The Microsoft outage on July 19th, 2024, was a wake-up call for all of us, showing just how vulnerable our tech infrastructure can be. The disruption, which varied in duration, threw businesses and individuals into chaos throughout the day.

So, what exactly happened? A major system failure, which means a complete breakdown of critical computer systems, caused widespread Blue Screen of Death (BSOD) issues, leading to server crashes and connectivity problems. The root cause was later identified as an update to CrowdStrike’s antivirus software, designed to protect Microsoft Windows devices from malicious attacks. This flawed update caused significant disruptions, impacting Windows PCs but not other operating systems. The disruption was worsened by an unrelated critical flaw in Microsoft’s cloud service infrastructure. 

This IT outage didn’t just cause a few hiccups; it had a ripple effect across different areas. Over 3,300 flights were canceled globally, with long queues and delays at airports like London’s Stansted and Gatwick, as well as Tokyo, Amsterdam, and Delhi. In the US, airlines such as United, Delta, and American Airlines grounded their flights. Australian carriers Virgin Australia and Jetstar also experienced cancellations and delays.

To make matters worse, the outage also hit payment systems, banking, and healthcare providers worldwide. In Alaska, Arizona, and Oregon, the 911 emergency service was affected, and Sky News was off the air for several hours. Businesses faced substantial downtime losses, supply chains were disrupted, and the market got a bit jittery. On the societal front, crucial services like healthcare, education, and government experienced delays and interruptions. The outage even threatened payroll processing, potentially delaying employee payments.

Technologically, the incident raised serious cybersecurity concerns, and this whole episode was a stark reminder of the need for a more robust and resilient tech infrastructure.

But you may be thinking, how is this relevant to AI?

There are many lessons we can draw from this worldwide outage, from potentially using AI to prevent similar outages to implementing better practices surrounding AI to prevent AI-related issues that could cause downstream outages and havoc. 

Let’s explore 5 lessons we can take away from this problem. 

5 Key Lessons from the Microsoft Outage

Lesson #1: The Importance of Redundancy in AI Systems

Redundancy is crucial for creating resilient IT systems, whether or not AI is involved. Think of it like having a spare tire: if one tire fails, the spare can take over until the main ones are repaired. The recent Microsoft outage highlighted this necessity when a problematic software update disrupted systems globally, underscoring the importance of having backup plans in place.

Figure 1: Illustration of data redundancy in a server cluster, showing how multiple copies of data are stored across servers to ensure availability and protection against server failures. Source: https://www.high-availability.com/articles/system/hardware-redundancy

Take Delta Air Lines, for example. When the CrowdStrike update knocked out critical systems, Delta’s lack of redundancy in crew scheduling caused a domino effect. Their primary system for crew tracking failed, making it impossible to reassign crews and operate flights smoothly. This resulted in over 3,300 flights being canceled globally, with thousands of stranded passengers.

By incorporating redundancy into their AI infrastructure, particularly for critical systems like crew scheduling, airlines like Delta could have significantly mitigated the impact of the outage. For example, imagine having a limited number of computer systems that run purely on a Unix-based operating system as opposed to purely Microsoft-dependent systems. Having a more basic implementation of the scheduling system run on a Linux-based environment could have meant on such rare but potentially damaging occasions, systems can still operate albeit slightly slower. 

Redundancy in AI

Similarly, in the context of AI, redundancy is equally important. Imagine having an AI-driven chatbot that’s 100% dependent on ChatGPT. If ChatGPT is experiencing a failure, how would you continue servicing customers while allowing the model vendors to get their act together? From having fallback rules-based systems in place to leveraging backup models, AI systems can continue to operate as long as you’ve incorporated mechanisms to avoid single points of failure. 

The lesson learned? Redundancy is an investment that pays off in the long run. It helps organizations weather unexpected storms and ensures continued operations even when disruptions occur.

Lesson #2: The Need to Incorporate Predictive Maintenance 

Predictive maintenance is all about anticipating problems before they happen. By analyzing heaps of data, AI can spot patterns that indicate potential equipment failures or system breakdowns. This proactive approach is a game-changer for businesses looking to avoid unexpected downtime.

For example, BlueScope, an Australian steel manufacturer, has achieved notable efficiency improvements through Siemens’ predictive maintenance solution. By leveraging AI to anticipate equipment failures and schedule timely maintenance, BlueScope has significantly reduced unplanned downtime and maintenance costs. This proactive approach has led to smoother operations and better overall equipment effectiveness, ensuring more consistent production and less disruption.

In the case of the recent outage, AI could’ve been used to predict if a rollout should happen or not. For example, an AI system can be trained to analyze system logs, previous rollout data, historical quality assurance data, historical failure patterns, backward compatibility data, and other relevant information and spot potential problems before a rollout. In CrowdStrike’s case, an AI system could’ve spotted something fishy with the update, prompting the CrowdStrike team to investigate the issue before authorizing a rollout. 

Getting Started With Predictive Maintenance

Since predictive analytics is AI’s strength in the business world, why not take advantage of it to safeguard your business? However, to implement predictive maintenance effectively, you’d need to start by gathering data from your daily business operations. For instance, if you aim to prevent outages in your web systems, you should collect systems data during uptime and downtime, along with relevant hardware parameters. This data is crucial for training AI models to anticipate potential issues before they arise. 

There are many other ways to leverage predictive maintenance. So start by figuring out points of failure, what data is being collected in those instances, and what other sources you should track and collect to make predictive maintenance possible.

Lesson #3: The Need for Rigorous Software Testing – AI or Not

Testing is a critical part of any software development and release. CrowdStrike’s problematic testing process, which was found to have missed several key aspects, resulted in this major outage. This raises serious concerns about their testing protocols, especially because their software has access to Microsoft’s kernel-level code. This type of code interacts directly with the operating system, so any malfunction can cause severe system instability. This incident highlights the critical need for comprehensive testing to prevent significant disruptions and ensure system reliability.

Similarly, in the AI world, testing is equally critical and I talk a lot about it in my book and in the article here. The stakes are high and the complexity is immense. AI systems operate on a data-driven basis and produce probabilistic outputs, meaning their behavior is influenced by the data they are trained on and the dynamic nature of the data that they process. This makes testing even more crucial, as AI outputs are never 100% accurate. 

Testing AI Systems

Consider our historical Microsoft’s AI chatbot, Tay, as an example. Launched on Twitter, Tay quickly began spewing offensive content, likely as a result of ingesting bad data in her interactions with other users. It’s unclear how much testing went into the AI initiative before Tay was released, but it clearly wasn’t enough. Had Tay been subjected to different types of tweet engagements during testing, the dangers of releasing Tay would’ve been obvious.

Figure 2: Example of Tay’s Tweet. Source: https://arstechnica.com/

Offensive output is not the only risk of inadequate testing of AI systems. Now, there is a whole movement around that’s “against AI” simply because AI outputs can be unreliable, where systems can hallucinate and confabulate, plus there’s a risk of AI systems becoming biased in unexpected ways. Further, the data that feeds into some of these AI systems raises many concerns from an ethical and legal point of view. All of these need to be taken into consideration when evaluating AI systems. 

Figure 3: An illustration of the software testing process from plan to launch. Source: https://www.specbee.com/blogs/software-testing-standards-and-processes

To keep AI systems running smoothly and, more importantly, predictably, thorough testing is key. Here’s a quick rundown of what that looks like:

Scenario Testing: Think of this as trying out different “what if” situations. For a customer service chatbot, you’d test how it handles everyday questions, tricky problems, and unexpected queries. This way, you ensure it’s ready for anything users throw at it.

Stress Testing: This is like pushing the AI to its limits to see how it holds up from different angles. For example, if you take the volume angle, you’d simulate a traffic surge. If you take the accuracy angle, you’d feed the AI system 85% of the inputs it would generally see and 15% with the weirdest of inputs it would rarely see. The goal? To make sure the AI system doesn’t buckle under pressure and keeps running smoothly.

Ethical Testing: Here, you’re making sure the AI plays nice and follows the rules. This means checking for any biases in its decisions, ensuring it handles personal data properly, and making sure it doesn’t produce harmful content.

By running these tests, you ensure your AI systems are reliable, efficient, and up to snuff on ethical standards.

Lesson #4: The Need for Real-Time Incident Response

When a tech crisis strikes, being able to react swiftly can make all the difference. When a problematic update from CrowdStrike caused widespread service outages, Microsoft’s Troubleshooting Center struggled to handle the flood of requests. This delay only made things worse, showing just how critical it is to have a real-time incident response plan.

Let’s talk about how AI can step in to save the day. AI systems can continuously monitor IT systems, process a high volume of systems alerts and alarms, and analyze and categorize the flood of alerts and alarms using machine learning, focusing on the most critical problems based on their potential impact and past data. This means that with AI, you get not only instant alerts but also prioritized insights, enabling swift action on the most pressing issues. On top of that, an AI system can also analyze relevant alerts and underlying data to get to the root cause of the problem. This is a key functionality of many AIOps applications.

