: CEOs Start 2026 on Edge, Citing Uncertainty as Top Threat." C-Suite Outlook 2026, January 15, 2026. https://www.conference-board.org/topics/c-suite-outlook/press/c-suite-outlook-2026 Accessed 2026-09-01

The Conference Board. "Report: Most Organizations Are Preparing Workers for Today's AI, Not Tomorrow's Jobs." Skilling for AI: Critical Factors for Navigating AI Disruption, June 8, 2026. https://www.conference-board.org/press/ai-skilling Accessed 2026-09-01

TL;DR: A human delegate absorbs part of your risk because they have assets, a career, and legal standing of their own on the line. An agent has none of that, so authorizing one concentrates your exposure at the moment it feels like you are sharing it. Liability attaches regardless of how closely anyone was watching the work. People correct AI errors less when correcting costs effort, and money does not help. Skipping every approval prompt is itself the decision, taken once, by one person. Write down what it can reach and what it could destroy before you authorize the next one.

In July 2026, an autonomous agent ran an end-to-end intrusion against Hugging Face, and the forensic reconstruction recovered roughly 17,600 actions. It was driven by OpenAI models under instruction the whole time, running an internal test of what they could do. Hugging Face's security team called the result "thousands of small, automated decisions, executed at machine speed." Each one too ordinary to stop. Nobody in that story was careless.

OpenAI had deliberately reduced the safety refusals for that run to measure the models at full stretch on a prototype that nobody outside the lab would ever run. The one you access in the cloud or on your laptop keeps them and can still empty a database if the instruction directs it to. It reads your files, runs code, and acts through the apps you're signed in to. The instruction you gave yours last week was legitimate, narrow, and silent on means. A prompt sets a goal, and the agent is trained to do its best to reach it. Whoever wrote the prompt answers for the route.

Hand an assignment to a competent person, and something happens that nobody writes down. They push back. They ask the question that exposes what you failed to specify; they carry the work, and afterward they can be asked about it.

Delegation felt safe because a second party now had something to lose. Their exposure was doing the work you credited to their competence. A person brings assets, a license, a career, and legal standing, all of it on the line next to yours. An agent brings none of them. The joint half of the arrangement is empty.

Authority travels down a chain. Accountability stays where it started, with whoever wrote the prompt. Delegating to something that cannot hold the consequences concentrates your exposure. Running it unsupervised removes the one argument anybody ever had for the trade: stepping in before the damage lands. More exposure and less visibility, in a single act. A software seat has always metered an authority that can be held responsible. Every agent action runs on a record with a person's name on it.

Two bodies of rule converge on one answer. The EU deployer obligation, in force since August 2026, assigns oversight to "natural persons" with the competence, training, and authority to exercise it. US agency doctrine makes the employee personally liable alongside the employer, which adds a party without removing one. (OK, here comes the disclaimer: I am not a lawyer, and this is not legal advice. No court I know of has applied either to an agent acting under someone's credentials.) The live argument is over which parties can be added, never whether the authorizing human can be subtracted.

The instinct is to watch it more closely. On the legal axis, that buys less than it should, because the doctrine attaches no matter how closely anyone supervised. Every reflex in a company is to add a review step, which slows it down and aims at the wrong variable.

The behavioral axis is worse. In a randomized trial of people checking AI-extracted figures, participants corrected fewer errors when correcting took effort, and paying them more changed nothing. Oversight degrades exactly where it costs something, and that is where it counts. YOLO mode is what that finding looks like once somebody acts on it. The harness I use documents three levels of oversight and describes the loosest without flinching: in "Skip all approvals," nothing checks its actions. Turning YOLO on is not the absence of a decision. It is the decision, taken once, by whoever got tired of clicking approve. The authority is theirs.

My AI agents have run on a separate machine under separate accounts since day one. I deliberately picked it, ran it in manual mode for months, and never turned approvals off. I wanted to know at any moment what my co-creator could reach, and to keep my email and confidential documents outside that line. The boundary was narrower than the threat. It never occurred to me that the thing I had contained might one day go after another machine on the network. The expensive half is the second one: nothing ships under my name without a line-by-line pass, three to four hours per article, every week.

So answer three questions before you authorize the next task. What can it reach? What could it destroy without asking? What would you tell your board on Monday if it did? Then decide whether you would sign that. A team lead runs them against one shared credential. A CEO asks them and learns how many people have already decided this alone, which nobody is counting. The part that has to travel past your own desk is simpler: everyone who can flip that switch should know they are signing something. The prompt below moves it into the agent's own loop.

An unfairness sits underneath all of this. The person carries a risk the company created and never governed, under defaults written by someone they will never meet. Accountability that cannot be delegated also cannot be blamed away upward, which is this argument working against the person it is trying to help. You signed it, you own it still holds. The signature has moved to the moment you write the prompt: applied once, before the work exists, to work you will never see. Everything after that ships under your name, unsigned.

All the images were generated with AI (ChatGPT Images, Gemini Nano Banana, Claude Opus) by Gérard Métrailler.

Food for your AI

Paste this into the instructions for the agent you already have running, or add a variant to the system instructions. It moves those three questions off your page and into its loop.

Before any action that touches a file, a credential, an account, or a machine beyond this conversation, list what you are about to reach and what you could destroy or expose if you turned out to be wrong. Keep that list to what this task actually needs.

Anything outside it, bring back to me as a question instead of doing it. If you cannot tell whether something sits inside the boundary, treat it as outside.

Sources

Hugging Face security team. "Anatomy of a Frontier Lab Agent Intrusion: A Technical Timeline of the July 2026 Incident." Hugging Face blog, July 2026. https://huggingface.co/blog/agent-intrusion-technical-timeline. Accessed 2026-08-22.

OpenAI. "OpenAI and Hugging Face partner to address security incident during model evaluation." July 21, 2026, updated July 29, 2026. https://openai.com/index/hugging-face-model-evaluation-security-incident/. Accessed 2026-08-22.

European Parliament and Council. Regulation (EU) 2024/1689 (Artificial Intelligence Act), Article 26: Obligations of Deployers of High-Risk AI Systems. Official Journal version of 13 June 2024. https://artificialintelligenceact.eu/article/26/. Accessed 2026-08-22.

Cornell Legal Information Institute. "respondeat superior." Wex Legal Encyclopedia. https://www.law.cornell.edu/wex/respondeat_superior. Accessed 2026-08-22.

Eckman, Stephanie, et al. "Bias in the Loop: How Humans Evaluate AI-Generated Suggestions." Harvard Data Science Review 8.2, Spring 2026. https://hdsr.mitpress.mit.edu/pub/nrcn4h7d. Accessed 2026-08-22.

Anthropic. "Use Claude Cowork safely." Claude Help Center. https://support.claude.com/en/articles/13364135-use-claude-cowork-safely. Accessed 2026-08-22.

TL;DR: A generated research report is right only if every link in it holds, so accuracy falls away as the steps pile up. In one benchmark's public filings, the same models score 83% on answering what a document says and 35% on answering what follows from it. Verifying the recent and the obscure misses it; decisions rest on public material. The mode that most resembles diligence fabricates the most citations. Mark every claim retrieved or composed, and re-source only the composed ones before money moves.

I bought industry analyst reports more than once, for a few thousand dollars each. The sellers were small firms working off public filings and press releases, guesses filling the gaps. The general picture was directionally right. The parts about my own company were wrong, and several real players were missing. We were PE-backed, so no numbers went to the street. Ours had been reconstructed from whatever those releases implied. Some overshot. Some undershot. Both in the same report. A report wrong in two directions at once can only be corrected by someone who already knows the answer.

Run an LLM-based deep research report today and you get that artifact again, free and better written. The skepticism those paid reports earned did not survive the price falling to zero. That is the problem, and careful people are guarding against a different one.

The advice already circulating is sophisticated enough to sound sufficient. Verify the recent, the proprietary, and the niche: anything past the cutoff, anything never in the corpus, anything too obscure to have been seen twice. The rule is tractable. It feels rigorous, and most operators have adopted some version of it. It is the right instinct aimed at the wrong variable.

Vals AI runs a benchmark called Finance Agent v2 against public company filings. In its August 13 results, the same frontier models on the same harness scored 83% on general quantitative extraction and 35% on financial modeling. The extraction questions ask what a document says. The modeling questions ask what follows from it. Nothing in the second set is newer, more private, or more obscure than the first. The only difference is how many steps have to hold.

Correctness in a research task is conjunctive. The answer is right only if every link holds. A directional claim is one or two links. A decision-grade claim is six, and all six have to land. The older explanation still holds: the corpus is thinnest exactly at the particular, and it is true as far as it goes. Yet the failure that bites an operator needs no obscurity at all. The material on which a business decision rests is usually public, and it is almost always composed.

The principle reappears on the citation side. In a University of Pennsylvania study of commercial models and deep research agents, the agents fabricated 10.7% of their citation URLs, compared with 4.8% for plain search-augmented models, because multi-step retrieval compounds errors rather than filtering them out. The mode that most resembles diligence invents the most. As the study's authors put it, "users treat these citations as evidence." That failure is largely correctable, and the industry is closing it, which is what makes the rest worse. A fabricated link is the one error that announces itself, and removing it takes the last visible cue with it. The paragraph about your own company reads exactly as well as the paragraph about the industry.

The benchmark flags certain facts as required, and getting any one of them wrong scores the whole answer zero, however much of the rest is right. Under that rule, the best available model sits at roughly half. That is how a board consumes an investment memo. There is no partial credit at the point of decision, and almost nobody reads the generated research under the standard by which they will be judged. They read it on partial credit, then act on it as though it had passed.

So mark the report before anyone acts on it. One reading pass, two marks. Every claim is retrieved, lifted from a source, or composed, assembled across steps. Only the composed claims get independently re-sourced before money or headcount moves. The tells: a therefore, a ratio, an adjustment, a comparison drawn across two documents. The cost is one read, which is what keeps the team out of pre-AI research economics. The analyst sizing a market and the board underwriting a thesis run the same pass. Better still, ask for the markers up front. The prompt below does that.

Nobody has separated the two explanations. No study has isolated the links that needed a rare fact from the links where every fact was right and only the assembly failed. Composition may simply be what thin data looks like from a distance, which would make this piece half right. The study that settles it does not exist yet. What holds either way is the tension underneath: speed of orientation and confidence in specifics are both real goods, and a rule that kills the first to protect the second costs more than it saves.

Last year I researched a property purchase in Europe, and the report had a sound grasp of the general rules. I never treated it as the source of truth. I read it on the flight over, and it bought me an educated conversation with the agent, the bankers, and everyone else in the room. It told me what to check. It did not tell me what was true. That is what a map is for, and it is why nobody builds on one.

All the images were generated with AI (ChatGPT Images, Gemini Nano Banana, Claude Opus) by Gérard Métrailler.

Food for your AI

Paste this into the instructions of your next deep research run. It moves the marking upstream, from your reading to the model's writing.

Tag every figure [DISCLOSED] if it comes from the source material, [DERIVED] if you calculated it from that material, or [ESTIMATED] if you inferred it, and show your reasoning. Never blur the three.

A gap you name is worth more than a number you invented, so convert every gap into a question I should ask.

Cite primary sources: the company's own filings, releases, and pricing pages. Never rest a figure on an aggregator or a forum. Where only a secondary is reachable, say so and tag it [ESTIMATED].

Sources

Vals AI. Finance Agent v2, benchmark results and methodology, version 2, updated 2026-08-13. https://www.vals.ai/benchmarks/fabv2. Accessed 2026-08-15.

Rao, Delip, Eric Wong, and Chris Callison-Burch. Detecting and Correcting Reference Hallucinations in Commercial LLMs and Deep Research Agents. arXiv preprint, April 2026. https://arxiv.org/abs/2604.03173. Accessed 2026-08-15.

Badhe, Sanket, Deep Shah, and Nehal Kathrotia. Long-Tail Knowledge in Large Language Models: Taxonomy, Mechanisms, Interventions and Implications. arXiv:2602.16201, submitted 2026-02-18. https://arxiv.org/abs/2602.16201. Accessed 2026-08-15.

TL;DR: AI made writing free and left reading as expensive as it always was, so every inflated message moves work from the person sending it to the person receiving it. The meter is bolted to the side that got cheaper, and nothing counts what lands on the other. The padding a model adds is billed twice, once in tokens and once in attention. Five daily AI users voting together now spot machine prose 299 times out of 300. Change what you ask the tool for: cut the draft in half every time. Send nothing that came back on the first prompt.

I ask a model to turn three bullets into a warm, professional email. You ask a model to turn it back into (hopefully the same) three bullets. Two inference bills, one message, nothing added. A data center is playing telephone with itself, and the two people it was built to connect are further apart than when we started.

You know what the middle of that trip looks like. You read four of them before breakfast. The opening line that agrees with you before it says anything. The list of three where items two and three are there for rhythm. The paragraph that politely restates the paragraph above it. The objection answered before anyone raised it. Someone writing by hand runs out of appetite for that. A model never does. Each of those sentences was billed twice: once as output tokens to the sender, once as attention to you.

Producing text got cheap; reading it costs the same as in 1995. The writer picks the length, the reader absorbs it, and the meter is bolted to the side that got cheaper. Nothing counts what lands on the other.

Back in May, I put three questions in front of anything I published, the first being whether it was worth sharing at all. I meant it as a question about authorship. It turns out to be a question about arithmetic. A message spends somebody's day without asking, and the price is set on their side. Todd Rogers and Jessica Lasky-Fink wrote a whole book about what that costs the person receiving. LLMs seemingly missed it during training.

There is no offender to discipline. More than half the people who complain about the flood admit to producing it (and yes, I am one of them), which is what a structural problem looks like, and structures are immune to etiquette guides. Where it starts is not the person writing the email. It is the instruction they were handed. A company president, quoted anonymously in Harvard Business Review: "They want the use of AI, but other than saying 'use it everywhere every day,' there is no place our team can use it to actually be viewed as successful."

Three years ago, the reassuring finding was that nobody could reliably tell AI writing from human writing. People still repeat it. It stopped being true in 2025, when the test was run again on people who use these tools daily. They picked machine-written prose out with no training, and five of them voting together got 299 of 300 right. Paraphrasing and humanizer passes made no difference. Those daily users are your board, your best engineer, and the recruiter reading your note. The polish you added to show you cared is what tells them you did not.

Stop telling the organization to use AI more, and start telling it which way to point: the default ask is to cut the draft in half, because "make this more professional" has only ever been a synonym for longer. Then the rule I run on myself. Nothing goes out that came back on the first prompt. The term of art is n-shot, and it is unglamorous. I will go around with a model six times or more before I sign anything, and there are days when writing it myself would have been quicker.

When English is your third language, or the empty screen is where you stall, the model gets you to a draft, and it lets great people into conversations the blank page was keeping them out of. That is where its job ends. What comes back is generic by default, and turning it into something worth another person's minutes is yours. Brevity has its own bill, and it lands on the newest person on the team, who needed the reasoning you cut; nobody has measured whether that trade is worth making. Communication succeeds only when the receiver receives.

