Two numbers, from the same body of research, describe the entire state of enterprise AI in 2026.

88% of organisations now use AI in at least one function. And 74% see no tangible value from what they have spent on it.

The gap between those figures is not a technology story. It is one of the clearest governance failures I have seen at this scale, and it is worth being precise about what the evidence actually shows before reaching for a prescription.

What the research says

The figures below are synthesised from Deloitte, PwC, McKinsey, Stanford HAI, Gartner, Forrester and more than a dozen further sources.

  • 97% of organisations struggle to demonstrate clear business value from what they have already built.
  • 74% of companies report no tangible value from their AI investment (BCG), against total spend running into the hundreds of billions.
  • 42% of companies abandoned most of their AI initiatives in 2025, up from 17% the year before.
  • Just 7% of organisations have fully scaled an initiative past pilot.
  • Fewer than half of UK enterprises offer AI training of any kind, whether company-wide or role-specific.

And the single most useful number in the set:

  • A formal AI strategy takes the odds of measurable impact from 20% to 60%, and most mid-market organisations do not have one.

That last figure deserves more attention than it gets. It is not a claim that strategy is nice to have. It is a tripling of the odds, attributable to a document and the accountability structure behind it, before a single additional pound is spent on tooling.

The 74% is the number that should worry a board

Three quarters of companies spending on AI cannot point to anything they got for it. That is the figure I would put in front of a board, and it is the one that gets quoted least, because it is the one that implicates the people in the room rather than the technology.

If outcomes were mostly a matter of access to models, they would not look like this. Everyone is buying substantially the same models at substantially the same price, and a quarter of them are getting something out of it while three quarters are not. The technology is the constant across both groups, so it cannot be what separates them.

What separates them is everything around it: whether the data is in a state to be used, whether someone senior owns the outcome, whether the use case was chosen against a business problem or against a demo, and whether anyone is measuring the result against the case that justified the spend.

That list is unglamorous. It is also, precisely, what the quarter has and the three quarters do not.

Why the value can’t be shown

When an organisation cannot demonstrate value from AI, it is almost never because the value is absent and being concealed. In my experience it is one of four things, and they are worth separating because the fix differs.

1. Nobody wrote down what success would look like. The pilot was approved on the strength of the demo. There was no baseline, so there is now nothing to measure against. The benefit may well be real, and it is unprovable, which commercially amounts to the same thing when the budget comes up for renewal.

2. The benefit is real but lands somewhere nobody is counting. Time saved across forty people at ninety minutes a week is a genuine number, but unless somebody converts it into either reduced cost or redeployed capacity, it never appears in a management account. It shows up as people being slightly less busy, which no finance function has a line for.

3. The use case was never material. A great many pilots automate something that was not costing very much. They work exactly as designed, and there is no P&L impact to find, because there was never enough at stake for one.

4. There is no owner. The pilot belongs to an enthusiastic team lead who has neither the authority to scale it nor the mandate to kill it. So it persists, in production, unmeasured, indefinitely.

Only the third of those is a strategy failure in the usual sense. The other three are accountability failures, and they are considerably cheaper to fix.

The governance point

The training figure and the strategy figure are the same finding from two directions. Both say that outcomes are determined by the structures around AI rather than by the models themselves.

This is where I will be blunt about something the market is confused on. Governance, in this context, is not primarily about compliance. UK GDPR obligations and the EU AI Act matter and have to be met, but a governance structure whose only function is to keep you legal will not move a single one of the numbers above.

The governance that moves those numbers is much more ordinary: a named senior owner per use case, a written business case with a baseline, a standing decision forum with authority to stop things, and a defined threshold below which something is not worth automating at all. The power to say no is the part most organisations skip, and it is the part that generates most of the value, because it is what stops the pilot backlog from growing faster than the capacity to prove anything.

What this means if you are in the 97%

You almost certainly are, and it is not a mark against you. The distribution says so.

The first move is not more AI. It is an inventory: everything currently running or piloted, with a named owner, the problem it addresses, what it cost, and what evidence exists that it worked. Most organisations doing this for the first time find items nobody can account for, several duplicates, and at least one thing that should have been switched off a year ago.

That inventory is uncomfortable and it is also the cheapest week of work available, because it converts an unmeasurable portfolio into a list you can make decisions about. Everything else needs it to exist first: strategy, governance, scaling.

None of this is a technology problem. It is a leadership and governance problem, at exactly the scale where boards have not yet decided who owns it.

From the AI in Practice series. Next: why the constraint is plumbing rather than models.