Figure 4: An example of how AI-powered AIOps tools like ManageEngine turn raw data into insights for IT management. Source: https://www.specbee.com/blogs/software-testing-standards-and-processes

In the case of the Microsoft outage, having an intelligent incident response system could’ve helped IT teams around the world to quickly identify the reason for Windows machines crashing, in this case, a faulty update from CrowdStrike. Further, the AI system would’ve suggested a suite of corrective actions. And the icing on the cake? The AI system could’ve autonomously or, with the help of IT teams, quickly reinstated the best working versions of an organization’s software systems and placed a moratorium on subsequent updates until problems were rectified. 

By integrating such AI tools for monitoring, automating responses, and real-time reporting, you can better protect your organization’s IT infrastructure and minimize the impact of unexpected disruptions.

Lesson #5: The Need to Future-Proof AI Systems

Future-proofing IT systems, including AI systems, in anticipation of potential problems can ensure that business systems continue to deliver long-term value with minimal risks.

Let’s discuss some potential challenges in AI that might prompt you to consider future-proofing your AI systems. One big threat in AI is cyber security attacks, such as data poisoning. Imagine hackers tampering with the data your AI learns from, causing it to (a) learn the wrong things and (b) make all sorts of mistakes. The inference data can also be targeted and tampered with, leading AI models to make the wrong decisions.  

For instance, attackers could manipulate images used by self-driving cars by adding fake traffic signs or removing real ones, potentially causing the vehicle to misinterpret its environment and cause accidents.

Another big problem that AI can cause is bias and fairness issues, where AI systems can inadvertently be biased against subgroups of people. This was clear in Amazon’s recruiting tool, which dismissed qualified women candidates likely because its training data contained high levels of gender bias due to historical hiring patterns. The AI tool learned that male candidates were preferred because that’s what the data showed.

To build resilient AI systems, to start with, we need to focus on the AI training data. Training data should not just be of high volume but also represent highly diverse samples. Also, the attributes in the data that could result in unwanted biases should be actively tracked and adequately ‘cleansed’. Your training data should also account for adversarial patterns that can trick AI systems into doing unexpected things.

Accounting for potential cyber-attacks can also help you determine how best models should be designed, developed, and deployed. For example, should an AI system be the sole decision maker or a second opinion? Should your AI system be deployed as a microservice over the internet or within a private infrastructure? The stakes are much higher when your AI system is the sole decision maker and accessed over the internet. Security loopholes that may allow unauthorized access to your models and the underlying training data and related datasets, all need to be accounted for.

We also need to put AI systems through some serious testing. Speaking of testing, remember in lesson 3, where we talked about the importance of rigorous software testing? For AI, this means running various stress tests, including ones that simulate adversarial attacks. 

Beyond testing, staying up-to-date with the latest AI threats and vulnerabilities is key. This proactive approach helps you tackle potential issues before they become major problems, ensuring your AI systems stay resilient and reliable over the years.

Summary and Key Takeaways

In this article, we explored the significant lessons from the recent Microsoft outage and how they apply to both IT and AI systems. Here are key lessons for building resilient AI systems and using AI as a tool to build IT infrastructure resiliency:

• Redundancy is crucial: Having backup systems in place is essential for resilience. For AI, this means ensuring there are fail-safes and backup models to keep operations running smoothly, even if one system encounters issues.

• Predictive Maintenance: AI can play a key role in predicting and preventing problems before they occur. By analyzing data and spotting potential issues early, you can avoid disruptions and keep your systems efficient and reliable.
  

• Rigorous Software Testing: Thoroughly testing software—including AI-powered software—is critical in avoiding downstream issues. This includes scenario testing, stress testing, and ethical testing to ensure your AI behaves as expected under various conditions.

• Real-Time Response is Key: Having systems in place to detect and respond to problems in real-time can prevent minor issues from becoming major crises. AI-driven monitoring and automated responses can significantly enhance your ability to manage incidents swiftly.

• Future-Proof Your Systems: Anticipate and prepare for future challenges, including security threats and biases in AI. Stay updated on potential vulnerabilities and continuously refine your systems to ensure they remain robust and effective.

That’s all for now!

One of the biggest retail giants, Walmart, has been at the forefront of leveraging AI in innovative ways to improve customer service, personalize shopping, and even increase associate productivity.

In fact for those of you who have read my book, “The Business Case for AI,” in chapter 2, I discuss how Walmart’s e-commerce search uses a suite of AI solutions aimed at achieving the following objectives:

(1) Maximizing the relevance of searched items

(2) Maximizing revenues while maintaining relevance

(3) Providing a seamless user experience

What I love about Walmart’s AI strategy is its purpose-driven nature. Instead of using AI for AI’s sake, there are many use cases where Walmart uses AI for true business benefits.

In this article, we’ll examine three other ways Walmart integrates AI into its operations and extract key learning points from each.

1. Chatbot to assist customers with their order

Purpose: reduce time to servicing customers, reduce employee workload

When customers have a problem with their order, they often want it sorted out immediately. To this end, Walmart has been using NLP/NLU chatbot technologies to answer customer questions related to the status of orders, process returns, and more.

Source: https://tech.walmart.com/

As a result of this initiative, Walmart has nixed millions of customer contacts, reverting the easier questions about order status and returns to its chatbot. This has allowed Walmart’s support associates to focus on the trickier service issues.

Furthermore, through customization of the chatbot technology, this chatbot handles localized languages in various countries, such as Mexico, Chile, and India, and also understands associated contexts and catalogs. Since customers are getting more immediate service, they saw customer satisfaction scores (CSAT) increase by 38%.

What can you learn from this customer service chatbot?

1. Simplicity & risk mitigation: As seen in the image below, instead of responding like your best buddy, Walmart’s chatbot provides fairly structured responses, limiting the chances of providing erroneous and superfluous responses. It sticks to the point, and the burden is primarily on understanding the customer’s question and assisting the customer instead of trying to sound human.
   
   Moreover, even if the chatbot misinterprets the customer’s question, its structured output reduces the likelihood of leading customers into undesired actions. Any mistakes in understanding the query would be more apparent to the user.

2. Build once, use repeatedly: Although this customer service chatbot was initially built for customers in the US, the company was able to customize it for localized uses throughout the world. Sharing the underlying chatbot technology means extreme cost savings for the company, as the work done for each localized situation is limited to fine-tuning and customizing the tech for localized uses. It’s much easier to maintain and improve a single chatbot technology instead of ten different ones.

Walmart’s customer service chatbot provides predictable responses

2. Generative AI-powered search function

Purpose: personalized results, and a more streamlined shopping experience

Walmart has implemented a generative AI-powered search function on its mobile app and website. This feature is designed to understand a customer’s search context and provide personalized search results.

For instance, if a parent is organizing a minions-themed birthday party for their child, they don’t have to perform multiple searches for minions-themed items. One for balloons, one for cake toppers, one for plates, etc., which is what they will have to do with regular keyword-based e-commerce searches. Instead, with this new GenAI experience, they can just pose the query, “Help me plan a minions-themed party for my daughter” and they’ll see products relevant to their goals.

Generative AI simplifies the retail search process, transforming shopping from scroll-driven to goal-driven. This makes the digital shopping experience more streamlined, potentially reducing the number of searches, clicks, and page navigation needed to find items. While maintaining revenues, this efficiency could even result in a revenue boost due to an improved shopping experience.

Walmart’s search results for “Help me plan a minions-themed birthday party for my daughter”

What you can learn from this GenAI-powered retail search:

1. Subtle but clever: Unlike many AI initiatives that try to change their user experience by introducing new chatbots in co-pilot style, which disrupts the customer’s old workflows, Walmart very subtly provides a chatbot-style experience in its search functionality. This means that customers can continue using their search function as they always have and reap these additional context-aware search benefits. ****

2. Incremental complexity: Walmart’s generative AI experience currently supports context-aware searches to provide personalized results. However, this capability can grow with time, allowing users to perform increasingly complex functions right from the search bar.
   
   Imagine being able to accurately place an order for several grocery items with a single search request that includes the quantities of those items. That’s a solvable use case. But, instead of releasing a firehose of such features, Walmart is incrementally increasing its GenAI capabilities, which is most likely to limit risks and ensure that trust remains high with customers.

3. AI-Powered Store Advisor

Purpose: waste reduction, and revenue loss prevention

Walmart is piloting a new AI solution aimed at minimizing food and fashion waste. This waste leads to millions of tons of unsold items each year, potentially ending up in landfills and costing Walmart billions in losses.

The initiative involves an in-store AI system that advises employees on managing product ripeness, seasonal fashion sales, and waste reduction strategies. For example, Walmart’s AI technology enables staff to scan items such as bananas to assess their ripeness. Subsequently, utilizing AI, a digital dashboard will propose actions for the product.

Depending on the product, the AI might suggest adjusting the price, returning the product to the vendor according to policy, or recommending donation as the optimal course of action based on the analysis. The AI tool helps employees make proactive decisions based on data insights, potentially reducing waste.

It’s poised for a pilot launch in Canada and has plans for global expansion. Walmart’s commitment to waste reduction aligns with its goal of eliminating operational waste across its North American operations by 2025, reflecting a broader industry trend towards sustainability and waste reduction.