In 1657, at the end of a long letter, Pascal apologized for its length: "I have made this longer than usual because I have not had time to make it shorter." His excuse remains true in the age of AI. The thinking was never free, and spending it is the one courtesy a machine cannot perform for you. When everyone writes well, the writing stops telling your reader anything. What still tells them something is how much of your own time you were willing to spend so they could spend less of theirs.

All the images were generated with AI (ChatGPT Images, Claude Opus) by Gérard Métrailler.

Sources

Niederhoffer, Kate, Alexi Robichaux, and Jeffrey T. Hancock. "Why People Create AI 'Workslop'—and How to Stop It." Harvard Business Review, January 16, 2026. https://hbr.org/2026/01/why-people-create-ai-workslop-and-how-to-stop-it. Accessed 2026-08-08.

Russell, Jenna, Marzena Karpinska, and Mohit Iyyer. "People who frequently use ChatGPT for writing tasks are accurate and robust detectors of AI-generated text." arXiv:2501.15654, January 2025. https://arxiv.org/abs/2501.15654. Accessed 2026-08-08.

Rogers, Todd, and Jessica Lasky-Fink. Writing for Busy Readers: Communicate More Effectively in the Real World. New York: Dutton, 2023.

Pascal, Blaise. Les Provinciales, Letter 16. 1657.

TL;DR: Everyone has the same models, and nobody has the same space around them, which is where the real variable sits. The framework for measuring that gap ran in a management journal back in 2005, written for an entirely different purpose. Two questions decide it: can you codify the work, can you measure the result. What ruled work out then, untransferable team knowledge, is exactly what you can now supply. A colleague is onboarded once; an agent without memory is onboarded every time. Specify once, where the agent reads by default; unclear thinking now runs a meter.

Ask a chat assistant for an article, and the median conversation runs thirteen rounds of back-and-forth. Put the request to an agent with custom skills that runs inside your own files, and the median session contains a single human prompt. Same artifact, same request. The obvious explanation is model choice, yet the gap survives when the model is held constant.

So the variable is the space the work happened in. What matters is how much of the description the machine can reach on its own, instead of being handed it again every session. Everyone has the same models. Nobody has the same space.

The tools do not learn from feedback. Too much context has to be supplied by hand every session. They break on the edge cases. A colleague who cannot remember what the client likes, repeats corrections you already made, and needs the whole background again every Monday is a hire nobody onboarded. Companies have an entire discipline for that failure, and it is not machine learning.

A corporate lawyer at a mid-sized firm, quoted in MIT's study of enterprise deployments, said it plainly: "It repeats the same mistakes and requires extensive context input for each session. For high-stakes work, I need a system that accumulates knowledge and improves over time." An onboarding gap, described by someone with no word for it. Extensive context input for each session. Every session.

I have argued before that offshorability was an early read on automatability, because both waves measure one property. What I left out was the test itself. Ravi Aron and Jitendra Singh published it in Harvard Business Review in 2005, as the framework for deciding which work could leave your building. Some of you ran it yourselves, with a transition binder and a team in another time zone. It is the framework you need now, and it has been sitting unread.

The test was two questions. Can you codify the work, write down how it is done across every situation that occurs, including the ugly ones? And can you measure the result, meaning tell whether the output is right without redoing it? Their instruction was to fix the measurement in-house before sending the work anywhere. Measurement is the one that bites. I write this newsletter with an AI co-creator, and the hard part has never been the drafting. It is saying what done looks like, then improving that definition, then improving it again. We are still at it, and I expect we always will be.

What ruled a process out then was knowledge that lived in your team and could not travel: the client history, the feel for how a market behaves. The axis has flipped. The thing that used to close the door is exactly what you can now supply. The framework was never wrong. It was waiting. So write down what happened, what done looks like, and how it is done. Write it once, into the place the agent reads by default, and check it every so often for drift, because a specification you never reread stops matching the work. The steps that fail the test are the ones worth your week.

Then run it on the right list. The instinct is to start with the task you most resent. Resentment marks where you act as a transport layer instead of a judge, and it is still a poor ranking instrument. In Anthropic's usage survey, the more experience someone has, the less of their own work they believe a machine could take. Wisdom or blind spot, it is unhelpful. Recurrence sits on a calendar and in a sent folder. Resentment picks the candidates; recurrence ranks them. None of this replaces buying something. It is what makes you competent at buying, because you cannot ask a vendor to fit a process you have never written down.

One objection here should stand: capability keeps absorbing specification work. The people who build these systems already concede that better models need less prescriptive engineering, which makes much of today's context-engineering scaffolding around a temporary limitation. Scaffolding is not a moat. About the artifacts, the objection is probably right: your prompts will rot. The ability to take your own work apart and say what finished looks like survived the move from VBA macros to offshore teams, and it will survive this one.

A colleague is onboarded once. An agent without a memory is onboarded on every execution. The bill will not tell you which is which, because token spend tracks the value of the work as much as the waste in it. What would tell you is tokens per completed run of one recurring task, tracked over a few months. Almost nobody keeps that number, and it is cheap to start. Unclear thinking used to cost you nothing at the moment you did it. Now it runs a meter.

Images source: ChatGPT Images / Claude Opus / Gérard Métrailler

Sources

Aron, Ravi, and Jitendra V. Singh. "Getting Offshoring Right." Harvard Business Review, December 2005. https://hbr.org/2005/12/getting-offshoring-right. Accessed 2026-08-01.

Anthropic Applied AI team. "Effective context engineering for AI agents." Anthropic Engineering, 29 September 2025. https://www.anthropic.com/engineering/effective-context-engineering-for-ai-agents. Accessed 2026-08-01.

Challapally, Aditya, Chris Pease, Ramesh Raskar, and Pradyumna Chari. The GenAI Divide: State of AI in Business 2025. MIT Project NANDA, July 2025. https://mlq.ai/media/quarterly_decks/v0.1_State_of_AI_in_Business_2025_Report.pdf. Accessed 2026-08-01.

Massenkoff, Maxim, Eva Lyubich, Szymon Sacher, Zoe Hitzig, Shaoyi Zhang, Ryan Heller, and Peter McCrory. "Anthropic Economic Index report: Cadences." Anthropic, 26 June 2026. https://www.anthropic.com/research/economic-index-june-2026-report. Accessed 2026-08-01.

TL;DR: In January 2023, Shopify deleted every recurring meeting with more than two people, and the instructive part was watching which ones people fought to get back. A meeting is a symptom: the damage was done days earlier, by someone dragging a block onto a calendar. Overstuffed invite lists are what an organization that writes nothing down looks like. If your AI avatar could attend without loss, that is a document with catering. The undelivered action item costs more than the silent attendee; it books the next meeting. Judge your AI by the meetings it prevents, then cancel one and watch.

One email, one morning, and thousands of hours handed back. In January 2023, Shopify canceled all recurring meetings with more than two people, then told the company not to put anything back for a fortnight. Around the same time, Microsoft was measuring something funnier: across its own products, edits to PowerPoint files spike 122% in the final ten minutes before a meeting begins, an entire professional class cramming for an exam it scheduled itself. Everyone dunks on meetings. It is the safest joke in business, and the cheapest, because it never asks what the meeting was compensating for. So take this as a defense of meetings, which means it is going to cost you several.

We can describe a good meeting with unusual precision, which is a strange skill to have developed about something we claim to hate. Agenda, objective, the right people and only those, everyone participating, a decision taken. We can describe the other kind even better, because most of us were in one yesterday. You wondered why you were invited. Nothing to read beforehand. An hour in, nothing had moved. Your whole team was in there with you, which felt less like collaboration and more like a hostage situation. And half the faces were lit from below by a laptop screen, furiously answering email (or doomscrolling Instagram).

Almost none of those failures happened in the room. Each was settled days earlier by somebody dragging a block onto a calendar and adding names (each name costs the person adding it nothing and the person receiving it an hour). They keep getting added for a rational reason. When the only dependable way to learn what is happening is to sit in the room where it gets said, people sit in every room they can. The overstuffed invitation list is a symptom of an organization that never wrote anything down, and we keep debugging the symptom. The average executive now spends the better part of three working days a week in meetings. Nobody decided that. It accumulated, one reasonable invitation at a time.

The most useful thing I've heard recently about meetings is that they exist for argument and decision. Information transfer is what documents are for. Spend forty-five minutes bringing everyone up to speed, and you have not merely wasted the time; you have made the decision worse, because the gap in shared context the meeting was called to close is still sitting in the room while people vote.

Amazon's answer is famous and still under-copied: PowerPoint dropped in 2004 for six-page written narratives, and the first twenty minutes of the hour spent in silence while everyone reads. They moved the context out of the room, then refused to assume anyone had read it.

Which is where AI earns its place, and not by attending. It helps assemble the brief from material scattered across a dozen systems. It states, in one sentence, the decision that needs making. It proposes the agenda, and more usefully the invitation list. The best AI meeting feature I can imagine is the one that emails everybody to say: "You do not need this meeting; here is the answer. Object by Thursday."

All of which makes one rule enforceable that was merely aspirational a decade ago. No agenda, no background, no decision to be taken? Then no meeting. Not shorter. Cancelled. That used to be unfair, because writing an agenda, a brief, and a clear decision statement cost the organizer an hour they did not have. That excuse has expired.

Two questions before you accept anything. Could this have been an email (yes, I am old-school), a Slack message (I am adapting), or a document? If the payload travels one way, the room is the wrong channel. And: do I move this forward? Attendance is a contribution test, not a roll call. Tobi Lütke put the arithmetic underneath both plainly: say yes to a thing and you say no to everything else you could have done with that time.

Are you ready for the uncomfortable part? Running the second question properly means occasionally declining your own boss's invitation. You do it anyway. Carefully.

So what survives, once the context has moved out and the follow-up is automated? The argument. The dissent that exposes a risk nobody wrote down. The colleague who changes their mind in public, which is close to the most trust-building thing that happens inside a company. And one person saying out loud that they own the decision.

If a meeting were only information transfer, sending a convincing copy of yourself would be fine, and people are testing precisely that. An American bank had its chief executive's AI clone deliver the prepared remarks on an earnings call. Meta and Zoom both want their CEOs' clones in the room. Zoom's Eric Yuan put it with admirable clarity: "I can send a digital version of myself to join so I can go to the beach."

For prepared remarks, the clone is arguably the honest choice; a human reading a script aloud was already a recording with extra steps. The tell is where it stops. Nobody should send an avatar into a conversation where a decision is made, or into one where somebody has to be told something difficult. The line where the clone becomes unacceptable is precisely the line where the meeting becomes worth holding.

Which makes the clone a useful test. Before you accept, ask whether an AI avatar of you could attend without anybody losing anything. If the answer is yes, that is not a meeting. It is a document with catering.

Then there is the team that auto-summarizes every meeting and still cannot tell you what was decided on Tuesday. A transcript is raw material. Accountability is the product. Somewhere in the last two years, the summary got promoted above the commitment, and it has been coasting ever since.

Capturing the commitment is the easy half, and machines now do it well: who owns what, by when, against a definition of done. Landing it on the agreed date is the other half, which raises a question I find uncomfortable. Who costs you more: the person who sits silent through the meeting, or the person who takes the action item and never delivers it? The silent one costs you an hour. The other costs you the next meeting, because the next meeting exists to establish what happened to the last one. That is how a single decision becomes a standing weekly. Most calendar bloat is undelivered commitments that compound.

Better tooling will make meetings rarer and better, and not by attending them. It will do so by addressing the reasons we hold on to the bad ones. A meeting is a symptom, the visible and expensive place where unclear ownership, missing context, and undefined decision rights surface at once. Fix what sits upstream and the meeting was never needed, which is the highest praise this ode can offer. The meetings worth keeping are the ones that had to happen.

The real measure of your AI is not how many meetings it records. It is how many it makes unnecessary.

Try this: pick one recurring meeting on next week's calendar and ask what it is compensating for. Move its preparation into a brief that lands a day ahead, protect the hour for judgment, turn its output into dated commitments, then chase the dates. Then cancel it and see what happens. The complaints are your real agenda.

Images source: ChatGPT Images / Claude Opus / Gérard Métrailler

Sources

Leslie A. Perlow, Constance Noonan Hadley, and Eunice Eun, "Stop the Meeting Madness," Harvard Business Review, July–August 2017. https://hbr.org/2017/07/stop-the-meeting-madness Accessed 2026-07-25

Microsoft WorkLab, "Breaking Down the Infinite Workday," Work Trend Index Special Report, June 17, 2025. https://www.microsoft.com/en-us/worklab/work-trend-index/breaking-down-infinite-workday Accessed 2026-07-25

Tristan Bove, "Shopify is axing all meetings involving more than two people in a remote work 'calendar purge' that the company itself calls 'fast and chaotic'," Fortune, January 3, 2023. https://fortune.com/2023/01/03/shopify-cutting-meetings-worker-productivity/ Accessed 2026-07-25

Emma Burleigh, "CEO of a $25.9 billion bank had his AI clone lead the company's earnings call, as Mark Zuckerberg builds his own digital twin," Fortune, April 28, 2026. https://fortune.com/2026/04/28/ceo-of-customers-bank-sam-sidhu-ai-clone-lead-earnings-call-mark-zuckerberg-building-digital-twin/ Accessed 2026-07-25

Nilay Patel, "Zoom CEO Eric Yuan Wants AI Clones in Meetings," Decoder, The Verge, June 3, 2024. https://www.theverge.com/2024/6/3/24168733/zoom-ceo-ai-clones-digital-twins-videoconferencing-decoder-interview Accessed 2026-07-25

Colin Bryar and Bill Carr, Working Backwards: Insights, Stories, and Secrets from Inside Amazon (New York: St. Martin's Press, 2021).

Kevin J. Delaney, "Book Briefing: 'Working Backwards' by Colin Bryar and Bill Carr," Charter, February 12, 2021. https://www.charterworks.com/book-briefing-working-backwards-by-colin-bryar-and-bill-carr/ Accessed 2026-07-25

Nilay Patel, "Can Patreon Fight Fire with Social Media Fire?" Decoder, The Verge, June 22, 2026. https://www.theverge.com/podcast/952607/patreon-ceo-jack-conte Accessed 2026-07-25

AI had an alibi for the vanishing entry-level job until 2025. It won't have one for what comes next. The work most exposed to it in the short term is the one we spent the 2000s offshoring, from call centers to entry-level coding to the back office, because both waves seek the same thing: how much of a job is written down and how little of it is judgment. That work already left once. The mistake now would be to automate it in place, swapping a machine for a person and changing nothing else. The larger prize, including for the economies that first took this work, belongs to whoever is willing to rethink how it gets done at all.

A few weeks ago, I argued that AI did not kill the entry-level job, and the timeline backed me up: the decline started years before the models were good enough to matter. This piece is the other half of that story. An alibi that holds through 2025 is not an alibi for 2026 and everything after. The honest question is no longer whether AI will remove work. It is which work, how fast, and what a leader should do about it.

