The question arrives in every boardroom eventually, and it is nearly always the same question: “Where can we use AI?” A director asks it after reading something on a plane. Investors ask because their own investors are asking – and management teams, because not asking has started to look like a decision. (It has.)
I answer some version of this question for a living, and my argument here is that it is the wrong question – wrong in a way that quietly produces a large share of the AI spending that later gets written off. The question I start with instead: where could this business become meaningfully better? Ask that first, and only afterward decide whether AI belongs in the answer. Sometimes it does. Often the honest answer is something less fashionable – ordinary automation, cleaner data, a process fixed by hand – and a fair amount of the time it is “not yet” or “not at all.”
What follows is how I work through that decision. You can run it on your own business this week, without me in the room.
The objection worth taking seriously
The obvious pushback first: isn’t this just semantics? Both questions end in a list of initiatives. Does the order really matter?
There is a stronger version of the objection, and it deserves a fair hearing. You cannot recognize an opportunity in a technology you do not understand. Nobody derived the spreadsheet from first principles by asking where their business could become better – they saw VisiCalc, and then understood their business differently. On this view, starting from the technology is not naive; it is how technological opportunity has always been found.
I concede most of that. Fluency in what the technology can and cannot do is genuinely necessary – I have spent years teaching it to non-technical executives. But fluency is an input to the decision, not a substitute for it, and the two questions produce different lists. “Where can we use AI?” starts from the tool and inventories everything that resembles the demo: a pile of use cases, mostly small, none of them ranked by whether the business would notice. “Where could this business become meaningfully better?” starts from the P&L and the workflow, and produces a short list with economics attached – whose striking property is that several of its best items turn out not to be AI at all.
The order matters. Start with the model and you will find uses for the model. Start with the business and you will find the two or three places where something – possibly AI – is worth doing at all.
Six questions, asked in a deliberate order
In practice I work backward through six questions. The order is the method.
Value comes first. Is the opportunity economically large enough to matter – would it move EBITDA, capacity, customer economics, decision speed, or leakage in a way the people running the business would notice without being told where to look? If not, stop. Nothing downstream rescues an opportunity that was too small at the start.
Then process. What, concretely, would have to change – which workflow, which decision, which handoff, which expensive block of human effort? AI creates value by changing how work gets done. If nothing about the work changes, the value is theoretical, and theoretical value has a way of staying that way.
Then data. Does the information the system needs exist, can it be reached, and is it trusted? Do the people who would use the output agree on what the underlying numbers mean? In my experience, this question ends more initiatives than any other – and answering it squarely is often where the real transformation begins, a year before anything resembling AI shows up.
Only then, technology. Fourth, on purpose. Asked without a thumb on the scale, the menu at this step is wide: a language model, a predictive model, plain deterministic automation, conventional software, a cleaner data foundation – or no new technology at all, because the process problem found at step two can be fixed by deciding to fix it. AI is a tool category. It does not get to skip the queue because it is the reason everyone is suddenly interested.
Adoption next. Whose behavior has to change, in exactly what way, and what will stop that change from happening – inability, unwillingness, friction, or the plain fact that the old way still works and the new way is on probation? A technically successful system nobody uses is not a partial success. It is a failed investment with good documentation.
And measurement last – which is to say, first. How will we know it worked? That gets answered before anything is built, while the answer can still be cheap and honest. Decide it afterward and the measurement will be designed, consciously or not, to flatter the money already spent.
Each of these compresses a dozen smaller questions – practitioners will notice how much I am flattening – but the order is what does the work. And none of it is exotic: it is the same discipline an operator applies to any capital decision.
The verdicts
Run a candidate initiative through those six questions and you get a verdict. The range I consider honest:
- Yes.
- Yes – once a specific prerequisite is fixed, and the prerequisite is the real project.
- The problem is real, but ordinary automation solves it cheaper and more reliably.
- The technology works; the economics do not.
- The economics work; the organization cannot absorb the change this year.
- The cost of the errors would exceed the value of the speed.
- The data this would need does not exist yet.
- Not at all.
Only one of the eight is an unqualified yes. That ratio is not the method being stingy – it is what the questions produce when they are allowed to answer. A useful AI strategy is not a list of AI projects; it is a prioritization system, and the no’s and not-yets are most of its output. They are also what make the surviving yes credible: to a board, to an investment committee, and to the team being asked to do the work. I have made the market-level version of this argument elsewhere – the value gets decided in selection, before anything launches – but at the level of one business it is simpler. Every no is a quarter somebody keeps.
I would rather disqualify a weak initiative on paper than watch a company spend money and management attention discovering the same verdict the expensive way. The strategies I distrust on sight are the ones where every idea survived.
This problem is older than the technology
I did not arrive at this method during the current wave. In 2019 – three years before the wave – I co-wrote a research study on AI in financial services, covered by the Financial Times and cited by the Bank of England and the European Parliament (institutions not famous for excitability). Its central argument fit in its title: “it’s not magic.” AI was pattern-finding in data, consequential precisely because it was unmagical – and the institutions getting hurt by it were mostly the ones expecting magic. In the same stretch I built and delivered AI education for executives: non-technical leaders learning to judge whether an AI project was worth doing at all.
Rereading that material now is instructive. The technology it describes is antique. The failure modes are not. What killed machine-learning projects in 2019 was rarely the model – it was unowned adoption, unmeasured outcomes, data that could not carry the use case, and decision-makers who had never been given a way to tell a good project from an impressive one. The current wave replaced the technology and kept the failure modes. That is why technology sits fourth on my list of six: five of the questions were the hard part before the wave, and they are the hard part still.
What building taught me
The other place the method comes from is building. I founded an AI software company and built its engine myself – a system that had to turn a language model’s judgment into scores a customer could act on. Shipping it taught me what advising alone had not: you design around where the model is unreliable, not around what it can do on a good day. The model supplied the judgment under uncertainty; everything that could be deterministic – the scoring, and the before-and-after comparisons customers made decisions with – stayed in ordinary software, because a model asked to do arithmetic will sometimes do it wrong. A language model does not know your business exists, and it will not give you the same answer twice unless you engineer for that.
The model is rarely the hard part. The operating system around it – the workflow it sits in, the checks on its output, the measurement that says whether anything improved – is where the work lives, and it is what the six questions are built to surface before money moves.
The mark of a real strategy
Back to the boardroom question. “Where can we use AI?” will keep being asked, and it will keep producing lists of uses. The question that produces a strategy is the other one – where could this business become meaningfully better? – followed by the discipline to accept the verdicts, including the unfashionable ones. If I could leave one test behind for judging any AI strategy, mine included, it is this: look at what it declined. A strategy with no no’s in it is a wish list that has not met the business yet. The harder question is whether we would recognize the difference in our own.
If you want the six questions run with outside judgment, the bounded version is The Portfolio AI Baseline: the same working-backward read, done across a fund’s portfolio. If you are the partner who owns value creation, it asks for under three hours of your own time – a working session and a playback, with the data request handled by your analysts and the desk work off your calendar. It is free for a stated reason: the baseline is how I choose the work, and money changes hands only if the fund decides to continue past it – which is what lets the verdicts above stay honest. Where the work goes deeper at one of the companies, it belongs to that company: its leadership engages it, owns it, and gets the credit for what it builds.