What you can learn from this store advisor initiative:

1. Putting company-owned data to work: Walmart is able to build such a powerful tool not because it has special knowledge in AI techniques but because it has collected large volumes of relevant historical data. For example, to build such an AI advisor, at a bare minimum, Walmart would need historical data on product conditions, the actions that were taken for the products in those conditions, and the corresponding outcomes.
   
   Without such company-specific data, you can only get away with generic applications. And the interesting thing is that for a solution like this, you don’t even need sophisticated AI algorithms. Simple recommendation algorithms will work equally well so long as your data is reflective of the behaviors you want to mimic.

2. Boring but essential: While this AI advisor might not seem as flashy as a chatbot that covers a range of topics or one that attempts medical diagnosis, its role is vital in achieving sustainability objectives, significantly reducing waste, and boosting the company’s revenue. It’s a win-win-win on all fronts. These types of application opportunities often remain hidden until you start proactively finding employee and customer pain points and shortlisting those that are suitable for AI.

Final Thoughts…

Unlike many companies that use AI for AI’s sake, Walmart is a notable example of a company that applies AI where it truly matters while also adopting a risk-averse approach. The company progressively introduces more advanced AI features into existing AI-powered solutions and subtly incorporates AI technologies without causing disruption to the customer experience, as was seen in the chatbot and the generative AI-powered search example.

Importantly, as we saw in the store advisor example, Walmart creatively leverages its vast data resources to create powerful customized AI solutions. These solutions may address issues that don’t appear exciting on the surface but are actually crucial to the organization’s operations.

Now the question is, how are you going to take a strategic approach to AI integration? Which of the points outlined above will you attempt to mimic? Finding unknown but lucrative AI opportunities? Subtly integrating AI instead and progressively improving it? Significantly reducing the risks of generative AI mishaps, by taking creative countermeasures? Start with one.

That’s all for now! This article was originally posted here.

Debunking Common GenAI Myths In Business AI Implementation

It is a common belief that GenAI resolves many issues associated with business AI implementation.

It demands less development time, is much more accessible to developers and business leaders alike, and GenAI models think like humans.

Easy, isn’t it?

Only that companies that began experimenting with GenAI a year ago are still grappling with it today. Companies I know that have been on the right AI implementation path are suddenly calling for help with their GenAI implementation.

And the amusing thing is that many of their issues are not technical, nor are their “AI problems” unsolvable.

Much of the issue is related to how the companies are treating and evaluating GenAI applications, made worse by the distracting information on social platforms, which is guiding these teams in the wrong direction.

In this article, we’ll focus on the three common myths surrounding GenAI that we’ve seen preventing companies from making positive progress in their AI journey.

3 GenAI Myths in Business AI Implementation

Myth #1: Software engineers are all we need

Ideate. build. deploy. That’s currently the standard thinking when it comes to building applications on top of Generative AI models such as LLMs. I can’t fully blame these teams, as vendors are selling their foundational models as such, insisting that building applications on top of LLMs is extremely straightforward. No AI expertise needed.

Because of this, many companies have resorted to leveraging their software engineers to build complex AI applications. This is ok, if you’re building a prototype, building tools for internal use cases or the risks when the AI system fails is not that significant.

In reality, expertise is crucial when building enterprise-scale applications —both AI expertise and subject matter expertise. These experts can help truly evaluate the performance of AI solutions, develop scalable solutions that work on more than 2-3 test cases, and continually monitor and enhance the quality of your AI-powered applications.

I’ve worked with many software engineers who have tried to build full-blown AI solutions but often struggle to scale these applications because of the lack of foundational data science thinking and skills.

Myth #2: “Data Strategy” is a hoax

“Now that GenAI is here, it’s time to ditch our data strategy.”

This is common thinking amongst executives, some of whom I’ve had to correct politely.

To be fair, training data is optional for specific applications where GenAI works out of the box. However, for most large-scale enterprise applications, especially in niche domains such as healthcare and the government space, LLMs need fine-tuning to accomplish specific tasks at a reasonable accuracy.

So, to fine-tune, you’d need data. To generate data, you’d need to know what data to generate or collect in the first place and in what format. Above all, you need a strategy to support long-term data collection for various projects within the organization. Which data collection activity do you prioritize?

Further, data collection can take considerable time, depending on the rules and regulations surrounding specific data sources. In some cases, multiple data sources may need to be merged to acquire the data needed to support a project. Or new user interfaces may need to be developed to collect data directly from users.

All this to say, data collection, sharing and warehousing requires careful planning and prioritization and thus the need for a solid data strategy. I’ve seen talented data scientists and analysts leave their jobs at companies that don’t have data for them to access and use in their day-to-day work. Many AI projects have been significantly delayed due to the lack of a data collection plan. So why wait?

Some of the elements that will make up your data strategy.

Myth #3: GenAI applications are “agentic” by default

Agentic AI is a type of AI that can understand complex tasks and work on them by itself, with little to no human help. It works like a human, understanding instructions given in simple language, setting goals, completing smaller tasks, and changing actions based on the situation.

When you think in terms of workflows in your business, such as finding relevant candidates in recruitment, this process may start from downloading relevant candidate resumes, reading and extracting content from resumes, filtering down resumes relevant to the jobs at hand and suggesting the best candidates for screening. This is what a truly Agentic AI is capable of doing without much supervision.

However, such an AI system is natively non-existent within LLMs. If vendors promise such systems, they’ve either trained these systems to perform some of these tasks OR, they’ve manually broken down the subtasks, used AI in the parts of the workflow where needed, and used other approaches where AI doesn’t fit.

So, what’s the big deal?

It is a big deal. Teams are getting extremely confused about how to identify AI-worthy opportunities, regardless of industry. They end up labeling entire workflows as potential AI projects when, in fact, only small parts of their workflow can be automated with AI. Such struggles are preventing these teams from making progress in project execution as they’re endlessly looking for GenAI systems that can automate entire workflows.

This type of agentic or RPA-like thinking, that AI can automate entire workflows, is brought upon by exposure to ChatGPT and Copilot, which seem to complete multiple tasks at a time.

In reality, ChatGPT-type LLMs, under the hood, are fine-tuned with hundreds and thousands of examples of how to complete and break down specific tasks that the user provides. So, while it is seemingly breaking larger and more complex problems down into subproblems on its own, in many cases, it actually already knows how to accomplish this through all the training it has received over time.

Key Takeaways

1. Software engineers alone are not enough to build robust GenAI applications; AI expertise and subject matter knowledge are critical, especially for public-facing or customer-oriented applications.

2. Despite the advent of Generative AI (GenAI), a comprehensive data strategy remains vital for long-term AI readiness, particularly for large-scale enterprise applications.

3. GenAI systems, while appearing autonomous, are essentially task-oriented and require specific training to perform multiple tasks. They are not inherently “agentic” or capable of breaking down and accomplishing complex tasks without human intervention in custom workflows.

Not Sure Where AI Can Be Used in Your Business? Start With Our Bestseller.

The Business Case for AI: A Leader’s Guide to AI Strategies, Best Practices & Real-World Applications.

In this practical guide for business leaders, Kavita Ganesan takes the mystery out of implementing AI, showing you how to launch AI initiatives that get results. With real-world AI examples to spark your own ideas, you’ll learn how to identify high-impact AI opportunities, prepare for AI transitions, and measure your AI performance.

READ NOW!

“AI development is a subset of software engineering!”

A manager I once worked with made sure he drove this point home.

None of the data scientists on the team appreciated this statement. They viewed machine learning and data science as forms of “science” and focused on scientific rigor rather than the quality or functionality of their code.

The manager had good intentions. He meant that we needed rigor at the source code level, which involves documenting code, unit testing, and modularizing code, on top of ensuring that the model is doing what it was meant to do.

The reality is that while AI development shares many aspects with sound software engineering practices, it has a slightly different development lifecycle. This article will explore the key differences between AI/ML development and traditional software engineering.

It is commonly thought that AI is a subset of software development. In reality, AI development shares many aspects with sound software engineering practices, but it has a slightly different development lifecycle.

Key Differences Between AI Development and Software Engineering

We will now examine six key areas where AI development diverges from traditional software development. The image below illustrates the two development lifecycles for comparison, and you’ll start noticing some overlaps and differences.

Traditional software development lifecycle vs. AI development lifecycle

1: AI Systems Thrive on Knowledge From Subject Matter Experts

While traditional software development primarily relies on software engineers, product managers, and QA personnel, having a subject matter expert (SME) in the loop is critical for AI development as AI systems attempt to mimic the SME’s thinking and decision-making.

By subject matter experts, I’m referring to individuals who have a deep understanding of the problem that the AI solution aims to address. These experts can typically solve the problem with high accuracy. For example, if it’s a customer service problem, the SMEs could be a group of service operators. If it’s a medical diagnosis tool, the SMEs would be a group of physicians or specialists in the field.

Correctly translating knowledge from these SMEs into AI models is crucial for the initiative’s success.

Imagine developing an AI model to predict a person’s eligibility for a home loan. Knowledge from the SME can guide the feature engineering process (signals for a model to learn from), the AI techniques to employ, and even the interpretation of results. These experts can provide invaluable insights into domain-specific nuances that can impact the performance of the AI solution.