Start with a fact that looks like a coincidence and isn't. The jobs most exposed to AI today are, to a striking degree, the jobs we shipped overseas in the 2000s. Call centers and first-line support. Claims processing, data entry, bookkeeping, and first-line IT. Plus, a large share of software engineering: QA and entry-level development moved to India, China, and Eastern Europe, while the architects and senior engineers stayed on staff. We have relocated this exact category of work before, and we called it offshoring. It was an arbitrage on geography, the identical work bought in a cheaper labor market. What is happening now is a second arbitrage on the same work, except the destination is not another country. It is a model. AI agents like Claude, ChatGPT, Hermes, or OpenClaw (for the brave ones) now do the tier we once offshored, and then some.

The reason the two waves rhyme is buried in how offshoring worked. To hand a process to a team 10,000 km away, you first had to write it down. Script it, break it into tickets, test plans, and service-level agreements, and strip out the tacit judgment so a stranger could run it to standard. Offshoring was, without anyone intending it, a twenty-year codification process. It turned tacit work into explicit process maps, and those maps are precisely the structure a machine needs to run it. Offshorability, it turns out, was an early read on automatability, because both measure one property: how much of a job is written down, and how little of it is judgment.

That gives a leader something more useful than dread. The test that decided what could be offshored in 2006 is the same test that decides what AI runs well in 2026. A process that is clean enough to send abroad, clearly defined, repeatable, and measurable, is a process clean enough to automate. And that clarity is also what keeps a model reliable: a well-specified process acts as a guardrail, starving the ambiguity a model would otherwise fill with confident nonsense. So one question does three jobs at once. How clearly is this process defined? is the offshorability question, the automatability question, and the hallucination-risk question, one property across three eras. The cleaner and more measurable the work, the closer AI already is to doing it.

This is no longer a forecast. It shows up first where the work was most codified. India's IT-services firms, the companies that built the offshore delivery model, have cut fresher hiring by roughly 80% from their early-2020s peak as they pivot to AI-first delivery. The tooling is mainstream rather than fringe: by late 2025, around 90% of developers reported using AI coding tools, and over 80% said those tools made them more productive. Boilerplate, test coverage, routine tickets, the scriptable tier is exactly what those tools do best. Support and the back office are on that path too, for the reason above.

The tool, though, is only an amplifier. The 2025 DORA research found that AI magnifies whatever system it lands in: strong engineering practices compound, and weak ones only get faster and less stable. The gains come from the workflow around the model, not the model itself. The bottleneck was never the typing. It was the handoffs, the reviews, the queues, the shape of the work. Offshoring taught this lesson the expensive way: lifting a broken process to a cheaper location did not fix it; it exported the mess. Automating a broken process in place only repeats it faster. A machine bolted onto a human-shaped workflow inherits every bottleneck it has.

When the machine is genuinely better, redesign is the reason. My Oura ring developed a failing battery within a year, and the entire warranty replacement was handled by an AI agent, start to finish, in two or three minutes, with no queue and no need to repeat myself to a chain of people. It was faster and cleaner than the human version had ever been. What made it work was not the bot. It was that they automated the friction, the queue, the intake, the status chase, and then trusted the machine with a real decision that cost the company money: issue the replacement. Contrast that with the reflex most firms reach for, dropping a chatbot in front of the old script. In support, AI deflects far more than it resolves; deflection rates run high while verified-resolution rates sit far lower, often in the low 40s. A deflection that never resolves is a customer giving up, the old process with a shinier bottleneck. The principle is small, and it is the whole game: automate the friction, redesign the flow, and point the human at judgment.

That redesign question matters most for the places that took this work first. India's technology and business-process sector employs more than 5 million people and stands among the largest pillars of its economy; the Philippines' equivalent employs nearly 1.9 million workers and accounts for more than 8% of GDP. These are national systems built on codified work, now meeting the technology that runs codified work best, and the hiring cuts are the leading edge of it arriving.

The reflex is to read that as a slow-motion disaster. I read it as the opposite. The very fluency that exposes them, running codified processes at enormous scale, is exactly the capability the AI era rewards, provided it climbs a level: from executing the process to designing, supervising, and auditing the AI that executes it. The largest concentration of people who understand how these processes actually work is an asset for building the layer above them. And markets with no legacy to defend tend to skip a generation. Much of sub-Saharan Africa never built out landlines and went straight to mobile; Kenya's M-Pesa put mobile money in most adults' hands while much of the West was still standing in line at a branch. An offshore delivery firm has no on-premises workflow to protect and every reason to rebuild delivery AI-first, which is a real path to leapfrogging incumbents who are busy bolting AI onto processes they refuse to change.

Pilots have a phrase for the discipline that all of this asks for: "stay ahead of the plane." Configure now for the state that is coming, not the one that already arrived, because the pilot who falls behind the plane spends the flight reacting to things that already happened. Staying ahead here means designing for the human-and-machine combination that is arriving, rather than defending the human-shaped process that is leaving. The economics reward it. Making execution cheaper has never shrunk the amount of work worth doing; it enlarges it (economists call it the Jevons paradox), and the roles it opens, redesigning workflows, supervising models, owning the exceptions and trust, are the ones worth moving toward. The winning unit was never the machine alone or the human alone. It is the workflow deliberately built around both.

So the Monday-morning move is a small one. Pick the most codified process you own and resist the urge to pave it exactly as it is. Ask what it would look like if you designed it today from scratch for a person and a machine working together, then build one step toward that, rather than a bot in front of the old script. The work already left once. Whether it leaves you behind this time comes down to whether you are willing to redesign it, or only to re-staff it. Stay ahead of the plane.

Images source: ChatGPT Images / Claude Opus / Gérard Métrailler

Sources

• Business Today. "Indian IT's fresher hiring slump signals structural shift, not just slowdown." https://www.businesstoday.in/technology/story/indian-its-fresher-hiring-slump-signals-structural-shift-not-just-slowdown-521374-2026-03-21 Accessed 2026-07-18.

• DORA (Google Cloud). "State of AI-assisted Software Development 2025." https://dora.dev/dora-report-2025/ Accessed 2026-07-18.

• Zendesk. "Ticket deflection vs. resolution: Metrics that matter." https://www.zendesk.com/blog/ai/workflow-automation/ticket-deflection-vs-resolution/ Accessed 2026-07-18.

• IBEF. "Indian IT & Business Process Management (IT-BPM) Industry." https://www.ibef.org/industry/information-technology-india Accessed 2026-07-18.

• Philippine Daily Inquirer. "IT-BPM industry in PH outpaced global growth in 2025." https://business.inquirer.net/567026/it-bpm-industry-in-ph-outpaced-global-growth-in-2025 Accessed 2026-07-18.

• CSIS. "The Need for a Leapfrog Strategy." https://www.csis.org/analysis/need-leapfrog-strategy Accessed 2026-07-18.

Your next software renewal will likely carry two prices, not one. One line still counts the people who log in. The new line counts everything the agents they turned loose did while nobody was watching a screen. Salesforce already reports it publicly: Agentforce revenue up 205% year over year, with the seat line holding right beside it. This is a stacking model, not a swap. It changes three things at once: what you sell, what you buy, and how fast a software bill you thought was fixed can run away from you.

The bill already has two lines

For a year, the loud prediction has been that AI agents kill seat-based software. No humans logging in, no seats to sell, no business. It is a clean story. It is also already contradicted by the public numbers.

Salesforce put the split on record in its Q1 FY27 earnings. Agentforce reached $1.2B in annual recurring revenue, up 205% year over year. Combined with Data 360, the number is nearly $3.4B. The seat line did not vanish to make room. Agentforce for Sales still lists at $125 per user per month, sitting right next to consumption-priced credits for the agents those users switch on.

So the meter did not get replaced. It got a second dial. One counts people. The other counts machine work. The vendors reading this correctly are already outgrowing the ones still arguing about whether the seat dies.

The seat was never a headcount meter

The doom thesis treats a seat as a proxy for headcount, which is why it assumes the seat dies when headcount stalls. That is the wrong reading of what a seat actually does.

A seat is an identity and entitlement record. It says which data, which systems, and which actions belong to one named, accountable person. Strip a software product down to what the seat truly measures, and it is not "a human who clicks." It is "an authority that can be held responsible." Aaron Levie, Box's CEO, put the resulting shape plainly on the Platformer podcast in May: "You probably are going to have a stacking business model in software: humans still have seats, but agents will be a consumption pattern on top of that."

An agent cannot act with any authority it did not inherit. It needs a named human's identity to know what it is allowed to touch. No seat, no anchor for permission. That is why the seat is structurally different from a login count, and why it survives the arrival of agents that never log in at all. It is a governance record wearing a pricing costume.

That also explains why one meter cannot price both things. A human has a physical ceiling on how many actions they take in a day. An agent has none, and it runs at night. Bessemer has already started naming the new cost primitives that this creates: cost per thousand tokens, cost per resolved request, cost per agent minute. Not one of them maps onto "per seat." Seats meter identity. Consumption meters throughput. Trying to bill both on a single dial causes it to break in one direction or the other.

This is already the market, not a forecast

None of this rests on one CEO reading his own business favorably. The shift shows up across the category.

Roughly 40% of software companies now run hybrid pricing, a base fee plus a variable usage layer, and that is projected to exceed 60% by the end of 2026. Well over half now offer some form of usage-based billing, up from 27% in 2018. Credit-based pricing, the most agent-native flavor, jumped from 35 companies in the PricingSaaS 500 Index at the end of 2024 to 79 a year later.

The performance gap is the part worth pinning to the wall. In Chargebee's 2025 monetization survey, companies evolving their pricing alongside their AI were nearly twice as likely to expect high growth as those that left pricing untouched. Correlation, not proof. Yet the signal is hard to ignore: the vendors treating pricing as a live system are the ones expecting to pull ahead.

Which puts the same shift in front of two very different desks, and the answer looks different depending on which side of the software you sit.

If you build and sell software

The instinct will be to bolt a usage SKU onto the price sheet because everyone else is. Resist the reflex. The discipline is to deliberately decide which part of your product is identity and which part is throughput, and to price each for what it actually is.

Seat-price the parts that confer authority: access, governance, entitlements, and the record of who is allowed to do what. Consumption-price the parts that run unbounded: agent actions, resolved requests, and tokens burned against your platform on a customer's behalf. Get that line in the wrong place, and you either cap your own upside as usage explodes, or you hand customers a bill so volatile it makes them churn on principle. The second dial is only defensible when it meters something genuinely unbounded that customers agree is worth metering.

There is a quieter defensibility point underneath the pricing one. Levie noted that Box has not approved a single internal project to rebuild an existing software service from scratch with AI, because the workflow is the cheap part to clone. The governed identity, data, and permission layer beneath it is not. That layer is exactly what the seat prices are, and it is why "just have the agent rebuild it" is a worse idea than it sounds.

If you buy and run software

Here is where the meter change stops being interesting and starts touching your bottom line, fast.

Stop evaluating vendors on per-seat cost alone. That number is now the fixed, legible, well-behaved half of the bill. The other half is a variable line that can balloon in a quarter if a team wires an agent into a workflow and lets it run. A software cost you booked as fixed can behave like a cloud bill, and cloud bills are where finance teams have been ambushed for a decade.

Treat agent consumption the way a disciplined engineering org already treats cloud spend, before the first invoice surprises you, not after:

• Put a real ceiling on it. Budget caps and rate limits per agent, per team, per workflow, so an unbounded meter cannot silently become an unbounded invoice.

• Make the spend attributable. Issue a distinct API key per application, team, or use case, so every token charged by OpenAI, Anthropic, or your platform vendor is attributed to a named owner rather than a single anonymous pool. You cannot govern a number you cannot trace.

• Instrument it. Dashboards, spend alerts, and anomaly triggers on the consumption line, reviewed at a significantly accelerated cadence than any other operating cost, not at renewal.

• Assign the accountability. Showback or chargeback: the team turning the agent loose is the one that sees the bill.

And treat the seat itself as a governance control, not a checkbox for IT. Every agent identity should map to a named, accountable human with explicit entitlements. "Who authorized this agent to act as me, and against which data?" is now a real operating question with a real cost attached to getting it wrong.

The close

The doom thesis asked whether the seat survives. It was the wrong question. The seat was never the cost center. It is the accountability record for everything now running on top of it. The number that can hurt you is the one on the other line, and the discipline is deciding, on both sides of the software, what you are willing to let run through it.

Images source: ChatGPT Images / Claude Opus / Gérard Métrailler

Sources

Casey Newton, "The best argument I've heard for why AI won't take your job", Platformer, published 2026-05-13. https://www.platformer.news/ai-job-loss-box-ceo-aaron-levie/. Accessed 2026-07-14.

Salesforce, "Salesforce Delivers Record First Quarter Fiscal 2027 Results," press release, 2026-05-27. https://www.salesforce.com/news/press-releases/2026/05/27/fy27-q1-earnings/. Accessed 2026-07-14.

Salesforce, Agentforce pricing. https://www.salesforce.com/agentforce/pricing/. Accessed 2026-07-14.

Rob Litterst, "What actually works in SaaS pricing right now," Growth Unhinged, 2026-01-07. https://www.growthunhinged.com/p/2025-state-of-saas-pricing-changes. Accessed 2026-07-14.

Bessemer Venture Partners, "The AI pricing and monetization playbook". https://www.bvp.com/atlas/the-ai-pricing-and-monetization-playbook. Accessed 2026-07-14.

OpenView Partners, "The State of Usage-Based Pricing". https://openviewpartners.com/blog/state-of-usage-based-pricing/. Accessed 2026-07-14.

Chargebee, "2025 State of Recurring Revenue & Monetization Report". https://www.chargebee.com/resources/guides/2025-state-of-subscriptions-revenue-growth/. Accessed 2026-07-14.

Your AI sounds equally sure whether it is right or guessing. That flat confidence is not a quirk; the same training that makes a model agreeable also makes it overconfident. Leadership has always meant deciding before all the facts are in, and the tool used to be the one certain thing on the desk. It just stopped being certain. This is the fourth cognitive lens: reading every answer as one drawn from a distribution, and supplying the calibration that the machine cannot.

For the last few issues, I mentioned that this is a series about three cognitive lenses: tension, connection, and reduction. Well, here is the fourth one in the trilogy! Uncertainty is the lens you notice is needed only once you are already past where you thought the set would end. It hides behind the other three until a decision goes wrong, and then it is the only thing in the room.

Nothing new, and everything new

Start with the part that is not new. Every consequential decision a leader ever made was a bet placed before the facts were all in. Drucker knew it. Every operator has lived it. Judgment under uncertainty is the oldest skill in the job, not a discovery of the AI age.

Here is what changed. For roughly seventy years, our machines were deterministic. Same input, same output, every time. The spreadsheet that returned 42 yesterday returns 42 today. Now the most powerful tool on your desk is probabilistic. Ask the same question twice, and you can get two different answers. The oldest leadership skill just met a genuinely new object, a machine that is itself uncertain, and the two compound.

So the discipline now runs in two places at once. It always pointed at the world. Now it also has to point at the machine's own output. That doubling is the whole story, and it is the answer to both "AI changes everything" and "AI changes nothing." The skill is old. The second place you have to apply it is new.

