Earlier this year I wrote about what boards need to understand about AI — the data reality, the accuracy problem, the generative shortcut. The response was interesting. I got a lot of messages along the lines of “I agree with all of this, but I still don’t feel like I know enough to challenge my CTO when they put an AI proposal in front of me.” Which is exactly the problem.
The gap isn’t governance instinct. Board members know how to evaluate a budget, stress-test a strategy, and sniff out an over-optimistic forecast. The gap is technical literacy — not the ability to build AI systems, but the vocabulary to evaluate claims about them. Without that vocabulary, you’re reduced to either rubber-stamping whatever the technology team puts in front of you, or saying no to everything out of caution. Neither serves the business.
In September, I’m running a three-part course with The Boardroom Network to close that gap. Here’s what we’ll cover, and more importantly, why.
Class 1 — What AI actually is (and mostly isn’t)
The first session is about cutting through the terminology fog. “AI-powered” has become the most overused and least informative label in enterprise software — it could mean anything from a genuine neural network to a database query with a dashboard on top. The label tells you nothing; the mechanism tells you everything.
We start with the matryoshka dolls: AI contains machine learning, which contains deep learning, which contains the generative models everyone’s currently excited about. Each layer is a subset of the one above, not a synonym for it. Getting this taxonomy straight eliminates about half the confusion in a typical board discussion about AI.
Then we get to the bit that matters most for governance: the power versus explainability trade-off. This is the central tension that defines the entire field. In the 1980s, expert systems could explain every decision they made — “I recommended this because rule 47 matched condition B” — but they were brittle, expensive to maintain, and couldn’t handle anything their designers hadn’t explicitly anticipated. Neural networks swung to the other extreme: extraordinarily powerful pattern recognition, but the internal reasoning is opaque even to the people who built them. Every AI governance decision a board makes is, at some level, a negotiation between these two poles.
We’ll also spend time on where AI genuinely works — pattern recognition, classification, optimisation, prediction with stable historical data — versus where it reliably fails. Novel situations, causal reasoning, small datasets, rapidly shifting domains. The evaluation framework I use gives board members a structured way to assess any AI proposal: is the data sufficient, is the problem suitable, is the vendor claim testable?
Class 2 — How it works, and how it breaks
This is the session I’m most looking forward to, because it’s where we get concrete. The core claim is provocative but defensible: all machine learning is curve fitting. The sophistication is in the number of dimensions, not in the concept. Claude Opus is, at a fundamental level, y = mx + c with billions of parameters. Saying this to a room full of board members tends to produce a visible shift in confidence — the technology becomes less mystical and more evaluable.
We’ll work through a loan default prediction example that makes the business implications tangible. The model outputs a probability — say, 73% chance of default. The question is: what do you do with that number? At what threshold do you decline the loan? 50%? 70%? 90%? This is where most AI conversations go wrong — the threshold is a business decision, rooted in risk appetite and strategy, not a technical one. The board should be setting it, not the data science team.
Then comes the live demo. We’ll train a handwritten digit recognition model in real time — the kind of thing that hits 99% accuracy and looks like magic. And then I’ll show the class the 1% it gets wrong: inputs where the model is 100% confident in a prediction that is completely incorrect. No hesitation, no uncertainty flag, just wrong. Add a tiny amount of noise to the input and accuracy collapses while confidence stays sky-high.
This is, I think, the single most important thing a board member can understand about AI: models don’t know what they don’t know. There is no built-in “I’m not sure about this one.” When your vendor tells you the model is 99% accurate, the follow-up question should always be “what happens in the 1%, and how would we know?”
We’ll close with case studies that make the risks visceral. Wells Fargo’s credit scoring system approved refinancing for 47% of Black homeowners versus 71% across all lenders — over 100,000 minority applicants affected. Amazon built a hiring tool trained on ten years of past hires that learned to penalise CVs containing the word “women’s” and all-women’s college graduates; removing the gender terms didn’t help because the model found proxies. These aren’t hypothetical risks. They’re the predictable consequence of training pattern-matching systems on historically biased data and then trusting the output.
Class 3 — Generative AI, honestly
The final session tackles the thing everyone actually wants to talk about. Generative AI has genuinely changed the economics of AI deployment — from £500K–£5M for a bespoke model to £1K–£50K to get started with foundation models. That’s transformative, and I don’t want to downplay it.
But the shortcut comes with four specific risks that boards need to internalise:
Repeatability. Ask the same question twice, get different answers. Fine for brainstorming. A serious problem if you’re building regulated business processes on top of it. Audit trails become complicated when the system is non-deterministic.
Hallucination. The model is optimised to produce text that sounds correct, not text that is correct. In 2023, a New York lawyer named Steven Schwartz used ChatGPT to prepare a legal brief and cited six cases. All six were fabricated. The model had invented them, complete with plausible case numbers and judicial reasoning. He was fined $5,000 by Judge Kevin Castel. This is not a bug that will be patched out — it’s intrinsic to how these models work.
Explainability. Good luck explaining to a regulator exactly why your AI system made a particular recommendation when you can’t look inside the model. The EU AI Act has views on this, and they’re not optional.
The predictive-generative conflation. These are different tools for different problems, but they both get called “AI.” If you need to predict which customers will churn next month, you still need a predictive model trained on your data. ChatGPT won’t help.
The case study I find most illuminating is the Chegg and Duolingo comparison. Chegg ignored generative AI entirely — their stock is down 99%. Duolingo went all-in, replacing contractors with AI — their stock dropped roughly 80%. Both extremes destroyed shareholder value. The lesson isn’t “adopt or die.” It’s that how you adopt matters as much as whether you do.
Why this matters now
I’ll be honest about my motivation for putting this course together. I’ve spent the last year watching well-intentioned boards approve AI projects they couldn’t meaningfully evaluate, and watching equally well-intentioned boards block projects that would have created genuine value — both for the same reason. The technical literacy gap turns every AI decision into a coin flip dressed up as governance.
The uncomfortable truth is that a lot of the AI industry — and I include some of my fellow practitioners in this — has a vested interest in keeping the mystique alive. Complexity sells. “You wouldn’t understand the technical details” is a convenient way to avoid hard questions about whether the thing actually works. The best defence against this is a board that knows enough to call it out.
The course comes with practical tools: an evaluation framework matrix for assessing AI proposals, a six-dimension risk assessment, and a twenty-question checklist that any board member can use in their next meeting. These aren’t theoretical — they’re distilled from real conversations with real boards about real AI projects, some of which went well and some of which went expensively sideways.
The winners in the AI era won’t be the fastest adopters. They’ll be the smartest ones. The course runs in September with The Boardroom Network — if you’re interested, get in touch.
This post also appears on my Substack.