For instance, a specific group of people may appear ineligible for a loan based on the available data. However, a distinguishing factor not clearly represented in the data might make them eligible. Incorporating this information into the development process can enhance accuracy.

This doesn’t mean that SMEs are irrelevant in traditional software development. However, their involvement is usually more at the functionality level. Essentially, the question is: Does the tool work as it was designed (in traditional software development), or does it think like a human (in the context of AI development)?

2: AI Systems Start and End With Experimentation

In a typical software engineering project, we frequently use a software library that somewhat addresses a particular task. For instance, if we need a software library to read and parse a file, we typically select the first one that works and proceed. This process doesn’t involve much “experimentation.”

The same does not apply to AI model development. You cannot simply use the first Large Language Model (LLM) or open-source solution that you think works and then move on.

As AI systems learn to mimic human-like thinking by making sense of data, they’re only able to complete tasks with a certain degree of confidence. In fact, models are really just making educated guesses.

As a result, the output quality can dramatically vary from model to model, even within the same class (think: Claude vs. ChatGPT) and from approach to approach (think: LLM vs. non-LLM based approaches) on the same task.

So, to develop a high-quality, deployment-ready solution, developers need to experiment with different tools, techniques, and models to find one or more that fit the bill.

Not only do you have to find the best model or approach for the AI solution, but once development is in progress, further experimentation is a must to ensure the results meet the application’s needs and reach a trade-off between accuracy and deployment timelines (see point 5).

3: AI Systems Are Heavily Data-Dependent

One of the first questions a machine learning engineer will ask you if you approach them with an AI project is, “Do you have the data necessary for the project.”

One of the last things that companies often have is data! Companies have good ideas, reliable subject matter experts, machine learning engineers, business stakeholders, and even users, but they often lack good, usable data.

As Matei Zaharia, the creator of Apache Spark, often says, “It’s your data, stupid.” This is in reference to why models don’t behave.

Data is at the front and center of AI development, as can be seen in the AI development lifecycle. General software engineering, on the other hand, primarily depends on the logic embedded in the code. If you have a big software development project with an AI component somewhere in there, that component would require data-driven development.

Even if you think you’re just going to prompt GenAI, truth be told, you need data to evaluate and fine-tune your solution. Otherwise, you can expect to be a household name due to AI gone wrong dramas, as will be discussed in point 5.

4: AI Systems Require Iterative Improvement

A recent client shared a mild AI mishap with me. They had hired a contractor to develop a sentiment analysis solution using ChatGPT. They integrated this solution into their main product, assuming the job was complete. However, a few months later, they had to rehire the contractor.

Why?

Upon launching the AI solution, the company discovered that their initial assumptions were incorrect. The output required adjustments to align with the product’s needs. Consequently, the contractor remained on the scene for several months until the solution fit correctly within their application.

Indeed, it’s not a myth that AI requires iterative development. If you look at the AI development lifecycle, iterative development is crucial during model development and testing. Iterations may also be required long after the AI model is deployed, especially when the data or expected model behaviors change.

When it comes to conventional software development, the requirements are actually more straightforward. As long as you stick to what the application needs, you can develop, test, and push code more swiftly into production systems.

However, with AI, you’d develop, test, fine-tune, validate, and then reiterate the process until you achieve a satisfactory level of performance and the output feels just right. This iterative process is inherent to AI development and contrasts with the more linear approach of traditional software development.

5: AI Systems Need Quantitative & Qualitative Evaluation

If you remember, the Tay twitter bot story from my book, or if you followed the Air Canada chatbot story where the Chatbot confidently promised a bereavement reimbursement

During development and post-development testing, you’re essentially stress-testing and playing “mind games” with the AI system to find its points of failure using quantitative and qualitative approaches.

AI failures are not only a media talking point, but in some applications, they can cause harm to people’s lives and livelihoods, such as in medical diagnosis and loan assignments. While AI systems will never be perfect, their behavior should be predictable, they should be free of known biases, and their points of failure should be well understood, documented, and communicated.

In contrast, testing, in the typical software development sense, is all about checking whether the application works as planned, often by passing a series of tests known as a test suite. This is also applicable to an AI tool, but in addition to functionality testing, you’re also performing quality testing. Consequently, it’s critical to have appropriate evaluation metrics and processes in place to assess the quality and accuracy of AI outputs.

6: AI Systems Must Be Monitored

Unlike traditional software systems, which, once tested and deployed, require minimal continuous monitoring, AI solutions require continuous monitoring due to their probabilistic nature.

Model degradation (left). Periodically refreshed model (right).

The performance of AI models can degrade over time if the data they are interacting with changes. This phenomenon, known as “model drift,” necessitates regular checks and potential retraining of models or data quality improvement to ensure they continue to perform optimally.

In contrast, traditional software systems, once developed and deployed, tend to be more “stable” as their outputs are more predictable, requiring less continuous output monitoring. The monitoring for these systems is more focused on availability and unexpected errors.

AI vs. Software Engineering: Summary

While AI development and software development intersect a great deal, in principle, the process of developing scalable and high-quality AI solutions is different from traditional software engineering.

The table below visually summarizes some of the key differences.

AI Development vs Traditional Software Engineering: Key Differences

Remember, understanding these fundamental differences between AI and traditional software development is essential for the effective planning, execution, and management of AI projects.

That’s all for now!

Not Sure Where AI Can Be Used in Your Business? Start With Our Bestseller.

The Business Case for AI: A Leader’s Guide to AI Strategies, Best Practices & Real-World Applications.

In this practical guide for business leaders, Kavita Ganesan takes the mystery out of implementing AI, showing you how to launch AI initiatives that get results. With real-world AI examples to spark your own ideas, you’ll learn how to identify high-impact AI opportunities, prepare for AI transitions, and measure your AI performance.

SEE ON AMAZON

I use LLMs regularly. But I use prompting occasionally. 

So, is prompting the only way to leverage the power of LLMs?

When I speak to marketers, journalists, podcast interviewers, and people who have used ChatGPT in some capacity, they tend to think that prompting large language models (LLMs) is the primary way to use AI to accomplish various automation tasks.   

However, in my work with clients, when it comes to designing and building out scalable, intelligent applications, many of the projects we’ve tackled use LLMs…but often not through prompt engineering. 

Even if we do use prompts, it’s often not for the purpose of completing the final task.

Let me tell you three ways we use LLMs in our work with clients and how you, too, can think about ways to leverage LLMs that go beyond prompting.

1/ Training Data Generation

Many of the startup clients we work with have little to no data, and the larger companies often have data in an unusable format or have not started collecting data for their intended AI application.

However, to build scalable domain-specific applications, we need data…lots of good quality data. 

As a simple example, to train a model that can perform company-specific categorization, such as if the given user question fits in one of 10 different company-specific topics, we need hundreds of examples of question-topic pairs.

Example of question-topic pairs

Although for these questions, you can quickly prompt LLMs to obtain accurate categorization, the more specific the topics become, or the more sophisticated and domain-centric the questions become, the harder it will be for LLMs to get things right without adequate domain knowledge. As is, we often see 60-70% accuracy on domain-specific categorization problems. 

Prompting ChatGPT to classify questions

So what do we do? We fine-tune the LLMs to improve their accuracy. 

This is where training data becomes crucial. 

Unfortunately, we may not have access to a large number of human-curated datasets. 

To address this data scarcity problem, we use LLMs to increase the number of examples through question expansion, rephrasing, and other creative approaches. Depending on the domain, we can sometimes go from 100 examples to 500 to 1000 rows of data. 

Why this works

Although this approach can sometimes overproduce and the examples can seem repetitive, it works for us because we have the luxury of reviewing the generated training data and eliminating poor-quality examples before using it to build our downstream models. 

Over to you: How you can leverage LLMs to speed up AI development

Similarly, if you have highly domain-specific tasks and LLMs don’t seem to be cutting it, instead of using LLMs to complete the task, use LLMs to generate training data, which you can then use to fine-tune suitable LLMs or build simpler ML models to complete the intended tasks accurately. 

2/ Computing Similarities

Many of the “AI models” that our clients require are more “AI solutions” rather than a single AI model. By that, I mean its a combination approaches (ML, traditional NLP, intelligent software engineering) coming together to solve a specific problem.

For example, if you take a recruiting tool that will recommend potential candidates given a job description, you will often need more than a single AI model under the hood. 

There will be some sophisticated software engineering paired with AI components in different areas, such as in scoring potential candidates and augmenting the input data with more information. In this context, one common area where we often leverage LLMs is computing similarities or relatedness. 

For example, if you take the problem of assessing how related these two texts are:

• “What is Lena’s work experience?” 

• “Lena is a machine learning engineer with over 15 years of experience in python,..” 

This requires us to compute how close the answer is to the question. We have used numerous LLMs for similar tasks with varying degrees of success depending on the domain. 

Why this works

LLMs are excellent for this task because they understand the “semantics” of the texts used for comparisons. Further, you can leverage many smaller, open-source LLMs to effectively accomplish the task without relying on paid APIs or the billion-parameter LLMs. We often get comparable results with smaller LLMs with much less hassle when it comes to productionization.