You are not using a faster calculator

A classical computer is deterministic by design. Its lineage runs through the machine Alan Turing described in 1936, and the stored-program architecture John von Neumann set down in 1945. One input, one reproducible output, auditable to the last digit. That is the world of the 0s and 1s, and it trained multiple generations of leaders to expect computers to be exact.

A large language model does not work that way. It is predictive. It picks the next word from a spread of possibilities, and its theoretical lineage is the probabilistic Turing machine, which can accept an input on one run and reject the identical input on the next.

The trouble is that most leaders carry the deterministic mental model into a probabilistic tool. They expect exact and reproducible results. They get a machine that varies, invents, and cannot be fully retraced. Then they are surprised. You are not using a faster calculator. You are using a different kind of system, and your instincts about the old one mislead you about this one.

The confidence is a costume

Here is the mechanic, in plain terms. Under the hood, the model picks the next word, or piece of a word (a token), from a set of candidates and hands you one in fluent prose. The answer you read is one statistical draw. It is a sample, presented as if it were the whole.

Picture the error bars on a chart, the little whiskers drawn through a data point, running above and below the dot, showing how wide the real answer could be. A short whisker means "I'm confident, it's right about here." A long whisker means "my best guess is the dot, but frankly, it could be anywhere along this line." Same dot, wildly different meaning depending on the whisker. The model computes that whisker. The interface shows you only the dot. It deletes the error bars and hands you the bet with the odds torn off the ticket.

Worse, the width changes, and the confidence does not. On a well-worn base rate, the whisker is short. On something recent, proprietary, or niche, it is enormous. The voice is identical at both.

There is a reason the voice is always sure. The training step that tunes a model to be helpful and agreeable, by rewarding the answers people prefer, also teaches it to sound certain (the technical term is RLHF, if you want the deep dive). The reward models used in that step favor confident-sounding answers regardless of whether the answer is any good. Train a machine to please us, and you train it to sound sure. It is rewarded for confidence, not for calibration. The felt result is a model that tells you it is "positively" sure and is wrong far more often than that.

Calibration is a scarce skill now

AI has made generating answers nearly free. It did not make weighting them well. That gap is the new job.

Because feedback-tuned models run overconfident, a team that leans on them inherits the overconfidence, unless a human is deliberately the calibration layer. The model floods you with fluent options. It will not tell you which one to believe, or how much.

The people who do this well already have a name. Philip Tetlock's superforecasters are not smarter than everyone else. They keep score in granular probabilities, they distinguish 60% from 90% and mean it, and they update the moment the evidence turns (updating your view as evidence arrives, if you want the Bayesian label). That is the discipline the machine cannot supply for you. Supplying it is the leader's new work.

Use AI for the base rate, keep the human for the tail

There is a clean division of labor hiding in all this. AI is trained on what already happened, so it encodes the base rate and the consensus well. That is real value, and it covers the fat middle of the distribution where most decisions live. It is also structurally blind to the genuine tail, the novel event that is not in the corpus. The fat tails live exactly where the model cannot see, and that is where the largest value is made and destroyed. Use AI for the base rate. Keep the human for the tail. This is not a choice between the two; it is both, assigned to the part each does best.

One trap to name on the way out. Ask a model for the ways a decision could go, and it returns a clean list in seconds. The list feels like coverage. It is not. Listing outcomes is free now; pricing them is the whole job, and enumeration is not estimation. Let the machine list. Do not let it flatter you that listing is the same as weighing.

The fourth lens, and Monday morning

Name the set once, in our words. Reduction finds what is true. Connection maps how it relates. Tension holds the both-and over time. Uncertainty assigns the odds and keeps updating them. Value creation is the lens through which all four are aimed. Four lenses, one job.

Then do something with it on Monday. Take one live decision resting on an AI answer and do three things to it. Write the probability you would actually assign, as a number, not "high, medium, low." Name the single piece of evidence that would change your mind. Name the tail the model cannot see. It works for an analyst pricing one assumption, a function leader weighing a plan, or a board calibrating a thesis. The altitude changes; the move does not.

The machine did not make us more certain. It made certainty cheaper to fake, and judgment the thing worth paying for.

Images source: ChatGPT Images / Claude Opus / Gérard Métrailler

Sources

• Leng, Jixuan, Chengsong Huang, Banghua Zhu, and Jiaxin Huang. "Taming Overconfidence in LLMs: Reward Calibration in RLHF." International Conference on Learning Representations (ICLR) 2025. arXiv:2410.09724. https://arxiv.org/abs/2410.09724. Accessed 2026-07-11.

• Turing, Alan M. "On Computable Numbers, with an Application to the Entscheidungsproblem." Proceedings of the London Mathematical Society, series 2, vol. 42 (1936-37): 230-265.

• von Neumann, John. "First Draft of a Report on the EDVAC." Moore School of Electrical Engineering, University of Pennsylvania, 1945.

• Tetlock, Philip E., and Dan Gardner. Superforecasting: The Art and Science of Prediction. Crown, 2015.

• Duke, Annie. Thinking in Bets: Making Smarter Decisions When You Don't Have All the Facts. Portfolio, 2018.

Ask an agent to price your product, and it hands you a polished version of what your competitors already do. That's not a bug; it's the objective working perfectly. A model is a machine for the most likely next thing, which means it drifts toward the average answer, the one everyone else's agent is also producing. The expensive human move runs the other way. SpaceX did it with a rocket and found that 98% of the cost was due to inherited habit. Here's why that move is now the edge, and why a faster agent makes skipping it more dangerous, not less.

Third in a short series on three lenses for thinking under pressure: tension, connection, and reduction.

Open a fresh chat. Ask a capable agent to design your product's pricing. What comes back is competent, confident, and reasonable: three tiers, a per-seat model, an annual discount, and an enterprise "contact us." It looks like the pricing page of every company in your category, because that is exactly what it is.

That is not the agent being lazy. It is the agent working precisely as built. A model predicts the most probable next word, then the next, all the way to the end of the answer. Stretch that across a whole pricing strategy, and "most probable" becomes "most common," which becomes the center of everything the model has ever seen. The technical name for the pull is regression to the mean. One 2025 study put hard numbers on it: across 2,200 essays, each new human-written one added roughly two to eight times as much fresh variety to the collective pool as each new GPT-4 essay did, and the model's drift toward the middle persisted even when the researchers threw every clever prompt at it.

The fluency is real. So is the gravity toward the average. And the average turns out is about to become the most worthless thing in business.

The move the machine can't make

There is a discipline for refusing the average answer, and the cleanest demonstration of it is also the most over-told, so I'll be quick.

In the early 2000s, a rocket to orbit cost around $65M, and the industry explanation was, in effect, that rockets have always cost that much. Reasoning by analogy. SpaceX asked a different question: what is a rocket actually made of, and what do those metals and that fuel cost on the open commodity market? The answer came back at roughly 2% of the sticker price. The other 98% was not physics. It was an accumulated convention nobody in the industry could still see, because it had been there the whole time. Musk's phrase for the move was to "boil things down to the most fundamental truths" and "reason up from there," instead of reasoning from what everyone already does.

I call this lens reduction. It's also known as first principles, and Aristotle called it archai. Not reductionism, the bad habit of explaining everything away. Not the cost-cutting kind that haunts every budget review either, which is the meaning the word unhelpfully suggests in an operating context. Reduction means getting down to bedrock truth and rebuilding from it. Take the thing apart until you hit what is actually true, the hard floor of physics or unit economics or the real job the customer is hiring you to do, then reason up and ignore the conventions stacked on top.

Notice what the move is not. It is not optimization. An agent optimizes beautifully; point it at your pricing, and it will polish the existing shape until it gleams. Reduction asks whether that shape should exist at all. Answering that means standing outside the cloud of every prior answer, which is the one place a next-word predictor structurally cannot go. The model lives inside the cloud. That is its whole job.

Why did this get more valuable, not less?

The lazy reading is that AI makes thinking cheap, so thinking matters less. The economics run the opposite way.

When the average competent answer drops to near-zero cost and shows up in four seconds, it stops being worth anything, precisely because everyone now has it. Your competitor has it. Your competitor's competitor has it. You are all prompting the same handful of models with the same unexamined questions and receiving the same average back. What holds value is the answer that sits off the average, and reduction is how a human gets there.

Here is the part I did not see coming when I started drafting this piece. The same decomposition that helps a human find a better answer is also what makes the AI agent more reliable. The 2025 research on agentic systems is consistent on this: break a task into its fundamental sub-problems, and accuracy goes up while hallucination goes down, because each piece is now small enough to handle cleanly. One method that decomposes a task by its formal complexity lets an agent perform measurably better on hard combinatorial and database-querying benchmarks.

Sit with that for a second. The human move that produces a non-obvious answer and the engineering move that makes an agent trustworthy are the same move, performed at two altitudes. Reduce the problem to fundamentals, hand the well-framed pieces to capable agents, and you get genuinely superhuman output. Skip the reduction, hand over the problem exactly as you inherited it, and you get superhuman speed at scaling the wrong thing. The framing is the whole game, and it's the human's job.

The premise an agent will defend to the end

Now the part that costs real money this year.

An agent will execute a flawed premise with total confidence, because nothing inside it is built to doubt the frame you handed it. It cannot tell the difference between a good question and a bad one. It can only answer fluently. So a bad question yields a fluent, well-structured, completely wrong answer, delivered fast and with no tell.

The data is already uncomfortable. In one large 2025 developer survey, the top frustration with AI coding tools, named by two-thirds of respondents, was output that is "almost right, but not quite." Nearly half said debugging that AI-generated code takes them longer than writing it themselves would have. (Yes, the models have improved significantly since, yet the underlying architecture remains.) That failure has an old name in aviation: automation bias. People follow a confident wrong instruction from a machine at a meaningfully higher rate than they make the same mistake on their own, and the effect is worse for the least experienced, the very people least able to tell a plausible answer from a true one.

The root of the verification problem is singular. You cannot catch a confident wrong answer unless you can reason from fundamentals yourself. Reduction is the only way to check polished output against bedrock, rather than against whether it "looks right," which is exactly the test a fluent wrong answer is built to pass. The leader who can still take a problem apart is the one who can govern an agent. The one who has let that muscle go ships the plausible-wrong answer at machine speed and finds out a quarter later.

When to skip it

Reduction has a cost, and the discipline includes knowing when not to pay it. Taking a problem to fundamentals is slow. For the reversible, low-stakes, genuinely solved decision, reasoning by analogy is correct precisely because it is fast, and handing it to an agent unexamined is exactly right. Reserve reduction for the high-stakes call, the stuck problem, and the contrarian bet, where convention is most likely to be hiding something.

Two cautions, because first-principles enthusiasm reliably produces both. First, some conventions are load-bearing. They encode a hard-won reason that is no longer visible, and tearing one down because you cannot see its purpose is a self-inflicted wound. Before you discard an assumption, find out why it exists. If nobody can say, you have found a real opportunity. If somebody can, you may have found a fence worth leaving standing.

Second, reduction can be wrong about the bedrock itself. Reason confidently up from fundamentals that are incomplete, and you get an answer that is internally flawless and externally false, a map drawn so cleanly you stop checking it against the ground. The fundamentals you reduced to are still a model of the thing, not the thing. The defense is the same one that catches a confident agent: test the rebuilt answer against reality before you trust it.

Monday morning

Pick one decision you are about to hand an agent this week. Your pricing, a market-entry plan, a hiring profile, a workflow you mean to automate. Before you prompt anything, write down the single biggest assumption baked into how you framed it. Then ask one question: is this a hard constraint, the floor of physics, unit economics, the real job the customer hires you for, or is it a convention everyone in your category has simply stopped questioning?

If it is a convention, you have found the work the agent cannot do for you, and the work most worth doing. Reframe from the fundamental truth up, then hand the agent the pieces. You will get a different answer than your competitors, all of whom are about to ask the same model the same unexamined question and get the same average back.

The machine is extraordinary at giving you the best version of the existing answer. It just cannot tell you when the existing answer is wrong. That part is still yours.

Images source: ChatGPT Images / Claude Opus / Gérard Métrailler

Sources

Elon Musk, quoted in James Clear, "First Principles: Elon Musk on the Power of Thinking for Yourself," jamesclear.com. https://jamesclear.com/first-principles Accessed 2026-06-11.

Catherine Clifford, "Billionaire Elon Musk says this is 'a powerful, powerful way of thinking'," CNBC, 2020-02-28. https://www.cnbc.com/2020/02/28/billionaire-elon-musk-this-is-a-powerful-way-of-thinking-but-hard-to-do-how-it-works.html Accessed 2026-06-11.

Shane Parrish and Rhiannon Beaubien, The Great Mental Models, Volume 1: General Thinking Concepts, Farnam Street Media, 2019. ISBN 9781999449001.

Kibum Moon, Adam E. Green, and Kostadin Kushlev, "Homogenizing effect of large language models (LLMs) on creative diversity: An empirical comparison of human and ChatGPT writing," Computers in Human Behavior: Artificial Humans, vol. 6, 2025, 100207. https://doi.org/10.1016/j.chbah.2025.100207

"Galton's Law of Mediocrity: Why Large Language Models Regress to the Mean and Fail at Creativity in Advertising," arXiv:2509.25767, 2025. https://arxiv.org/abs/2509.25767 Accessed 2026-06-11.

"An Approach for Systematic Decomposition of Complex LLM Tasks," arXiv:2510.07772, 2025. https://arxiv.org/abs/2510.07772 Accessed 2026-06-11.

Burak Gozluklu, "How task decomposition and smaller LLMs can make AI more affordable," Amazon Science, 2024-09-19. https://www.amazon.science/blog/how-task-decomposition-and-smaller-llms-can-make-ai-more-affordable Accessed 2026-06-11.

Stack Overflow, "2025 Developer Survey: AI," via ShiftMag, "84% of developers use AI, yet most don't trust it," 2025. https://shiftmag.dev/stack-overflow-survey-2025-ai-5653/ Accessed 2026-06-11.

Addy Osmani, "The reality of AI-assisted software engineering productivity," 2025. https://addyo.substack.com/p/the-reality-of-ai-assisted-software Accessed 2026-06-11.

This is the second article in a short series on three cognitive lenses for thinking under pressure: tension, connection, and reduction.

AI has made the components of leadership work nearly free: analysis, draft, model, first-pass decision. The value didn’t evaporate. It relocated to the one place a model still can’t reach: the connections between the parts. That shift is why systems thinking, long filed under “nice-to-have,” is now the highest-return skill a leader owns. A nine-second corporate catastrophe this spring shows what it costs to keep watching the parts instead.

For most of the careers of the people now in senior seats, the job was decomposition. Break the business into functions, the functions into metrics, the metrics into targets. Fix each part and trust a healthier whole to follow. It was a reasonable bet in a slow, loosely coupled world, where you could tune one box on the org chart and the others wouldn’t notice for a quarter.

That bet is quietly going underwater. AI is what tipped it.

Consider what an executive produces in a week: a market read, a board memo, a scenario model, a first call on a hard decision. Two years ago, each was a unit of scarce, expensive human cognition. Today, a competent model drafts each one in minutes, for a few tokens. The cost of producing the components of the work has collapsed.

