The 39% Problem

AI return is usually decided before the tool is bought

There is a pattern I keep seeing with AI: A business starts with the product. Which model. Which vendor. Which licence. Which use case looks impressive enough to get through the next leadership meeting. That feels practical. It is usually the first mistake.

The better question comes earlier: where, exactly, does the business need to change? Not where could AI be used. Not where could a pilot be demonstrated. Where does margin leak? Where does risk sit? Where does service strain? Where does the management team already know the organisation is slower, weaker or less certain than it should be?

If that question is not answered first, the buying decision has no anchor.

The research points in the same direction. MIT’s work on enterprise generative AI found that around 95% of pilots produced no measurable profit and loss impact. McKinsey’s 2026 read is less severe, but still uncomfortable: only about 39% of organisations can show AI reaching earnings.

That is the 39% problem: AI is being adopted almost everywhere, but return is being proved in far fewer places. The gap is not mainly a technology gap. It is an order-of-operations gap. Take a mid-market services firm. It buys an AI writing assistant because competitors seem to be doing the same. Six months later, the marketing team is producing more copy. The tool works. The team is faster.

But nothing material has changed in the numbers: the real constraint was elsewhere: finance still spends days reconciling data by hand, delivery teams still lose time moving between disconnected systems, and the management team still lacks a clean view of which work makes money. The writing assistant did what it was meant to do. It just was not aimed at the part of the business where value was trapped.

That is why I am wary of AI plans that begin with a shopping list. Every tool bought without a business reason still has to be governed. It needs data. It needs ownership. It needs policy. It needs budget. A pilot that proves nothing does not disappear quietly. It leaves behind a small administrative tax. Enough of those taxes become drag.

The firms getting value tend to do something less glamorous first: They decide the sequence. They start with the business agenda. They name the constraint. They work out what would have to change for that constraint to cost less, close faster or become easier to govern. Then they ask whether AI is the right instrument. Sometimes it is. Sometimes the answer is better data. Sometimes it is a process change. Sometimes it is simply stopping work that nobody should have been doing in the first place.

That is why sequence matters more than the individual tool decision. Point AI first at finance reconciliation and you may clean the data foundation that later decisions depend on. Point it first at marketing copy and you may improve a visible symptom while the underlying constraint remains untouched. Same budget. Same enthusiasm. Very different result.

This matters most for owner-managed founders and investor-backed leadership teams, because they do not have the luxury of confusing activity with value. A founder does not care which model tops a benchmark. They care where the next point of margin comes from. An investor does not care how many AI pilots exist. They care which ones change earnings, risk or enterprise value.

AI should be treated as a business resource, not a purchase: That sounds simple, but it changes the work. The first document is not a vendor comparison. It is a map of the business problem. The first meeting is not a demo. It is a sequencing conversation. The first decision is not what to buy. It is what to fix first.

The market is splitting on this point: One group keeps buying tools, running pilots and reporting activity. The other starts with the business, sequences the work and can explain where the return came from. The distance between those two groups will widen, because the second group compounds learning while the first compounds clutter.

So the question I would start with is not: Which AI should we buy?

It is this: If you removed every AI tool bought in the last two years and started again from the business, what would you do first, and why?

If that answer is clear, the technology decision becomes much easier.

If it is not clear, another tool will not rescue the plan.

That is the work: not smarter tools, but a clearer order.