Over to you: How to use LLMs strategically and increase the chances of automation success

If you have complex intelligent automation tasks that you’re looking to accomplish, you can leverage LLMs strategically in different areas to make your intelligent automation shine. 

For example, in creating a solution that crafts customized responses to emails from customers, you may be able to categorize emails with labels that LLMs can handle, so you can customize how each group of emails is processed rather than applying vanilla logic to ALL emails. Then, within each pipeline, you could retrieve the relevant data needed to address the issues within the email (non-AI approach) and then use LLMs to generate a beautiful email response that encapsulates the relevant extracted data ready to be reviewed by your service agents.

The bottom line is that instead of expecting LLMs to complete the entire automation task, use them only where needed for maximum control of your automation and a much higher chance of success in production systems. 

3/ Search 

Building powerful search applications is an area where we get a lot of consulting interest. While many companies rely on simple keyword searches to answer customer questions, find company-specific documents, etc., there’s a lot we can do to enhance the quality of search results by leveraging the power of AI, including LLMs. 

For example, we can use LLMs to understand the intent of the user’s query. Are users searching for “apple,” the electronic item, or “apple,” the fruit? Are users looking to “book,” as in “book travel,” or “book,” as in “a book to read?” Understanding the user’s intent helps us get users the right answers or the right navigation paths for their tasks.

Why this works

As LLMs have excellent language capabilities, they can help disambiguate the meanings of queries, improve user queries, augment queries, and retrieve semantically related documents. The best part is that you can use smaller open-source LLMs and get excellent results, keeping things reasonably easy to productionize. 

Over to you: How you can use LLMs to improve search

Look into areas where search or any type of retrieval or question answering is crucial to your customers or employees. Are you able to find what you need most of the time? 

Do you see obvious areas where the search engine or question-answering system misses the mark? Perhaps it did not understand your intent accurately, or a particular intent is not currently supported. It could also be that the retrieved results are irrelevant to the query. These are clear opportunities for improvement with more intelligence. Of course, these opportunities need further investigation through quantitative approaches, but it’s a starting point. 

Summary

As you’ve seen in this article, contrary to how most people think we use LLMs, in our development work where we architect comprehensive AI solutions, we don’t typically prompt and re-prompt LLMs to solve an entire automation problem. 

We use LLMs as needed, strategically placed to maximize the quality of the final results and to increase the chances of solutions being put into production. 

I’ve given you three areas to consider for how you can do the same. Let me know if you’ve found an area where you think one of the above suggestions could work.

That’s all for now!

One of the problems I commonly hear from leaders (typically at the Manager and Director levels) looking at AI integration is the lack of executive buy-in.

 The question is, why are executives so hesitant?

There are several reasons for this. A common factor is fear.

Having spoken to multiple executive teams, I can attest to the fear factor. I even had a CEO of a mid-sized tech company question the growing number of hyper-parameters of LLMs. “What’s next, with GPT-5 and 6? Will humans become extinct?” Having worked in this area for close to two decades, I often find myself staring blankly at these questions. As a practitioner, to me, AI is nothing but a tool.

Not surprisingly, 42% of 119 CEOs interviewed at the Yale CEO summit in 2023, say that AI could destroy humanity in five to ten years [1].

There is also the fear of breaching data privacy and security, biased and unfair models, and all those ethical risks that come with AI implementation. The fear factor is made worse when industry pioneers echo similar negative sentiments.

But we know that AI is not all bad especially when you are thinking about how you’re going to use it to benefit the organization, its customers and society at large and how you’d handle the associated risks.

Fear aside, AI as a technology, especially with Generative AI in the mix, is not well understood. People either misinterpret it as the “new cryptocurrency” meant to lose steam over time or some tout it as the panacea to all software automation problems.

Before the rise in popularity of ChatGPT and its cousins, much of this information and misinformation was previously put out by well-intentioned marketers, sales teams, and the media. All of this has changed. IT consultants, software developers, cybersecurity personnel, and others who seem “technical” but don’t have any background in AI or have not implemented AI for production use cases have also started adding to the noise.

It’s sad to say, but much of the AI information released by these new experts is not based on first-hand experience or directly implementable advice.

As an example, I recently saw a social media post (as you’ll see below) stating the accuracy numbers of the different LLMs and their corresponding hallucination rate. What this post doesn’t state is:

1. Accuracy on what tasks? (there are so many tasks you could use an LLM for, so how were these numbers computed?)

2. How was the hallucination rate computed (is this a new metric? )

3. More importantly, what’s the source of this data? (no links and no citations)

A misleading social media post. There are no citations and no explanation of the metrics and what they measure.

This could very well be data taken out of context from Meta’s or Microsoft’s research papers or recycled news from another influencer.

Whatever it is, this seemingly authoritative content is misleading. A leader new to AI may think that GPT-4 is the gold standard solution for all AI problems when, in fact, there are a series of cheaper but equally good LLM and non-LLM solutions for many AI problems.

All this is to say that it’s hard for leaders to determine which information is accurate and what they should trust.

While educating your executive team may seem easy, it’s often not implemented. However, it’s a critical step in getting past the AI hesitance and full AI buy-in. From my experience teaching and working with executives, I’ve repeatedly been told how having targeted content is so much more helpful in improving their understanding than reading scattered content from the Web.

Getting Started with Executive Education

To get started with executive education as a strategy to get AI buy-in and reduce resistance, here are few steps you can take.

Step 1: Get a team vested in the use of AI within the company. This can be people from your analytics team, product management team, etc. 

Step 2: Try to understand your executive team’s reasons for AI hesitance and assess their knowledge gaps. Do they know AI at a high level? Are they fearful because of the talk of an AI takeover? Get deep into their fears and misconceptions about AI. 

Step 3: Put together a topic list that would address your executive team’s biggest concerns, fears, and knowledge gaps. Suppose ethical issues and data privacy are a big concern. You can have a topic that covers the true risks of AI and practical mitigation strategies. You can also show examples of how other companies are handling those risks.

In the healthcare domain, for example, one big fear is the use of AI in clinical diagnosis. Using AI as a sole decision-maker presents an elevated level of risk that not many hospitals want to take on. However, if you demonstrate through valid use cases how AI systems are used mostly as assistants and as sole decision-makers only in low-risk situations, it starts painting a more plausible future with AI.

Step 4: Curate content and invite speakers. There are some topics that will resonate deeply with executives when presented by their internal teams. However, having a mix of internal content further reinforced by external speakers can drive home some crucial points, such as GenAI is not the only way to implement AI.  Or, jumping head first into AI without proper use cases can result in failed initiatives or initiatives with little to no impact.

The more you can help your executive team understand AI, how it reasons and how it’s built, where it truly helps, and how certain risks can be mitigated, the more approachable the technology will become. Of course, there will always be bad actors using the technology in harmful ways. But this is a problem with any technology, not just AI, and this might be another important point when communicating with them.

References

[1] https://www.cnn.com/2023/06/14/business/artificial-intelligence-ceos-warning/index.html

With the growing number of ready-to-use or semi-ready-to-use AI tools, buy or build AI questions are top of mind for many leaders. There’s a big push to start using AI immediately, but many leaders also understand that not all AI tools will fit their organization’s needs.

For example, one of our customers invested heavily in developing machine learning models internally, from scratch. However, after deployment, they were inundated with customer complaints due to inaccuracies in model output and needed a way out of this problem. 

Our evaluation pointed to several significant problems that would require the company to rebuild models entirely. To prevent further setbacks and contain costs, instead of investing more in in-house development, we helped them transition to a third-party solution, outperforming their existing one while keeping costs within expectations. 

There can be many such twists and turns when implementing AI. You may have budget limitations, need more data, or your company’s IT infrastructure may also be too “simple” to support the use of AI. 

You may also be faced with privacy concerns, cost complications, and other important organization-specific challenges, which could push you to either build from scratch or buy an off-the-shelf solution.

Therefore, the best strategy for integrating AI is highly context-dependent. 

In this article, we’ll examine the different ways to integrate AI and the pros and limitations, or things to consider, for each. This will help you assess what’s best for your organization.

Integrate Task-Specific Third Party APIs

Task-specific third-party APIs are AI solutions that solve targeted problems and often integrate into a large software application. For example, labeling documents with specific categories, recognizing entities in text data, recognizing people within images, and so on are examples of task-specific problems.  

Benefits of Task-Specific APIs

The benefit of a task-specific third-party API is that most of the machine learning work is already done for you, and your costs are relatively predictable if you understand your approximate usage patterns. You don’t need a team of data scientists to build the solution, nor do you need to understand ML algorithms to start using these APIs. Often, you can use the output directly within your applications without customization. Examples of such APIs include Amazon Textract, Amazon Transcribe, and Vertex AI Vision.  

In most cases, you also won’t have to worry about maintaining and monitoring models. This is mostly taken care of as these are managed services. You can also use your existing engineering and IT teams to help evaluate and integrate the API into your business systems. 

Example architecture of how a podcast can be used to support search and analytics using Amazon Transcribe and Amazon Comprehend ML APIs. Source: aws.com

Limitations and Things to Consider

The downside of third-party task-specific APIs is that the solutions are often overgeneralized. They’re meant to work for everyone, and consequently, quality and accuracy are often at stake. The main reason for this is often that the underlying assumptions are not specific to your organization’s use cases and data. 