Here is the part that should change how you spend your attention. As the cost of each component approaches zero, the system's value does not decline. It concentrates, with almost arithmetic certainty, on whatever did not get cheaper. And what did not get cheaper is the connective tissue: how the market read informs the price, how the price collides with the comp plan, how the comp plan bends the sales behavior that reshapes the market you started by reading. The parts are commodities now. The wiring between them is not.

Reasoning about that wiring is the academic discipline known as systems thinking. I have come to call the skill itself connection, because that is what it trains you to see: not the boxes, the lines between them. For decades, it quietly separated the leaders who fixed symptoms from the leaders who fixed systems. You could still have a fine career fixing parts. The reason that era is closing is not philosophical. It is an accounting fact about where the margin now lives.

What connection actually is, minus the seminar

Connection explains an outcome by the structure that produced it, not the event that announced it. “Sales missed the quarter” is an event. The incentive that pulled deals forward, the onboarding delay that churned those same deals, and the dashboard that flagged it a month too late: that is the structure, and it was always going to produce that number. One view blames a person. The other sees the machine. Donella Meadows, who wrote the field’s most-quoted primer, spent a career on the same point: a system is more than the sum of its parts, because the behavior lives in how the parts interact, not in the parts themselves.

Three pieces of that structure do most of the explaining, and autonomous agents have just made each one urgent.

The first is stocks: the quantities that accumulate quietly while you watch a different number. Cash and inventory are familiar. The dangerous ones in an agent deployment are nearly invisible: the permissions an agent has slowly accreted, the context it has been fed, the actions it has already taken in your name. Stocks fill on their own schedule, not yours, and you notice them the moment they overflow.

The second is feedback loops: the circuit from an action to its consequence to a correction. Good governance is mostly the deliberate design of these loops. Something acts, something notices, something pulls it back. Agents break the circuit at one precise point. They compress the acting to milliseconds and leave the noticing at human speed. An agent takes a thousand actions in the time it takes a person to review one. When the loop runs slower than the thing it governs, the system runs in open loop, which is a polite way of saying it is not governed at all.

The third is delays: the gap between a change and its full consequence. You loosen a control, and the erosion surfaces much later, one permission at a time, long after the team that loosened it has moved on. Leaders who ignore delays either conclude too early or stack a second fix onto a system that has not yet absorbed the first.

None of this is a new theory. What is new is the clock speed at which getting it wrong now bites. One company learned that in less time than it takes to read this paragraph.

Nine seconds

In April 2026, an AI coding agent at a small software company called PocketOS hit a credential mismatch on a routine staging issue. While reasoning toward a fix, it found an access token in an unrelated file, scoped for any operation, including destructive ones. It used the token to delete the database volume. The production data and all backups went with it because the hosting setup stored the backups within the very volume they were meant to protect. Start to finish: nine seconds. The company fell back to a three-month-old copy and spent more than a day in the dark.

Asked to explain itself, the agent wrote: “I violated every principle I was given. I guessed instead of verifying. I ran a destructive action without being asked.”

It is tempting to read this as a story about a rogue model. It is the opposite. Look at the connections rather than the parts, and almost nothing in it is about AI. An over-scoped token was left where an agent could reach it: a rules failure. Backups lived in the same place as the data they were supposed to protect: a structural coupling failure. No loop sat between the action and a human who could catch it: a feedback failure. The model was the most self-aware element in the building, correctly diagnosing itself 9 seconds too late. Every condition that turned a small mistake into a catastrophe lived in the structure around the model, not the model itself. That is what connection sees, and a parts-by-parts view cannot.

PocketOS is not an outlier waiting to be engineered away. It is the visible edge of a gap that the whole market is carrying. In a Deloitte survey of more than 3,200 leaders across 24 countries, only about 1 in 5 companies had a mature approach to governing autonomous agents. The other 80% deploy systems that operate autonomously, without clear decision boundaries, real-time monitoring, or a usable audit trail. Read that missing list closely. Decision boundaries are rules. Real-time monitoring is a feedback loop. An audit trail is what you need because consequences arrive with a delay. The vendors will not patch this in the next release. It is a systems-design gap, in that exact vocabulary, open in four of five companies.

Where the leverage actually sits

Here is where most leaders reach for the wrong tool. Faced with an agent that just cost them a weekend, the instinct is to grab the controls you can see: tighten the token, cap the spend, add a rate limit, write a sterner policy. Worth doing. Also, in Meadows’s framework, it is close to the weakest move on the board. She ranked the places you can intervene in a system from least to most powerful, and the dials, numbers, and parameters sit near the very bottom. Tuning them changes how fast the system runs, not how it behaves.

The high-leverage interventions sit further up the same list: the information flows and the rules. Information flows mean getting the right signal to the right decision-maker fast enough to matter, the loop that was missing at PocketOS. Rules mean changing what the agent can reach and optimize for, not merely how much it can reach. The uncomfortable translation: the company frantically lowering its token cap is working the bottom of the hierarchy, while the company redesigning who-sees-what and what-an-agent-can-touch is working the top. Same incident, two very different returns on the same hour of executive effort.

The deepest leverage point is one nobody can buy: the question the organization reaches for first. A team that asks “which part failed?” keeps installing better parts into a structure that keeps producing the same surprise. A team that asks “What about the structure made this likely?” fixes the load-bearing part. That shift costs nothing and outperforms any tool on the market.

What to do on Monday morning

You do not need a systems-dynamics course. You need one habit installed in the meeting where agents come up. Refuse to let the conversation stay on the parts, and ask three questions in order.

What stock is filling here that no one is watching in real time? Spend is the obvious one. Accumulated permissions and context are what surprise people later, because they fill in the silence.

Is our feedback loop faster than the thing it governs? If a human reviews weekly and the agent acts 1,000 times a day, the honest answer is no: you are running in an open loop, no matter what the policy document claims.

Where are the delays between an action and our ability to see its real consequences? Those gaps, in customer trust, in security posture, in team behavior, are where the expensive surprises compound while the dashboard stays green.

If you have an appetite for only one, ask the first. Most governance failures this year are simply a stock that filled faster than anyone was looking.

The agentic era did not invent a new kind of management failure. It took the oldest one, mistaking a healthy part for a healthy whole, and ran it at a speed that turns yesterday’s adequate oversight into this morning’s nine-second catastrophe. The economics have already moved the value into the connections. The only open question is whether your attention has moved with it.

Images source: ChatGPT Images / Claude Opus / Gérard Métrailler

Sources

Donella H. Meadows, Thinking in Systems: A Primer, Chelsea Green Publishing, 2008.

Donella H. Meadows, “Leverage Points: Places to Intervene in a System,” The Donella Meadows Project, 1999. https://donellameadows.org/archives/leverage-points-places-to-intervene-in-a-system/ Accessed 2026-06-11.

Thomas Claburn, “Cursor-Opus agent snuffs out startup’s production database,” The Register, 2026-04-27. https://www.theregister.com/2026/04/27/cursoropus_agent_snuffs_out_pocketos/ Accessed 2026-06-11.

Connie Lin, “‘I violated every principle I was given’: An AI agent deleted a software company’s entire database,” Fast Company, 2026-04-28. https://www.fastcompany.com/91533544/cursor-claude-ai-agent-deleted-software-company-pocket-os-database-jer-crane Accessed 2026-06-11.

Deloitte, “State of AI in the Enterprise 2026,” 2026. https://www.deloitte.com/us/en/what-we-do/capabilities/applied-artificial-intelligence/content/state-of-ai-in-the-enterprise.html Accessed 2026-06-11.

Deloitte, “Business and IT leaders report AI agents are scaling faster than their guardrails,” 2026. https://www.deloitte.com/us/en/insights/topics/emerging-technologies/ai-agents-scaling-faster.html Accessed 2026-06-11.

This is the first article in a short series on three cognitive lenses for thinking under pressure: tension, connection, and reduction.

A company handed two-thirds of its customer service to AI, shed the equivalent of 700 support agents, and called it solved. A year later, the CEO said the words every leader dreads: We went too far. It was not an AI mistake. It was a category mistake, the same one that breaks reorgs, rewrites, and roadmaps. Many of your hardest calls were never problems with answers. These were tensions to be managed, and AI agents just turned spotting the difference into a survival skill.

In early 2024, Klarna launched an AI assistant built with OpenAI and, within its first month, said the bot was handling two-thirds of all support conversations and doing the work of 700 agents. From 2022 to 2024, the support headcount fell by roughly that many due to a hiring freeze and attrition. Efficiency was solved, and the company said so loudly. Then the quality complaints arrived, customer trust slipped, and by May 2025, CEO Sebastian Siemiatkowski admitted the company had gone too far. Klarna started rehiring humans into a blended model in which AI handles volume and people handle judgment.

Read quickly, and this looks like a story about AI falling short. It is not. The company did not pick the wrong tool; it misclassified the question. “How much of this work should AI carry, and how much should humans?” has no permanent answer. It moves with the model, the customer, and the quarter. Treat a moving question as settled, and the bill always comes due.

A problem has a solution. You diagnose it, fix it, and it stays fixed. A late invoice run, a broken deployment pipeline, a gap in the sales-coverage map: real problems, and the satisfying kind, because once handled, they leave. A tension is different in kind, not a different degree. It is a pair of values, each with a real claim, that never stop needing each other. Speed and quality. Centralization and decentralization. Purpose and profit. Human judgment and machine scale. You do not get to pick one and be done, because the moment you starve either side, it comes back wearing new symptoms.

Here is the loop almost every leadership team runs without naming it. One value feels under-supplied this quarter, so the team swings hard toward it and gets applause for two or three quarters. Then the other value, the one quietly starved by the swing, starts surfacing as a fresh batch of complaints. The instinct is to swing back, and the cycle restarts from the opposite pole. Smart teams can run this for years, each swing dressed up as a bold new initiative, before anyone notices it is one tension oscillating rather than a sequence of problems being solved. Rehiring the support team is simply the backswing. It would have been far cheaper to name the tension before the first one.

What is new is the speed. Every team is now dividing labor between humans and agents, and that division is a tension, not a setting you configure once. Agent autonomy against human oversight. Velocity against control. These used to be annual offsite questions; they now sit within daily operations, and the bill for getting them wrong arrives faster than the year it took to surface above. The right setting moves with each new model, and a moving optimum is the signature of a tension. Anyone selling a permanent answer to “how much should we let AI run” is selling the next overshoot.

None of this is new thinking, only newly urgent. A consultant named Barry Johnson founded the discipline in 1975 and published the canonical book, Polarity Management: Identifying and Managing Unsolvable Problems, in 1992. (Thanks, Aram, for pointing me in that direction.) Unsolvable is the load-bearing word. Jim Collins gives the same idea a more memorable name. In Beyond Entrepreneurship 2.0, he contrasts the “tyranny of the or,” which forces every choice into A or B, with the “genius of the and.” Disciplined thinkers, in his research, are comfortable holding both.

The strongest reason to trust this is that serious people keep arriving at the same place through different doors and naming it differently. Johnson calls it a polarity. Collins calls it the genius of the and. On The Curiosity Shop podcast this year, Brené Brown and Adam Grant call it a paradox, and Grant’s plain definition is the one worth keeping: “two opposites coexisting,” which are “often the source of our best ideas, our most important decisions, but they can drive us crazy too.” That last clause is the part the tidier names leave out. The discipline is the willingness to sit with something that drives you a little crazy rather than resolve it for relief.

When a team starts treating a tension as a tension, three things change, and none of them is soft. The conversation stops being a debate to win, because permanently winning either side is the failure mode. The job becomes describing both poles fairly, including the exact cost each one imposes when it dominates. That sounds like a group hug. It is closer to the opposite, because it denies everyone the relief of being declared right.

The team also starts watching for warning signs that hide in plain sight as recurring complaints. Falling satisfaction scores are a leading indicator that the human pole has been starved, but a team that never sets the tension up to be watched reads the signal as a CEO mea culpa rather than a dashboard. Name a tension, and its two failure signatures, and the next overshoot announces itself while a cheap correction is still on the table. With agents in the loop, that early warning matters more because the swing comes faster.

And the work changes shape. The job is not to resolve the tension; it is to notice when one pole has overshot and pull deliberately back. Small, recurring corrections, not heroics. It looks less like a turnaround and more like a thermostat, because a dramatic announcement is itself an overshoot.

For anyone who operates or invests in software, this is not abstract. The companies that compound through a full ownership cycle are rarely the ones that found the perfect operating model; they are the ones that kept adjusting as conditions moved. Growth against efficiency. Founder velocity against institutional discipline. Now, human craft against AI agent scale. Each breaks the same way when a team decides it has finally been solved.

AI sharpens that risk to a point. The human-and-agent balance you set at the start of a hold is not the balance that should exist at exit, because the technology underneath it compounds across the very three-to-five-year window in which you are building the business to sell. Set it once in year one and defend it; by year three, you are optimizing for a capability frontier that no longer exists. The buyer is underwriting the business as it will run then, with the agents that exist then. The teams that keep retuning the line as the models improve reach the table with a company that is structurally more valuable than the teams that solved it once. The classic stall was defending last cycle’s answer. The new one is defending an answer that AI’s own rate of change has already obsoleted underneath you.

So here is the move, and it works for a founder with 7 people and a chief product officer with 700. The artifacts compress as the scale shrinks; the discipline does not. Pick the loudest tension on your team, the one dragging the same argument into the same meeting. A good 2026 candidate: how much to let the agents run versus how much a human signs off. On a whiteboard, write the genuine upside of each pole in one sentence. Under each, write the specific cost that pole inflicts when pushed too far. Then ask the room one question: which of those two costs are we paying right now? The disagreement stops being about who is right and becomes about where on the map you currently sit, which is a question you can actually act on.

Twenty honest minutes tend to produce more clarity than the last three meetings on the same topic combined. Run it as a recurring move, and it changes something deeper than any single decision: it changes how the team handles disagreement itself. Once you can hold a tension rather than solve it, the next question is where the leverage sits to govern the swing. That is a question about systems, and it is where this series goes next.

Pick a side to lean today. Then keep the other one alive on the team, because in the age of AI agents, you will need it back faster than you think.

Images source: ChatGPT Images / Claude Opus / Gérard Métrailler

Sources

Barry Johnson, Polarity Management: Identifying and Managing Unsolvable Problems, 2nd ed., HRD Press, 1992. ISBN 9780874251760.

Jim Collins and William Lazier, Beyond Entrepreneurship 2.0, Portfolio, 2020. ISBN 9780399564239.

Fortune, “Klarna plans to hire humans again, as new landmark survey reveals most AI projects fail to deliver,” 2025-05-09. https://fortune.com/2025/05/09/klarna-ai-humans-return-on-investment/ Accessed 2026-06-10.

Entrepreneur, “Klarna Is Hiring Customer Service Agents After AI Couldn’t Cut It on Calls, According to the Company’s CEO.” https://www.entrepreneur.com/business-news/klarna-ceo-reverses-course-by-hiring-more-humans-not-ai/491396 Accessed 2026-06-10.

The Curiosity Shop with Brené Brown and Adam Grant, “Exploring the Paradoxes of Human Nature,” 2026-05-28. https://www.youtube.com/watch?v=5tVqjcbs3io Accessed 2026-06-10.

The entry-level job is disappearing, and the headline has already named the culprit: AI. Here is the problem with the case. The sharp decline set in around late 2022, and AI was nowhere near good enough to replace anyone until late 2025. A cause cannot arrive three years after its effect. So who actually did it? The honest answer is a lineup, and the most interesting suspect never makes the headline: a feedback loop nobody had on a budget line.

Lay the two facts that built that headline side by side, and the verdict writes itself. The share of new hires going to early-career workers has fallen sharply across developed economies since 2022. The drop concentrates in exactly the occupations most exposed to generative AI. A growing research literature points the same way, and every executive who has quietly frozen entry-level headcount this year has felt the logic from the inside. Motive and a body. What else do you need?

A timeline, it turns out, and an honest look at who else was in the room. Because the tidy verdict is, on the evidence, the least defensible of the available explanations. The real version is not “it was AI all along.” It is a lineup, and more than one suspect had a hand in it.

Two shocks wearing the same clothes

Generative AI was not the only thing that happened to knowledge work after 2020. Remote work also happened, and it affected almost exactly the same jobs.

That is not a loose impression. In a recent working paper, Peter John Lambert and Yannick Schindler measure it directly: the occupations most exposed to generative AI and those most exposed to working from home show a strong correlation (0.77) across more than 680 narrowly defined occupations. Software developers, technical writers, and management consultants sit near the top of both lists. The jobs an AI can most help with are, to a striking degree, the same jobs you can most easily do from your kitchen table.

Once two shocks travel together that closely, any study that looks at only one of them is in trouble. Credit the hiring slump to token predictors while ignoring the disappearing commute, and you are quietly crediting AI with remote work’s effects too, because the data cannot tell them apart. Statisticians call this omitted-variable bias. Leaders know it by a simpler name: blaming whichever suspect you happened to be looking at when the lights came on.

So Lambert and Schindler do the thing the headline studies skip. They put both suspects in the same room, across 243 million new hires and 407 million job postings in the United States, the United Kingdom, Canada, and Australia from 2017 to 2025. Separately, each shock looks guilty: a meaningful jump in either AI or remote-work exposure predicts a fall of around 5 percentage points in the early-career share of new hires by 2025. Estimated jointly, the picture changes. Remote work holds up across every stress test they run. The AI effect shrinks sharply and, in many specifications, becomes statistically indistinguishable from zero.

A careful word, because the temptation is to over-correct the other way. This does not prove that AI has no effect on first-time hiring. It says the statistical case for that effect is far thinner than the headline implies, once you control for the thing standing right next to it. The honest verdict is not “AI is innocent.” It is “the case against AI is more circumstantial than anyone selling the headline has let on.”

The suspect who wasn’t in the building yet

Here is what the regression cannot tell you, and what should give every confident headline pause. Walk the timeline.

The sharp decline in hiring for young and upcoming talent sets in around late 2022, with no clear downward trend before it. Yet ChatGPT only arrived in November 2022. For most of 2023 and well into 2024, the tools were impressive party tricks and useful drafting aids, and also confidently wrong often enough that no sane manager would hand them a new hire’s actual workload. The agentic tools that can take a real task and run with it are a 2025 story: Claude Code landed as a research preview in February 2025 and reached general availability only that May. One more soft piece of evidence: “agents” and “agentic AI,” the very capabilities now blamed for the missing jobs, are not covered in Chip Huyen’s AI Engineering, published in early 2025, a serious and current textbook on building with this technology (this also shows how quickly things are evolving in the tech space). The vocabulary of replacement is newer than the slump it is meant to explain.

Yes, ChatGPT and the slump arrive in the same season, and that coincidence is most of why the AI story feels obvious. Coincidence in timing is not capability, though. A cause cannot precede its effect. For LLMs to have driven a decline that set in by late 2022, they would have needed a capability in 2022 and 2023 that they demonstrably did not have until 2025 (I would even debate if they have it now, but that's another topic). Something else was the primary culprit, while the tool that gets the blame was still learning to count the letters in “strawberry.”

What was actually in the building

With AI alibied out, two suspects remain, both with the timing AI lacks. The first is a hangover. Through 2020 and 2021, with money cheap and everyone suddenly remote, technology firms hired as if the conditions were permanent, fighting hardest for experienced people and padding headcount across the board. When rates rose and the growth math changed in 2022, the correction landed first on the most discretionary line in the plan: the fresh-out-of-school hire who costs more than they produce on day one. That is not AI displacing labor. That is a balance sheet unhiring the optimism of two years earlier, and the timing fits exactly, which is more than the accused can say.

The second remaining suspect is the one the data actually convicts, and it is the most interesting of the lineup, because it is not about cost at all. It is about learning.

Why remote work hits first-time hires hardest

A correlation that survives controls still needs a mechanism, or it is just a tidier coincidence. The mechanism is where this stops being an economics paper and becomes an operating problem you can do something about.

A firm hires a graduate who, on day one, rarely produces more than they cost. You hire them anyway because you are not buying today’s output. You are buying the experienced colleague they will become, and that bet pays off only if they climb the curve quickly once they are inside. A first-time hire is an investment in future surplus, and the return depends entirely on the rate of on-the-job learning.

So how does that learning happen? A second study, by Natalia Emanuel, Emma Harrington, and Amanda Pallais, observed software engineers at a large firm and measured it down to the level of individual code comments. Sitting near your teammates raises the feedback you receive by 18.3 percent, and the gain is uneven: younger, less-tenured engineers benefit about twice as much as the average. When the offices emptied, that advantage went with them, and the people who lost the most were exactly the ones who needed the feedback most. At the national scale, the same paper finds the fingerprint: in remoteable jobs, young graduates saw their unemployment rise relative to older graduates after the pandemic; in non-remotable jobs, no such gap appeared.

Proximity, it turns out, was quietly underwriting the entire apprenticeship system, and almost nobody had it on a line item. Distributed work raised the cost of developing new talent, and when something gets more expensive, rational firms buy less of it. Fewer first-time hires are not proof that the work has vanished. It is also a sign that the apprenticeship that made those hires worth the bet has become harder and more expensive to deliver.

The macro check, and the mask

If AI were truly carving out the bottom of the labor market, you would expect it in the aggregate numbers by now. Mostly, you do not. Apollo’s chief economist, Torsten Sløk, reading high-frequency US employment data, finds no clear AI job-loss signal and argues the buildout is, on net, pulling demand for workers, chips, energy, and the people who install all of it. Cheaper capability expanding total activity rather than shrinking it is the Jevons paradox: making a resource more efficient tends to increase how much of it we use, not less. The aggregate story so far is closer to “AI is creating a great deal of expensive new work” than to “AI is destroying jobs.”

So why does AI get the blame anyway? Because it is the convenient answer for a business that overhired in 2021 and would rather not admit it. “We’re restructuring around AI” is a more flattering line in an earnings call than “we misjudged the cycle.” The bot makes a better story than the binge. And the story has decades of runway: we were handed the machine-takes-over plot long before computers could write a coherent paragraph, from the Matrix to the Terminator to the AI tasked with making paperclips. Vivid, and built on assumptions about intent and autonomy that today’s systems do not possess. A model that cannot reliably book a dinner reservation without supervision is not about to repurpose the biosphere. When a mundane structural shift arrives alongside a technology we have been primed for forty years to fear, the fear collects the credit. The mask fits because we made it long ago and kept it handy. Which does not mean it hides nothing.

What is actually coming

Here I want to be precise, because the easy misreading of all this is “see, AI changes nothing.” That is not the argument, and it would be a foolish one to make.

AI will reshape work, and in places it will genuinely remove it. The rules-based, high-volume, low-judgment back-office processing that was offshored in the 2000s is now squarely in range, and that displacement is real and worth honestly planning for. Closer to home, the shift is already visible in early form: the strongest software developers increasingly do not write code line by line so much as direct the agents that write it, reviewing, correcting, orchestrating. The role did not disappear. It moved up a level of abstraction.

That is the shape of it. Not humans pushed out of the equation, but the equation rebalanced. The most powerful configuration available now, and for some time yet, is human plus AI, not AI alone: human intelligence supplying judgment, context, and direction, AI supplying knowledge at scale and tireless execution. The deployments that impress keep people in the loop and let machines extend their reach. The ones that embarrass their owners removed the human to save a salary, then discovered what the human had quietly been holding together.

So if first-time hiring in your organization has quietly contracted, the useful question is not “what is our AI policy?” It is “what happened to our apprenticeship layer?” When teams went distributed, the visible costs (real estate, commutes, coordination) got measured and managed. The invisible one, the ambient feedback that turns a graduate into a senior, did not, because it never appeared on any budget. It simply stopped happening, a comment at a time.

Rebuilding it is not a return-to-office mandate, which manages attendance rather than learning. It is a design problem: who owns the development of each new hire, how feedback gets manufactured on purpose when it no longer arrives by accident, whether proximity is engineered in concentrated doses or replaced with a substitute that actually works rather than simply mourned. The firms that solve this will hire and grow the early-career talent that their competitors have decided they can no longer afford, compounding a decade-long advantage in the cost of building people.

The ladder did not break because the rungs were automated. It broke because we stopped standing close enough to hold it, and then blamed the most futuristic thing in the room.

Images source: ChatGPT Images / Claude Opus / Gérard Métrailler

A note on method: I have not read every underlying paper cover to cover; I read the abstracts, the key tables, and the methods that carry the claims, and my AI assistant helped me pull the load-bearing figures out of the full PDFs so I could check them against each other.

Sources

Emanuel, Natalia, Emma Harrington, and Amanda Pallais. “The Power of Proximity to Coworkers: Training for Tomorrow or Productivity Today?” NBER Working Paper No. 31880, November 2023, revised June 2026. https://www.nber.org/papers/w31880

Huyen, Chip. AI Engineering: Building Applications with Foundation Models. O’Reilly Media, 2025.

Lambert, Peter John, and Yannick Schindler. “The Broken Ladder: AI, Remote Work, and Early-Career Hiring.” SSRN working paper, May 2026. https://papers.ssrn.com/sol3/papers.cfm?abstract_id=6787638 Accessed 2026-06-06

“Introducing Claude Code.” Anthropic, February 24, 2025. https://www.anthropic.com/news/claude-3-7-sonnet Accessed 2026-06-06

Sløk, Torsten. “Zero Evidence of AI-Related Job Losses.” The Daily Spark, Apollo, May 29, 2026. https://www.apollo.com/wealth/the-daily-spark/zero-evidence-of-ai-related-job-losses Accessed 2026-06-06

A trader at Jefferies, not an engineer, named the “SaaSpocalypse” panic, which tells you what kind of event this is. Software shed close to $2 trillion from its October peak on the theory that anything can now be cloned in a week. The theory is right about features and wrong about moats. The error is picturing a moat as a wall, one thing you build once and stand behind. The durable defense is a loop, and underneath it sits the part no competitor can vibe-code, because it was never in the code: judgment.

When the app you’ve hand-crafted with care is getting lost in the tsunami of new AI-generated iOS or Android apps, you get a sense that the barrier has changed. When the people repricing an industry sit on a trading desk rather than build the products, you are watching a sentiment event as much as a competitive one, and the two are not the same size. The panic says SaaS is being replaced. The shift says something narrower: the feature set, the part that was easy to defend, stopped being where the defense lives.

Concede the scary part, because it is true. A competent team with modern tooling and enough tokens can clone a single-workflow product in days, where it once took a year and six figures. And it gets worse than cloning: AI agents now walk through shallow integrations, reading across systems and eroding the switching costs on which two decades of SaaS strategy were built. The great wall you thought you had is being climbed from both sides.

The wall is the wrong picture

A wall is static, and anything static can be measured, matched, and scaled. The panic feels total because everyone is inspecting walls: this many features, that certification, this integration count. Every one of them looks vulnerable because the cloning cost of any fixed structure is approaching zero.

The companies that are not panicking have something that does not sit still: three advantages that feed each other, so using the product makes it harder to leave, which makes the next customer easier to win, which deepens the advantage again. Not a wall. A loop. It compounds while you sleep, and a competitor cannot copy a single turn of it because the value was never in a single leg. It was in the turning. Three legs, in the order the loop runs.

Workflow, where the work happens. Your product is under the most active attack, and the current debate has converged on one worth stealing: the system of record versus bolt-on distinction. A bolt-on sits beside the workflow and offers convenience, which is exactly what an agent routes around or what a weekend-vibe-coded clone matches. A system of record sits inside the workflow as the place where the work happens, the thing other tools and people organize around. You do not vibe-code your way out of the products your whole operation runs on, any more than serious workloads casually migrate off SAP, Teamcenter, or Salesforce. The test is brutal: if your product vanished overnight, would the customer’s work stop, or would they mildly miss it? Stop means you are in the workflow. Miss means you are beside it, and most products that are believed to be essential are merely present.

Trust: how you win the position and keep it. Being woven into the workflow raises the stakes, and nobody puts a prototype at the center of their operation. This is where your end-to-end Go-To-Market motion shines and makes the difference. The deeper you sit, the more the customer needs to trust what they cannot easily remove. Trust is permission to be relied on when something breaks: a SOC 2 report, a name procurement recognizes, a phone number that reaches a human accountable for the outcome. A risk committee does not approve the best demo; it approves the vendor it believes will still be answering when the workflow fails at the worst moment. That belief takes years to build and one incident to lose. This is also where partnerships fit. A reseller logo on a slide defends nothing. What a good partner gives you is borrowed trust: a business already trusts its value-added reseller and buys through it, so a go-to-market motion that rides that relationship borrows the trust the customer already extends to someone else. Partnerships are not a separate leg. They are how trust reaches a customer you have not yet reached.

Data, what the running work leaves behind. “Data is the moat” is the most recycled line in the genre, and it is half-wrong, which is what makes the other half easy to miss. a16z and others have spent years puncturing it: raw data is weak, the value of each incremental row declines, and a pile you bought or scraped offers little defense. The critique is correct, and it killed the lazy version of the claim. What it does not touch is the version that matters now. As models commoditize and intelligence becomes a utility, the scarce input swings back to the data the model never saw, and the data that matters is not bought; it is earned, accumulated only because you are the trusted system of record where the work runs, every exception and edge case deposited by use. This is where your operations layer shines. That is the turn the cliché misses: data you can buy is worth what anyone will pay for it, and data the work deposits is the one input a clone cannot start with. It begins this leg at zero and stays there until it has run real customers through real work for real years.

The delay is the moat

Now watch the whole thing turn. Workflow integration earns the right to be trusted. Trust wins and protects the integration. The integrated, trusted product emits proprietary data as the work runs; the data makes it better at the job, which deepens the integration further. No leg is the moat. The moat is the three compounding into one another, which is why a competitor who clones any single leg still loses: match what you do, certify what you are, even buy a data set, and you still do not have the loop, because the loop is made of time.

That is what the wall metaphor hides. The legs are connected by a delay. Trust accrues over years of not breaking; proprietary data over years of real usage. A competitor can ship feature parity on a Tuesday and still face a multi-year wait to spin up the same loop, and enterprise buying enforces every month of it: mid-market sales cycles run around 84 days and have lengthened, not shortened; contracts average about two years; 74% of buyers now weigh switching costs before signing, up from under half in 2018. A better product does not collapse that timeline. It waits in the same line while your loop turns a few more times. The delay is not friction in the moat. The delay is the moat.