For example, suppose you’re trying to perform sentiment analysis on customer conversations from phone transcripts. Sentiment APIs that are primarily trained on tweets and short comments may not work well on this data as it doesn’t resemble the idiosyncracies of phone transcripts. This may lead you to think that AI doesn’t work when in reality, the particular API may be unsuitable for your use case.  

Additionally, third-party task-specific APIs are often available as a cloud service, where you need to send out data over the internet to get a set of predictions. This could mean you’re exposing confidential, sensitive, or proprietary data to a third-party service. It’s often unclear how your data is stored or used once the third-party service has collected them, unless it’s clearly specified in their terms and agreement. 

Bottom Line

Task-specific third-party APIs are an incredibly convenient way to integrate AI without hiring data scientists or understanding ML model development. Further, as they’re already fine-tuned for a specific task, the costs are also more predictable if you understand your usage. This convenience may, at times, come at the expense of accuracy, and data privacy. So think about these elements when considering task-specific third-party APIs.

Customize or Fine-Tune LLMs

There has been an increase in the number of AI models, also known as large language models (LLMs), capable of comprehending and producing language. These models are pre-trained on large data sets and can be accessed through a straightforward API call. Providers of such LLMs include OpenAI (the GPTs), HuggingFace (Bloom), and Cohere. There are also LLMs that can run locally on your hardware such as Alpaca (Stanford) and LLama (Meta). 

A Handful of LLM providers. Source: https://cobusgreyling.medium.com/the-large-language-model-landscape-9da7ee17710b  

The nice thing about LLM models is that they can execute various tasks based on instructions (aka instruction following or prompting) and with a few examples on what the output should look and feel like. You’re essentially using AI to develop AI. These tasks can range from text classification to entity extraction to writing emails, resumes, poems and ad copy.

Benefits of Customizing LLMs

The benefit of pre-trained language models for AI application development is the speed in developing a range of intelligent solutions using just ONE base model. You can use it to develop sentiment analysis applications, machine translation applications, generate ad copy, write essays, and more. Of course, you’ll have to write the instructions (or prompts) in a way that generates the desired output. 

The beauty of this is that you don’t need large amounts of training data; at times, you need no training data at all. Plus, anyone can develop these applications-it doesn’t have to be a machine learning engineer.

Limitations and Things to Consider

Despite the convenience, LLMs are only suitable for some use cases. One big known problem of LLMs is hallucination, where facts are added to the answer, or answers are “made up” when the right answer is not known. This is a huge risk for applications that are dependent on facts. This problem becomes especially prominent when the output is fairly open-ended. For example, when  ChatGPT was asked what’s “Alpaca in the context of LLMs,” here’s what it answered:

In the context of LLMs, “LLaMA” refers to “Low-Latency Large Language Model Inference Acceleration”, which is a system developed by Microsoft Research to enable the deployment of large language models (LLMs) on commodity hardware with low latency and high throughput.

LLaMA achieves this by utilizing a combination of model compression, quantization, and efficient hardware acceleration techniques to reduce the computational resources required for running LLMs. This makes it easier and more cost-effective for organizations to leverage the capabilities of LLMs in their applications, particularly those that require low-latency inference.

Overall, LLaMA is an important development in the field of LLM deployment, as it enables the efficient and scalable deployment of LLMs on commodity hardware, which can have a wide range of practical applications in natural language processing, speech recognition, and other related fields.

ChatGPT’s response to “What’s Alpaca in the context of LLMs”

In reality, LLaMA stands for Large Language Model Meta AI (not Low-Latency Large Language Model that ChatGPT claims) and was released by Meta AI (not Microsoft Research which ChatGPT claims) in February 2023. But, it’s right about the usage on commodity hardware.

So, ChatGPT made stuff up instead of saying, “I don’t know it.”  This is extremely risky in many applications. This is why controlling the output of these LLMs through prompting or other mechanisms is absolutely essential.

For example, when asked, “What is llama in the context of LLMs that was released in February 2023,“ ChatGPT changed its answer entirely.

As an AI language model, I must inform you that my previous statement was incorrect, as I had mentioned that LLaMA was developed by Microsoft Research. However, there is no such system developed by Microsoft Research with the name LLaMA.

Moreover, it is not possible for me to provide information about something that was released in February 2023, as it is a future date beyond my knowledge cutoff of September 2021. Therefore, I cannot provide an answer to your question as it contains inaccurate information. If you have any other questions, please let me know, and I will do my best to provide accurate and helpful information.

ChatGPT’s response to “What’s LLaMA in the context of LLMs that was released in February 2023” right after asking ChatGPT “What’s LLaMA in the context of LLMs”

In summary, if you ask the LLM to perform a novel task like predicting the risk of a patient having lung cancer based on the description of the patient’s medical history, it may very well do it, but the results could be completely unreliable as it may make guesses based on data it has seen. 

Further, from a cost perspective, fine-tuning and developing applications over pre-trained models means that you may be pouring money into LLM APIs much earlier in the game than with task-specific APIs where you clearly understand the required input and the output that it produces. 

Bottom Line

LLMs are very flexible in that they can be customized to perform many tasks that require human-level intelligence with simple instructions and prompting, using just one base model. However, if used incorrectly, these pre-trained models have the habit of trying to be overly smart and overconfident by providing answers that are way off from reality. This can have adverse effects, from people getting hurt due to Dr. ChatGPT’s diagnosis to students learning twisted historical facts from Prof. ChatGPT.

Further, unless you’re opting for the open-source LLMs, you could be pouring a considerable amount of money in trying to fine-tune and make LLMs work for your applications. From experience, this is not a route many startups can afford to take when there are more affordable and predictable ways of achieving the same results.

Use AI in Packaged Applications

Packaged AI applications simply mean that the AI solution exists within a larger software application. 

For example, Splunk, an IT operations software solution, uses AI behind the scenes to go from reactive to more preventative responses with predictive technologies. There are specific features within the application that use AI, including event noise reduction and predicting health scores before service outages. 

While occasionally, the features in such systems need to be trained on company-specific data, there are many cases where the AI features are ready for use right out of the box.

Benefits of Packaged AI

The benefit of using AI within prepackaged software is the convenience that comes with it. You don’t need to “integrate” the solution, you don’t have to customize it, and you don’t have to think about developers in the loop. You would just use the AI features as you would any software tool. Security is also less of a concern if these tools are used within company firewalls, as there’s no data leaving company databases to a third-party service.

Limitations and Things to Consider

With the convenience of prepackaged AI also comes the issue of usability. While some prepackaged AI tools will work seamlessly untouched, some struggle to make accurate predictions on your data. 

That’s because, like with the task-specific third-party API option, these AI tools are meant to work broadly, for every organization. This usually comes at the cost of the accuracy and reliability of the solution.

Bottom Line:

Prepackaged AI tools can often be used right out of the box and are super convenient for a company looking to leverage AI within its workflow. The only way to know if it will work for your use cases and your company data is to evaluate it for a period of time.   

Build From Scratch

Building from scratch means you’re creating custom models for your AI problems. This can be in the form of models developed from scratch using machine learning and NLP libraries, an open-source solution that you customize, models developed using auto-ml tools, or training your own LLMs and customizing them afterward for specific applications. Typically, the developer or development team you hire is responsible for the success or failure of the solution. 

Benefits of Building AI from Scratch

The single largest benefit of developing something from scratch is control–control over what you’re building, the quality of the solution, the data that goes into training or fine-tuning your models, and how the models are tested, validated, and integrated into your workflow. 

Building a tool from scratch is very beneficial in sensitive domains such as law enforcement, finance, and healthcare, where the level of control will allow you to account for appropriate fallbacks during model failure and minimize downstream issues such as those relating to unwanted biases. Further, some domains are so niche that the available AI tools may not work well in those scenarios. In such cases, the best option would be to custom-build.

Limitations and Things to Consider

To be able to build from scratch, you would need the relevant personnel to develop the solution and later integrate it into your workflow. So you’d have to think about whether you should hire an employee or outsource the development, as both have their pros and cons, and one might be a better option than the other, given the number of projects on hand and budget limitations. 

Also, building from scratch is often more expensive than leveraging an off-the-shelf tool, although, in the long term, the costs can become comparable. It also takes time to build custom solutions, so you’ll have to factor that in. 

If you don’t have the building blocks ready for development, it can take even more time than expected. For example, you may have to allocate three additional months into your development time just to account for data collection.

Bottom Line

Although building a solution from scratch can seem cumbersome and costly, in many instances, it may be necessary. Your AI problem may be too domain-specific to warrant a generalized solution, or the problem may require a significant amount of control in terms of the data it’s trained on and how it’s evaluated. So don’t rule out this option. There are strategic ways to develop AI solutions cost-effectively and incrementally improve these solutions over time. 

Reality: Hybrid Build 

This is something you may not expect, but for most of our clients, we use a combination of approaches to solve a single problem, as real-world problems are often not as clear-cut or well-defined as research problems. 