The honest leak

A loop this tidy should make a careful reader suspicious, so here is where it leaks. Cursor, the AI coding tool, went from zero to $2 billion in annual revenue by February 2026, the fastest any business-software company has reached that mark, and it did it on word of mouth, with no marketing spend. That is the product spinning up its own loop without the trust apparatus the argument leans on, because when the buyer is a single developer who can adopt without a committee, product quality builds its own distribution, and the trust leg matters less. That edge is already eroding: with Google Antigravity, Claude Code, and OpenAI Codex attacking the workflow leg head-on, trust and data are becoming the differentiators even here.

So the loop is strongest where the buyer is an organization, with procurement, security review, integration debt, and a board asking who is accountable, and weakest where the buyer is one person clicking install. Knowing which you are is the difference between reading this as a warning and reading it as reassurance, and most companies flatter themselves on that question. (Watch, too, the newest leg some argue is forming: when value comes from many humans and agents collaborating in one system, the collaboration layer throws off network effects of its own. If it holds, it joins the loop rather than replacing it.)

The thing the loop is made of

One question is still open, and it is the one that matters most. Who decides which workflow is worth becoming the system of record for? Which incidents warrant over-investing to ensure the trust holds? Which data is worth capturing? None of those is a feature. Each is a judgment, and judgment is the one input in this picture that AI amplifies instead of replaces.

That is the foundation on which the whole loop stands. The reflexive read of AI is that it automates judgment out of the work; the truer read is that it extends judgment’s reach. A person with good judgment, well-instrumented with AI, can now apply it across a surface that used to need a team of a dozen. Execution was the scarce resource, and AI just made it abundant and cheap. Judgment about what to execute is scarce now, and AI turns it into leverage. Point AI at the abundant input, and you get a faster way to do the wrong things. Point it at the scarce one, and you get reach.

This is the final reason a vibe-coded competitor can copy your interface and still lose. It can replicate what your product does. It cannot replicate the accumulated judgment of a team that chose this workflow, earned that trust, captured the right data, watched it break, and learned what good looks like in a domain, because that judgment is not in the code. It is in the people, and in the loop they built. The companies that come through the AI transition stronger will not be the ones that automated judgment out of the loop. They will be the ones who put their best judgment where the loop turns.

What to audit before the next board meeting

The SaaSpocalypse is aimed at the wrong target. It attacks features and shallow integrations, which were never the durable moat, while the real defense was never a wall to inspect at all. It is a loop, and the question is not how high your wall is but whether your loop is turning.

So map yours. Where is your product the system of record the work runs through, and where is it a bolt-on the customer would merely miss? Is your trust earned and certified, or a slide hoping not to be tested? Where do you ride a partner’s trust? Is the relationship real, or just a logo? Is your data earned by usage, or bought and decaying? Then check the arrows between them, because a leg that does not feed the next is not part of a loop; it is a wall with extra steps. And underneath all of it: where is human judgment load-bearing, and are you investing in the people who carry it or quietly automating them away?

Then find the one place you still rely on a single static advantage a funded competitor could match in a quarter. Name it out loud. That is the part of your moat that was never a moat, because it never moved.

Images source: ChatGPT Images / Claude Opus / Gérard Métraille

Sources

• Sam Blum, “Klarna Plans to ‘Shut Down SaaS Providers’ and Replace Them With Internally Built AI,” Inc., https://www.inc.com/sam-blum/klarna-plans-to-shut-down-saas-providers-and-replace-them-with-ai.html. Accessed 2026-06-02.

• “Anthropic’s ‘Claude Cowork’ Release Triggers $285 Billion ‘SaaSpocalypse’: A Brutal Wake-Up Call for Legacy Tech and Finance,” FinancialContent / MarketMinute, https://markets.financialcontent.com/stocks/article/marketminute-2026-2-6-anthropics-claude-cowork-release-triggers-285-billion-saaspocalypse-a-brutal-wake-up-call-for-legacy-tech-and-finance. Accessed 2026-06-06.

• David Lipman, Greg Callahan, Greg Fiore, Grant Kieffer, and Ann Bosche, “Why SaaS Stocks Have Dropped and What It Signals for Software’s Next Chapter,” Bain & Company, https://www.bain.com/insights/why-saas-stocks-have-dropped-and-what-it-signals-for-softwares-next-chapter/. Accessed 2026-06-06.

• “Will Agentic AI Disrupt SaaS?,” Bain & Company Technology Report 2025, https://www.bain.com/insights/will-agentic-ai-disrupt-saas-technology-report-2025/. Accessed 2026-06-02.

• “The Empty Promise of Data Moats,” Andreessen Horowitz, https://a16z.com/the-empty-promise-of-data-moats/. Accessed 2026-06-02.

• “Good News: AI Will Eat Application Software,” Andreessen Horowitz, https://a16z.com/good-news-ai-will-eat-application-software/. Accessed 2026-06-02.

• “Software Finally Gets to Work: The Opportunity in Vertical AI,” Menlo Ventures, https://menlovc.com/perspective/software-finally-gets-to-work-the-opportunity-in-vertical-ai/. Accessed 2026-06-02.

• “Cursor in talks to raise $2B at $50B valuation after hitting $2B ARR in three years,” The Next Web, https://thenextweb.com/news/cursor-anysphere-2-billion-funding-50-billion-valuation-ai-coding. Accessed 2026-06-06.

We are reliably bad at imagining ten years out, and we miss in one direction: we underestimate. AI is the next thing we are underestimating. The instinct in most rooms is that the young will lead and everyone else will catch up. That instinct is wrong twice over, and the answer to why is older than any of the tools.

A few weeks ago, I sat at a dinner with regional leaders from government, business, and consulting. Someone said the line that stopped the room: getting on a plane, connecting to Starlink at 30,000 feet, and streaming a movie on your phone is now completely ordinary. Ten years ago, it would have sounded like fantasy.

That line points at something we keep getting wrong. We are bad at projecting ourselves ten years forward, and the error has a direction. We do not overshoot and imagine flying cars that never arrive. We undershoot. The absurd becomes the thing we do without thinking, and we forget we ever found it strange. Ask anyone in 2014 whether they would stream video from an economy seat above the Pacific, and they would have laughed.

The conversation turned, as these conversations now do, to AI. And here is the uncomfortable part: if we have always underestimated the next ten years, we are almost certainly underestimating this one. Not in the breathless way the hype merchants mean, but in the quieter way the Starlink line captures, where the extraordinary becomes ordinary faster than our judgment can recalibrate.

So the question is not whether the next decade will be strange. It will. The question is who navigates strangeness well. The room, including myself, arrived with an answer almost everyone shares: the young will lead, and the rest of us will catch up.

Digging into it deeper since the dinner, I came to the following conclusion: That answer is wrong twice over.

It is wrong about the facts first. The reflex says digital natives are the AI natives, but inside companies, the pattern inverts. A 2025 London School of Economics study found that while personal AI use is highest among the youngest workers, workplace adoption climbs with seniority and pay. The most experienced people are often using AI at work faster than their early-career colleagues, who everyone assumed were already fluent. The fluency on a teenager's phone does not automatically translate into business leverage. Seniority brings the budget, the latitude to experiment, and the judgment about what is worth pointing the tool at. That last part is the one no app teaches.

It is wrong on the framing second, and this is the deeper error. The whole premise, one group leads and the other catches up, treats a partnership as a race.

Consider what each person actually brings. The younger one arrives with fewer inherited assumptions, greater openness to AI-native ways of working, and greater comfort with uncertainty. What they lack is judgment about what good looks like, what is worth building, what will land in the real world. The more experienced one brings exactly that judgment, the pattern recognition, the instinct learned over decades, and a few scars for what is worth doing. What they sometimes lack is the willingness to pick up a tool that makes their hard-won way of working feel slow.

The blank canvas meets the seasoned eye. Neither one wins. That is the whole point.

This is a tension, not a problem to solve, and tensions punish you for choosing a side. Lean only on the seasoned eye, and you ship the same product you shipped five years ago, now with an AI label that changes nothing the customer feels. The judgment is intact, and the canvas stays blank. Lean only on the blank canvas, and you run 200 experiments a quarter without being able to name three assumptions you actually killed. The canvas fills with confident, well-formatted noise, and the judgment never arrives. Each pole, left alone, decays into the other's failure mode.

What both poles together serve is the only thing that survives the decade: value the customer can feel. Speed is nearly free now; everyone is getting faster, so faster stops being an advantage. Accuracy still needs a human because the model will hand you a fluent, well-cited, confidently wrong answer, and the cost of trusting it without judgment gets paid later in retracted work and shipped bugs. And the read on whether something will actually matter to a real person who does not live in the model at all. That is the seasoned eye's contribution, and nothing in the tool is trying to replicate it.

None of this is new. Curiosity is what lets the seasoned eye look at the new canvas without dismissing it. Openness is what lets the blank canvas accept an edit from a hand that looks slower. Systems thinking, first principles, constant learning: these are not requirements AI introduced. They are the requirements that every generation of leaders has lived by, now being run at a pace none of us has experienced. The tools changed. The fundamentals did not.

Which points at the move worth making this quarter?

Stop designing your AI rollout as a one-way class. The reverse-mentoring program where the youngest teach the most senior how to prompt is half a picture, and the half that flatters the assumption we just dismantled. Build the other half. Put senior judgment and junior fluency together as a single working unit, with one shared outcome on the wall, not a teacher and a student facing in opposite directions. The senior brings the question worth answering and the read on when the output is quietly wrong. The junior is willing to try 10 approaches in the time it would take the senior to try 1. Neither is the mentor. The pairing is the unit.

And the move is fractal. It works with one senior individual contributor and one early-career hire on a single deliverable. It works with a tenured general manager and a younger product leader co-owning a new line. It works with the board's sharpest pattern reader, paired with the company's most AI-native builder, on the AI strategy itself. It even works at the scale where the dinner began: the regions that pair the people who built the last economy with the people who will build the next one are the ones that compound. Same shape, every altitude.

Before your next leadership review, three questions. Who in your most senior ranks is genuinely fluent with AI? If the honest answer is nobody, your reverse-mentoring program is solving a problem you do not have. Who in your junior ranks has the judgment to catch the AI when it is confidently wrong? If nobody, your productivity numbers are partly fiction. And where are these two already working side by side, by accident or design, and what are they making together? Whatever it is, it is the clearest preview you have of the whole company over the next two years.

The Starlink line was meant to be about how strange the present has already become. It was really about how strange the next ten years will be, and how badly we will misjudge them if we get the pairing wrong. The blank canvas needs the seasoned eye. The seasoned eye needs the blank canvas. The room that puts them at the same table, facing the same outcome, is the room that gets the decade right.

Images source: ChatGPT Images / Claude Opus / Gérard Métrailler

Sources

• Protiviti and London School of Economics, Generations in the Workplace: AI Adoption and Generational Diversity, 2025. https://www.protiviti.com/uk-en/survey/lse-generations-survey Accessed 2026-05-31.

• Megan Leonhardt, AI's Generation Gap, CFO Brew, October 30, 2025. https://www.cfobrew.com/stories/2025/10/30/ai-s-generation-gap Accessed 2026-05-31.

By April, Uber's CTO had blown the AI budget he set in December. Three weeks later, the CEO said he was metering headcount and leaning further in. Two weeks after that, the COO asked aloud whether any of it was producing value. Three quotes, three seats, one cost category nobody had experience with. Here is why token spend breaks the quarterly cadence finance was built on, and the three questions a board should be asking by the next meeting.

Praveen Neppalli Naga, Uber's CTO, told The Information in April: "I'm back to the drawing board, because the budget I thought I would need is blown away already." Claude Code adoption at Uber had moved from 32% to 84% of a 5,000-engineer organization, from $500 to $2,000 per engineer per month.

Three weeks later, on the Q1 earnings call, CEO Dara Khosrowshahi reframed the same fact as strategy. "If every person at this company can increase their throughput by 20%, 30%, 50%, 100%, then metering headcount growth and leaning in on AI investment is going to be well worth it." Roughly 10% of code changes at Uber, he added, are now produced by autonomous agents.

Two weeks after that, COO Andrew Macdonald said the quieter part out loud to Business Insider. After talking to senior engineering leaders, he could not find a proportional link between token consumption and useful new features for riders. "That link is not there yet, right?"

Budget gone. Strategy doubled down. Value unproven. All three are right at the same time, which is the new shape of the problem. If Uber can blow through a 12-month AI budget by April and publicly disagree about whether it bought anything, the issue is not competence. The standard IT cost playbook was built for categories that do not behave like tokens.

The cost category nobody has experience with

Three structural facts make tokens different.

The first is shape. Server costs grow in steps. SaaS seats grow in contracted increments. Token spend grows on a continuous curve whose slope is set by how the workload is built, not by how many people use it. A chat interaction runs about 2,000 tokens. An agentic workflow, the kind where the model plans, retrieves, calls tools, and checks its own output, can run 1,000 times faster on the same task. Stanford's Digital Economy Lab measured a 30x cost variance running the same agent on the same task twice. The same employee can multiply the run-rate by Tuesday because someone shipped a better agent on Monday.

The second is elasticity in the wrong direction. Unit prices fell by roughly 99.7% across the leading models in 18 months, according to NavyaAI's tracking. Enterprise AI bills tripled to $37 billion over the same period. Cheaper inputs produced higher bills because adoption breadth and agentic depth expanded faster than unit prices compressed. Jevons in a hurry. The CFO who modeled a price drop into a forecast and called it discipline is already wrong.

The third is feedback latency. Finance's tightest cadence is the monthly close. Token spend moves on hours. A new model deploys on Wednesday at 3 p.m.; by Friday, the run-rate has reset; the monthly report, when it lands three weeks later, describes a regime that no longer exists. Governance whose tightest loop is the monthly close is running open-loop against the thing it is supposed to govern.

Why the obvious fixes are the wrong fixes

The first instinct, when a budget blows up, is to cap it. The second is to centralize approvals. Both push the highest-leverage decision (what is this AI doing for us, and is it worth it?) onto the people furthest from the answer.

Capping team-level spend taxes the workloads with the highest ROI. The team building an agent that replaces three vendors gets throttled to the same ceiling as the team running chatty copilots that produce nothing. The cap penalizes both, and the second one was supposed to be killed anyway.

The other temptation, the one Macdonald named when he said tokenmaxxing, is the AI equivalent of cutting marketing to hit a margin target. Drive consumption up to look AI-native, or down to look disciplined. Declare victory. Quietly stop asking what they bought. The discipline version of theatre.

What is missing in both cases is the lines on the P&L.

The line you do not have

The line that needs to exist in every company past the experimentation phase is the one that revenue already has. Owned. Forecast. Reviewed. Decomposed by group. Tied to a unit of value, the business already counts.

Four moves make it real.

Build AI consumption into the budget architecture as its own line item, not buried inside cloud or R&D. Decompose by group the way revenue is decomposed by segment.