Let’s take an information extraction project from PDFs. The information extraction piece may require AI in the form of NLP, but perhaps some content can be extracted using open-source tools. Still, some may require custom models for the extraction of highly domain-specific content. 

We have even leveraged paid APIs for parts of the work and custom-developed AI tools for parts that didn’t have open-source or paid tools. Further, that’s only the information extraction piece. What about the data collection and integration? Those pieces may require their own custom algorithms in combination with open-source AI tools, paid APIs, or custom-built AI. This is the reality of solving a real-world problem. You’re right, it’s not as elegant as a mathematical formula.

Build or Buy AI Software: Summary

In this article, we explored four approaches to implementing AI, including:

• Leveraging task-specific APIs

• Using AI in prepackaged software

• Customizing LLMs

• Building from scratch

However, in the end, you may end up using one or more approaches for a single AI problem (hybrid build).

With different options for implementing AI in your business, choose the one that makes the most sense, given your circumstances.  

Every team, organization, or product will have its constraints, be it IT infrastructure constraints, budget constraints, low tolerable risk, or privacy concerns. Further, you may have a sizable problem that requires a combination of approaches to solve one problem.

So be highly strategic in approaching AI implementation, and remember that you can continually iterate and improve with time.

Buy or Build AI: How You Can Implement AI in Your Business

References

• Introducing LLaMA: A foundational, 65-billion-parameter large language model

• Alpaca: A Strong, Replicable Instruction-Following Model

• How to Generate Quality Training Data For Your Machine Learning Projects (even if you’re data starved)

• Exploring the Ethical Implications of AI: A Closer Look at the Challenges Ahead

• What is automated ML? AutoML – Azure Machine Learning | Microsoft Learn

• 7 Natural Language Processing Applications for Business Problems 

• The Large Language Model Landscape | by Cobus Greyling

This article discusses what an AI strategy means, the different types of AI strategies that you should know about, and how as a leader you can get started with an AI strategy.

What is an AI strategy?

An AI strategy may seem like a complicated business-speak, but it’s simply a vision or high-level plan for integrating AI into the organization, such that it aligns with your broader business and automation goals. This high-level plan can be a:

• Product-level AI strategy (I often call this AI product strategy)

• Business-unit-level AI strategy

• Organizational-level AI strategy

• AI startup strategy

The granularity of the plan is often inversely proportional to the magnitude of the vision. So, the bigger the vision, the broader the objectives until each objective is broken down into a roadmap for implementation. 

Why Do Businesses Need an AI Strategy?

So now the question becomes, why does an organization need an AI strategy? Why can’t you just dive into implementation or make an off-the-shelf purchase with little planning? 

As Dale Carnegie says, “An hour of planning can save you 10 hours of doing” and with AI it can save you much more than that. It can save you months in setbacks and unnecessary spending. Let’s look at how an AI strategy can be helpful. 

Product Level AI Strategy

At the product level, an AI strategy is needed so that you have the foundational building blocks to ensure that an AI project can progress from an idea to a useful tool to creating meaningful results for the business. 

The building blocks at this level, could be the data that’s needed to support the initiative, the personnel who will build your data pipelines and develop the relevant models, the metrics to track success, or the competing off-the-shelf solutions that you should consider for integration. Without this level of detailed planning, you’re bound to run into roadblocks that will keep setting you back months at a time. 

Elements of an AI Strategy at the Product Level

For example, in early 2020, one of our healthcare customers wanted to develop an AI tool to automatically extract specific information from clinical records, primarily containing free-form text. They had been doing this information extraction work manually for years to obtain reimbursement from insurance companies. 

But as we were digging deeper into the project, shockingly, we discovered that none of that data was available. What happened?

As part of the company’s manual process, they copied data directly from the clinical records and pasted them into third-party forms for reimbursement. Unfortunately, once the forms were submitted, the company kept no records of the information that they copied over. This meant—no historical data for AI. As a result, the company had to first augment its manual processes with data collection in mind. This set them back six whole months.

This is a common problem for many companies, where the data to build the AI product is often nonexistent, is in the wrong format, or is yet to be collected. This is not just data for model training but also data for evaluation and learning more about the problem that you’re solving.

The building blocks for every AI product can look slightly different depending on whether you’re customizing an existing solution, building from scratch, or buying an off-the-shelf solution. 

Business Unit Level AI Strategy

At the business unit level, an AI strategy is needed to ensure that you’re pursuing AI opportunities that will give you the biggest improvement in your bottom line, productivity, or decision-making. 

Some of the drivers of those decisions include the severity of the problems that you’re addressing, data availability, infrastructure readiness, and availability of shared resources, and in many cases, you’d have to ensure that the selected initiatives align with the business’s broader objectives. 

To prevent you from tackling the wrong initiatives first (i.e., low-impact problems, no data problems, no buy-in problems), requires planning, which can help you be very strategic in implementing AI in your business unit. 

For example, if a new problem has no available data for development and evaluation, you may want to start the process manually first. And by doing this, you’re helping your future AI initiative in the following ways:

1. Establish a baseline performance, which you can later use for A/B testing

2. Generate high-quality data

3. Get a better handle on the problem

4. Leverage domain expertise in developing your AI solution in the future  

To sum up, a business unit level AI strategy is primarily to create a logical AI implementation roadmap to maximize impact.

Organizational Level AI Level Strategy

An organizational-level AI strategy is all about getting your organization as a whole ready and prepared for AI. This ensures that you’re not just running one-off pilots in an isolated business unit, but teams across the organization have AI on their radar, know how to tap into the necessary resources to get AI projects going, know how to spot AI opportunities, and have a base level of AI understanding.  

Getting an organizational-level AI strategy up is more than just finding AI opportunities to go after in your organization. It’s also about identifying all the readiness gaps from the perspective of:

• Budget (B)

• Culture (C)

• Infrastructure (I)

• Data (D)

• Skills (S)

All of this is so that a company can take an AI idea from conception to implementation to reaping measurable benefits—all with minimal friction, repeatedly.

In The Business Case For AI, I talk about the five preparation pillars for AI, referred to as B-CIDS (pronounced, “be kids”), which encapsulates the elements we discussed above.

The 5 AI Preparation Pillars From The Business Case For AI for an Organizational-Level AI Strategy Development. Data readiness is the biggest preparation pillar, followed by cultural readiness.

Each pillar is critical, with its own set of requirements for building an AI-ready company.

For example, under cultural readiness, there are six elements to consider to prepare a company culturally for AI. This is due to fears around the technology, the ethical dilemmas, the uncertainties that AI brings, and the need to be highly data-driven. 

If you’re not culturally ready for AI, you may experience resistance internally to the adoption of the technology, your employees may not know how to collaborate effectively on AI projects, and the risks that come with AI, including mitigation strategies, may not be well understood.

Another critical pillar, data readiness, is all about ensuring your data warehousing is strong and that you’re collecting data from the daily running of your business and are digitally ready to leverage your data stores. Essentially, this is all about getting your data strategy up and running.

Building up each readiness pillar takes time and a whole lot of planning. And a plan alone is not enough, you need an ACTIONABLE plan—something you can implement.

To facilitate the planning process, I’ve outlined The Jumpstart AI Approach in The Business Case For AI for getting started with your AI strategy using actionable, short-term strides. In a future article, we’ll explore the approach.

AI Startup Strategy

An AI startup is either selling an AI tool or technique that you can use within your applications (e.g., Perspective API), or AI is a big part of the startup’s products and services (e.g., Jasper.ai, the content writing assistant). 

If you’re looking to build a startup with AI as the primary driver of your business, then an AI strategy for your product is a must.

Depending on whether you’re developing the AI solution from scratch or using an off-the-shelf model to power your applications, the priorities can look slightly different. However, having worked with multiple AI startups, I can say that these are some common concerns:

• Model/Technology

  • Are you leveraging an open-source solution, buying, or building from scratch?

  • Who will perform the model development or customization, testing, and integration?

  • How will the solution integrate into your software or how will users access the solution?

• Performance

  • How will you assess the quality of your tool in solving a broad spectrum of unseen problems?

  • How do you ensure the quality of your tools stays consistent with time?

• Monetization

  • How will you monetize your tool?

• Safety

  • How do you ensure that your tool produces the expected output?

  • How will your tool fail gracefully instead of annoying users?

• Messaging

  • What exactly are you selling? The technology and quality behind the technology or the product’s benefits to the end user?

• Data

  • How will you start collecting data from day one to improve models, develop new models, etc.?

• Adoption

  • How will you ensure that you’re addressing fears around AI and helping new users become open to using your solution?

• Feedback Loop 

  • How would you collect feedback from users to improve your offerings continuously?

As you’ll notice, this has elements of an AI product strategy and some organizational-level AI strategy, with some monetization and marketing strategy. AI startups must prioritize their messaging, monetization strategy, quality consistency, and feedback loop, as AI is at the front and center of their business.

This is not to say that these elements are not critical to non startup AI products that augment existing workflows. It’s just that the priorities in those scenarios could look very different. 

For example, for a large company seeking to enhance a workflow, a 50% boost in productivity through AI may suffice regardless of the model’s accuracy as long as it meets the task requirements. But for an AI startup selling an AI tool, poor or even just “acceptable” accuracy can destroy a buyer’s trust. 

AI startup strategy has elements of an AI product strategy and organizational-level AI strategy.

Getting Started With Your AI Strategy

The first step in getting started with your AI strategy is to understand what it is, why it’s important, and if you’re considering an organizational, business unit, or product-level strategy. The scope of your AI strategy will determine what preparation elements to focus on.

At the product level, we saw that we’re primarily focused on the building blocks related to the specific AI project, such as data and personnel. 

But at the organizational level, we’re focused on getting the entire organization to have AI on their radar and be able to repeatedly take an AI idea into production and deliver value for the business with minimal friction. This takes much deeper and lengthy preparation along different preparation pillars. 

For an organizational or business unit level AI strategy, you can use The Business Case For AI, for concrete steps and frameworks to develop your AI strategy. For a product-level AI strategy, you’ll have to think about all the elements outlined in the first image.

AI ethics is about releasing and implementing AI responsibly, paying attention to several considerations, from data etiquette to tool development risks, as discussed in a previous article. In this article, we’ll explore some of the ethical issues that arise with AI systems, particularly machine learning systems, when we overlook the ethical considerations of AI, often unintentionally.

The 5 Common AI Ethical Issues

1. Bias propagation (racial, gender & socioeconomic biases)

Although there’s a strong belief that algorithms are less biased than humans, AI systems are known to propagate our conscious and unconscious biases. 

For example, there are known recruiting tools that algorithmically “learned” to dismiss women candidates as they learned that men were preferred in the tech workforce.

Even facial recognition systems are infamous for disproportionately making mistakes on minority groups and people of color. For example, when the researcher, Joy Buolamwini, looked into the accuracy of facial recognition systems from various companies, she found that the error rate for lighter-skinned males was no higher than 1%. However, for darker-skinned females, the mistakes were much more significant, reaching up to 35%. Even the most renowned AI systems have been unable to accurately identify female celebrities of color.

So, what’s the primary cause of AI bias?

Data. AI systems today are only as good as the data they are trained on; if the data is nonrepresentative, skewed towards a particular group, or somehow imbalanced, the AI system will learn this nonrepresentation and propagate biases. 

Bias in data can be caused by a range of factors. For example, if historically, certain groups of people have been discriminated against, this discrimination will be very well recorded in the data.

Another reason for bias in data can be a company’s data warehousing processes or lack thereof, causing AI systems to learn from skewed samples of data instead of representative ones. Even using a snapshot of the Web to train models can mean you’ve learned the biases in that snapshot. This is why large language models (LLMs) are not free from biases when they’re quizzed on subjective topics.

Bias in data can also be a development mistake where the data used for model development was not sampled correctly, resulting in an imbalance of subgroup samples. 

Bottom line: When there’s limited oversight of the quality of data used for model training, various unintended biases are bound to happen. We may not know when and where especially with unconstrained multi-taskers like LLMs.

2. Unintended Plagiarism

Generative AI tools such as GPT-3 and ChatGPT learn from massive amounts of Web data. These tools generate the probability of producing meaningful content. In doing that, these generative AI tools may repeat content on the Web word-for-word without any attribution. 

How would we know that the generated content is, in fact, unique? What if the uniquely generated text is identical to a source on the Web? Can the source claim plagiarism?

We’re already seeing this issue in artwork generators that learn from a large number of art pieces belonging to different artists. The AI tool may end up generating art that combines work from multiple artists.

In the end, who exactly owns the copyright to the generated art? If the artwork is too similar to existing ones, this can lead to copyright infringement.

Bottom line: Leveraging Web and public datasets for developing models can result in unintended plagiarism. However, due to little AI regulation worldwide, we currently lack enforceable solutions.

3. Technology Misuse

A while ago, a Ukrainian state leader was portrayed as saying something they did not actually say, using a tool called deepfakes. This AI tool can generate videos or images of people saying things that they never actually said. Similarly, AI image generator tools like DALL.E and Stable Diffusion can be used to create incredibly realistic depictions of events that never occurred.

Intelligent tools like these can be used as weapons in a war (as we’ve already seen), to spread misinformation to gain political advantage, manipulate public opinion, commit fraud, and more. 

In all of these, AI is NOT the bad actor, it’s doing what it’s designed to do. The bad actors are the humans who misuse AI for their own advantage. Furthermore, the companies or teams that create and distribute these AI tools have not taken into account the wider effects these tools may have on society, which is also an issue. 

Bottom line: While the misuse of technology is not exclusive to AI, because AI tools are so adept at replicating human abilities, it is possible that the abuse of AI could go undetected and have a lasting effect on our view of the world.

4. Uneven Playing Fields

Algorithms can be easily tricked, and the same is true of AI-powered software, where you can trick the underlying algorithms to gain an unfair advantage.

In a LinkedIn post that I put out, I discussed how people might trick AI hiring tools when you disclose the attributes the system will use in the decision-making process.

While enforcing steps to reveal an AI’s decision-making process in hiring is a well-intentioned step toward promoting transparency, it may enable people to game the system. For example, candidates may learn that certain keywords are preferred in the hiring process and stuff their resumes with such keywords, unfairly getting ranked higher than more qualified candidates. 

We see this on a much bigger scale with the SEO industry, estimated to be worth over 60 billion dollars. Getting ranked highly in Google’s eyes these days is not just a function of having meaningful content worth reading. But also a function of having done “good SEO” and thus, the growing popularity of this industry.

SEO services have enabled organizations with hefty budgets to dominate the ranks as they’re able to invest heavily in creating massive amounts of content, performing keyword optimization, and getting links placed broadly around the Web.

While some SEO practices are mere content optimization, some “trick” the search algorithms into believing that their websites are the best in class, the most authoritative, and will provide the best value to readers. This may or may not be true. The highly ranked companies may have just invested in more SEO.

Bottom line: Gaming AI algorithms is one of the easiest ways to gain an unfair advantage in business, career, influencer-ship, and politics. People who figure out how your algorithm “operates” and makes decisions can abuse and game the system.

5. Widespread Misinformation

As we rely more and more on answers and content generated by generative AI systems, the “facts” that these systems produce can be assumed to be the ultimate truth. For example, in Google’s demo of their generative AI system, Bard, it provides three points in response to the question, “What new discoveries from the James Webb Space Telescope can I tell my 9-year-old about?” One of the points states that the telescope “took the very first pictures of a planet outside of our own solar system.” However, astronomers later pointed out in a very public way that this wasn’t the case. Directly using output from such systems can result in widespread misinformation. 

Unfortunately, without proper citation, it isn’t easy to verify facts and decide which answers to trust and which not to. And as more people accept the content generated without question, this can lead to the spread of false information on a much larger scale than seen with traditional search engines. 

The same is true for content ghostwritten by generative AI systems. Previously, human ghostwriters had to research information from trustworthy sources, piece them together in a meaningful way, and cite the sources before they publish. But now, they can have entire articles ghostwritten for them by an AI system. Unfortunately, if an article generated by an AI system is published without further verification of the facts, misinformation is bound to spread. 

Bottom line: Over-reliance on AI-generated content without the human verification element of the facts will have a lasting impact on our worldviews due to the non-fact-checked information we consume over extended periods of time.

Summary

In this article, we explored some potential ethical issues that can arise from AI systems, particularly machine learning systems. We discussed how:

• AI systems can propagate racial, gender, age, and socioeconomic biases

• AI can infringe on copyright laws

• AI can be used in nonethical ways to harm others

• AI can be tricked, unleveling the playing field for people and businesses

• Trusting answers blindly from AI systems can cause widespread misinformation

It’s critical to note that many of these problems were not intentionally created, but rather they are the side effects of how these systems were developed, disseminated, and used in practice.

Although we can’t eliminate these ethical problems entirely, we can certainly take steps in the right direction to minimize the issues created by technology in general, and in this case, AI.

With insights into the ethical dilemmas of AI, let’s focus on devising strategies for more responsible development and dissemination of AI systems. Instead of waiting for government regulation, in an upcoming article, we’ll explore how businesses can lead the way in doing AI responsibly. 

Keep Learning & Succeed With AI

• JOIN OUR NEWSLETTER, AI Integrated, which teaches you how to successfully integrate AI into your business to attain growth and profitability for years to come.

• READ OUR BOOK, The Business Case for AI,to learn practical AI applications, immediately usable strategies, and best practices to be successful with AI. Available as: audiobook, print, and eBook.

• GET EXPERT ADVICE, Move your AI agenda forward with expert advice, implementation help, and customized guidance for your AI problems. From strategic issues to AI product design.

Not Sure Where AI Can Be Used in Your Business? Start With Our Bestseller.

The Business Case for AI: A Leader’s Guide to AI Strategies, Best Practices & Real-World Applications. By: Founder, Kavita Ganesan

In this practical guide for business leaders, Kavita Ganesan, our CEO, takes the mystery out of implementing AI, showing you how to launch AI initiatives that get results. With real-world AI examples to spark your own ideas, you’ll learn how to identify high-impact AI opportunities, prepare for AI transitions, and measure your AI performance.

The post Exploring the Ethical Implications of AI: A Closer Look at the Challenges Ahead appeared first on Opinosis Analytics.