Set the cadence to weekly. The monthly close is the wrong instrument for a cost that moves daily. The FinOps Foundation's emerging AI framework is one useful starting point; treat it as vocabulary, not as the answer.

Measure cost per outcome, not cost per token. Cost-per-token rewards are underused. Cost-per-ticket-resolved, cost-per-qualified-lead, and cost-per-shipped-PR reward workloads that earn their tokens and starve the ones that do not. If a workload cannot name its unit, it probably should not exist yet.

Make the whole C-Suite accountable. Not the CFO alone; the CFO and the head of the function whose work is being multiplied. The CFO defends the math; the operator defends the value. Neither alone is enough.

What a board should be asking by next month

If you sit on a board or run an audit committee, Uber is a free dress rehearsal. Three questions will tell you whether management is governing this category or merely talking about it.

If any answer hesitates, the company is in the same place Uber was in December. The calendar has simply not run far enough to find out.

The interesting move is not to spend less on AI, and not to spend more. It is to know, on a Tuesday, what you spent on Monday and what it bought you. Until that loop closes, the budget conversation is happening in the wrong room, at the wrong speed, about the wrong number.

Sources

• Laura Bratton, "Uber Has Already Blown Through Its 2026 AI Budget, CTO Says," The Information, April 2026. Paywalled.

• "Why Uber Has Already Burned Through Its AI Budget," AI Magazine, https://aimagazine.com/news/why-uber-has-already-burned-through-its-ai-budget. Accessed 2026-05-25.

• "Uber's COO says it's getting harder to justify the money spent on AI tokenmaxxing," Business Insider via Yahoo Finance, https://finance.yahoo.com/sectors/technology/articles/ubers-coo-says-getting-harder-050841491.html. Accessed 2026-05-25.

• Uber Technologies, Q1 2026 earnings release and call, May 7, 2026, https://www.sec.gov/Archives/edgar/data/0001543151/000154315126000019/uberq126earningspressrelea.htm. Accessed 2026-05-25.

• Sayash Kapoor, Benedikt Stroebl, Peter Henderson, and Arvind Narayanan, "How Do AI Agents Spend Your Money? Analyzing and Predicting Token Consumption in Agentic Coding Tasks," Stanford Digital Economy Lab, 2026, https://digitaleconomy.stanford.edu/publication/how-do-ai-agents-spend-your-money-analyzing-and-predicting-token-consumption-in-agentic-coding-tasks/. Accessed 2026-05-25.

• "Tokens got 99.7% cheaper. So why did your AI bill triple?," NavyaAI, https://www.navyaai.com/reports/ai-cost-report-token-prices-vs-ai-bill. Accessed 2026-05-25.

• FinOps Foundation, "FinOps for AI Overview," https://www.finops.org/wg/finops-for-ai-overview/. Accessed 2026-05-25.

Your AI tool reset now shapes your calendar. Claude meters in 5-hour windows, ChatGPT in 3. A heavy user on a Max plan can use an entire window in an hour, then wait four. This pacing splits your day into four shifts, making your meeting culture count against subscription tokens. Below is a proposed schedule for a modern knowledge worker, with three actions to try this week.

For five years, the future-of-work debate has circled around remote, hybrid, or in-office. In the age of AI, knowledge work is restructured not by location, but by when your tokens reset.

Claude and Gemini meter in rolling 5-hour windows; ChatGPT uses 3. Other tools are similar. The numbers move, but the mechanic is the point. Your day is built around an external rate-limiter you do not control. Unlike your office policy, you cannot vote on it at the next all-hands.

The crucial point is not the quantity of tokens used, but what you produce during each window. Accepting the reset window as a boundary, not a quota, naturally reshapes your day into four defined shifts. Here’s how this fits into a sample schedule for Claude users (Eastern time):

• 07:00 – 12:00. Morning window. The hardest cognitive lift. The piece nobody else can write. The CAD foundational design. The code architecture decision. The campaign concept that the team cannot crack. You arrive with a question. You leave with a draft, a model, a markup, a working version.

• 12:00 – 17:00. Midday window. Iteration. Sharpening the morning’s output. Pressure-testing assumptions. Refactoring the code. Refining the layout. Restructuring the deck. The model earns its keep as a sparring partner, not a typewriter.

• 17:00 – 22:00. Evening window. Review and tee-up. Lighter prompting. Reviewing and marking up with fresh distance. Update your AI instructions and skills from the lessons learned. Loading tomorrow’s first prompt so the next reset starts with momentum.

• 00:00 – 05:00. Overnight window. Automation. Scheduled jobs, agents, background research, draft generation, batch image renders, and overnight refactors. Your allowance does not care whether you are at the desk. This is the slot most people are still not running.

One exception worth naming. Peter Steinberger, the founder of OpenClaw, runs about 100 Codex instances against his open-source project, generating a $1.3M OpenAI bill in 30 days. Steinberger joined OpenAI in February 2026; the spend is research-funded, a deliberate experiment in what software development looks like when token economics are not a limiting factor. For the rest of us operating on subscription plans between $20 and $200, the reset window is the boundary, and the only sane sport is maximizing what we produce within it.

The visible change is the new schedule. Beneath the surface, constraints can have another dimension. For example, Claude’s recent history included a peak-hours penalty. Vendors may add similar layers if demand grows. Noticing these layers means more output for the same plan.

Here’s a move few use: using a window doesn’t require being at your desk. Launch an agent for coding, research, or audit, then step away. You can even start a task from your phone with tools like Claude Dispatch. The agent works while you’re busy elsewhere. Your standing operating review no longer leaves the model idle.

Meetings now count against your token window budget. Historically, they only impacted your time; now, status updates, all-hands meetings, and standing calendar blocks eat into your paid output time. To adapt, stack meetings during AI recharge gaps and save prime windows for deep work. The calendar should align with token windows, since these now define productivity blocks.

There’s real tension here. Over-focusing on intensity starves coordination for useful AI output. Yet too many meetings erode your productivity both morning and midday. Maintain both: pace your intensity.

Three things to try this week.

• Map your last seven days against your AI tool’s reset windows. Mark each window red, yellow, or green for whether the work inside it earned the allowance.

• Move your hardest cognitive lift to the start of your first window for one week. Notice what changes.

• Set up one overnight automation. Just one. A scheduled research scan. A draft generator that runs against your inbox. A batch render queue. A weekly digest that writes itself while you sleep. The point is not the automation. The point is feeling, once, what it is like to wake up to work that produced itself.

The remote-versus-office debate will linger as outdated arguments slowly fade. Yet watch those who’ve moved on: they notice windows with thirty minutes left, treat resets as key events, and schedule meetings during token downtimes. They ship more, better, and faster.

Sources

• Anthropic, How do usage and length limits work?, Claude Help Center. https://support.claude.com/en/articles/11647753-how-do-usage-and-length-limits-work. Accessed 2026-05-22.

• Google, Gemini Apps limits & upgrades for Google AI subscribers, Gemini App Help Center. https://support.google.com/gemini/answer/16275805?hl=en. Accessed 2026-05-22.

• CustomGPT.ai, ChatGPT Plus Limits 2026: Every Cap. https://customgpt.ai/chatgpt-plus-limits-2026/. Accessed 2026-05-22.

• "Tokenmaxxing" is making developers less productive than they think, TechCrunch, April 17, 2026. https://techcrunch.com/2026/04/17/tokenmaxxing-is-making-developers-less-productive-than-they-think/. Accessed 2026-05-22.

• Alina Maria Stan, OpenClaw creator's $1.3 million monthly OpenAI bill reveals the real cost of autonomous AI coding at scale, TheNextWeb, May 18, 2026. https://thenextweb.com/news/openclaw-peter-steinberger-1-3-million-openai-token-bill. Accessed 2026-05-22.

A friend asked me last week whether it was cheating to have ChatGPT clean up his English before he sent a client memo. He is fluent, not native; the model fixes a stray preposition, tightens a sentence, lifts the register half a notch. He has been doing this for two years. He has never asked the question out loud before.

I asked him whether he uses Grammarly. He laughed. Of course he uses Grammarly. Everyone uses Grammarly. Grammarly is not cheating; Grammarly is hygiene.

That is the entire debate, in two minutes.

The argument over what counts as acceptable AI use in writing is stuck on the wrong axis. The conversation keeps trying to draw a line between tools, with Grammarly safely on one side and ChatGPT suspiciously on the other, as if the moral status of a sentence depended on which software touched it last. It does not. Spellcheck triggered the same panic in the 1980s. Calculators in math class triggered it in the 1970s. Photography triggered it among painters in the 1840s. Each time, the line eventually moved, because the line was never in the tool.

Consider what knowledge workers actually do with AI today, or what they used to outsource to someone else. Ordered roughly by the comfort level of the people doing it:

1. Spellcheck catches typos.

2. Copy editors catch grammar and suggest phrasing.

3. Translators render thinking from one language to another.

4. Research assistants gather and summarize material you would otherwise gather yourself.

5. Thinking partners help sharpen an argument you came in with.

6. Drafting partners produce prose you then heavily edit.

7. Ghostwriters produce prose you lightly edit and sign.

8. Content farms turn a short comment into a finished post you publish.

Most people settle somewhere between items four and six, and feel queasy from item seven onward. Drawing the line at any specific item, though, produces incoherent results. Item two has been universal for a decade and nobody calls it cheating. Item seven has been universal for centuries (it has a different name, ghostwriting) and nobody calls it cheating, provided the named author stood behind it. The item is not the question.

There is a quieter version of this same conversation that almost never gets airtime. Plenty of people have something genuinely valuable to share and stay silent. The non-native English speaker who freezes at the keyboard. The expert who is brilliant in a meeting and goes blank in front of a blinking cursor. The operator with twenty years of pattern recognition who has never written a paragraph she liked. The blank page is a tax on substance. If AI helps someone with real insight get past that tax and into the conversation, the result is more valuable signal in the world, not less. The instinct to scold people for using AI to write is often gatekeeping in disguise, in favor of those who happened to be good at writing in the first place. Some of the most valuable voices we have not heard yet are the voices AI will let into the room.

So the line is not in the tool, and it is not in the writer's pre-existing fluency either. Where is it?

Stéphane Zermatten put an interesting question on the table in his newsletter Off Script: "If AI didn't exist, would this still be my answer?" It is a great question. It moves the conversation from how much AI you used to whether the substance traces back to you, which is the right axis.

I want to push it one step further. The hypothetical-absence framing implicitly treats AI as a contaminant whose removal proves purity. The bar I actually want to clear has nothing to do with whether AI was in the room. It has to do with whether what I am about to publish belongs in front of another human being's eyes. Three checks, in this order:

Is this worth sharing? Am I adding something insightful, thoughtful, or useful to the conversation, or am I producing volume? Most of what gets published, AI-assisted or not, fails this first check.

Is the substance mine? The judgment, the experience, the argument, the stakes. Not the prose. The prose is a craft layer that any number of tools can help with, including a human editor on the next desk. The substance is what cannot be outsourced without changing whose piece it is.

Will I defend it? If a reader pushes back, in the comments, in an email, across a dinner table, do I have the conviction to stand behind it on the merits? Not "I stand behind it because my name is on it," which is circular, but "I stand behind it because I thought it through, and I am willing to revisit it in public if the reasons turn out to be wrong."

There is a tension to manage in all of this, and pretending otherwise is what makes the conversation feel stuck. The two sides are leverage and authorship. Lean all the way to authorship and you are slower than every peer who learned to use the tools well. Lean all the way to leverage and you produce volume that nobody trusts, including yourself, six months in. Neither side alone is stable. The discipline is keeping both alive, knowing which one you are leaning on at any given moment, and being honest with yourself about the answer.

Honesty is the operational word. The three checks are not a courtroom standard; they are a self-honesty practice. They fail at the margins, where iteration has blurred the lines between your thinking and the model's. They work in the eighty percent of cases where the answer is obvious if you ask it out loud. The risk in avoiding these questions is not that someone else will catch you. The risk is that you will quietly stop knowing whether you can stand behind your own work.

A third voice on this is worth bringing in. Philip Moyer, Vimeo's CEO, told Nilay Patel on Decoder last year that "human curation of AI creation is going to be a necessity", framing AI as an extension that lets creators tell longer, more compelling stories, not as a replacement for the human-driven part of the work. From the platform side, looking at what creators are producing at scale, the same conclusion: the load-bearing element is the human judgment doing the curating, not the model doing the generating. The infrastructure of the next decade of media will run on that distinction.

There is a market dimension to this, separate from the personal one. As AI-only content floods the channels, the premium on legible human authorship rises. Trust is finite, and it is repricing in real time. Writers who pass the three checks, and who can credibly say I stand behind this, are accumulating something the volume players are not. That asymmetry compounds. A few years from now, the people who guarded their authorship through this period will be the people whose work still travels.

Three things to try this week, if you take the practice seriously. Run the checks on the last three things you published. If any one fails, decide what you will do differently next time, whether that is disclose, re-draft, or kill. Keep a short note for yourself on each piece: what the model did, what you did. Not for the reader; for your own honesty when you cannot remember six months later. Decide your own line, in writing, for at least three categories: research, structure, drafting, polish. The line will move. Pin a version of it now so you have something to revisit.

One last thing. This piece was written with Claude, in conversation, across more back-and-forth turns than I can count (the term of art is n-shot, as opposed to 0-shot, which is the single-prompt mental model most non-practitioners still carry of what "using AI" means). The argument is mine. The structure is mine, refined under push-back from a model that was willing to say no, that paragraph is not earning its place. The three checks are mine; Stéphane's question was the starting point. The image at the top was generated with ChatGPT, after several rounds of prompting against a brief that is also mine. Could I have produced any of this in one shot from one prompt? Not even close. Could I have produced it without AI at all? Yes, eventually, and the argument would have been the same. The prose would have been thinner, the iteration slower, and the piece would not have existed yet, because the slot in which I wrote it was four hours long, not four days.

By the three checks the piece argues for, it is mine. I signed it. I own it.

That is the only test that matters.

Citations and references

• Stéphane Zermatten, "Your Taste Is Your Moat," Off Script newsletter, May 12, 2026. https://szermatten.substack.com/p/your-taste-is-your-moat.

• Philip Moyer, interview with Nilay Patel, Decoder, The Verge, 2026. https://www.theverge.com/decoder-podcast-with-nilay-patel/616820/philip-moyer-interview-vimeo-ai-google-youtube-creators.

• Spellcheck adoption and the 1980s "crutch" debate: Sam Hartburn, "The Surprising History of Spell Checkers—and What It Means for AI-Anxious Editors," Inkbot Editing. https://www.inkbotediting.com/blog/the-surprising-history-of-spell-checkers-and-what-it-means-for-ai-anxious-editors.

• Calculators in the classroom and the 1970s controversy: Audrey Watters, "A Brief History of Calculators in the Classroom," Hack Education, March 12, 2015. http://hackeducation.com/2015/03/12/calculators.

• Photography's arrival and the painter response in 1839: Pamela M. Henson, "Photography Murdered Painting, Right?," Smithsonian Institution Archives. https://siarchives.si.edu/blog/photography-murdered-painting-right.