Why AI Transformations Really Fail

This is the month boards approve next year's spending, and for many of them the line marked "AI" will have gone up again. The record says that money deserves harder questions than it usually gets.

Across thousands of organisations, the pattern is now consistent: investment rises, and returns do not follow. When a pattern holds at that scale, it stops being bad luck and starts being a governance finding. And the finding, once you look at the evidence, is uncomfortable for anyone who signs off a technology budget. AI transformations fail for the same reasons every transformation fails — and almost none of those reasons is the technology.

Orgvue's survey of more than 1,000 senior leaders, published in May, found that 78% of organisations have had AI projects either fail (35%) or remain stuck in pilot (43%) — and that 83% still plan to increase their AI investment this year. PwC's Global CEO Survey, drawn from 4,454 chief executives in 95 countries, found that only about one in eight has seen AI deliver both lower costs and higher revenue; 56% report no significant financial benefit to date.

None of this means AI does not work. It means that, for most organisations, the way AI is being introduced does not work. That is a different problem, and it belongs to a different part of the organisation: not the IT department, but the people who decide what gets funded, who owns it, and what counts as success.

Spending rises. Returns do not. At that scale, it stops being bad luck and starts being a governance finding.

The most useful number in this whole debate is not a failure rate. It is a ratio.

When Boston Consulting Group asked 1,000 executives across 59 countries what was actually holding their AI efforts back, roughly 70% of the challenges were about people and processes, 20% about technology, and only 10% about the algorithms themselves. BCG found that 74% of companies had yet to show tangible value from AI — and the small group that had pulled ahead were not the ones with better models. They were the ones that had put most of their effort into the 70%.

McKinsey's global survey points the same way. Its "AI high performers" — the roughly 6% of organisations attributing 5% or more of their earnings to AI — are nearly three times as likely as everyone else to have fundamentally redesigned their workflows, rather than layering AI on top of the old ones. And MIT's NANDA initiative, whose 2025 study reported that around 95% of organisations saw no measurable profit-and-loss impact from generative AI pilots, located the problem not in model quality but in a "learning gap": tools that never fitted the workflow, and organisations that never learned how to use them.

Three different firms, three different methods, one conclusion. The technology mostly works. The organisation around it mostly hasn't changed.

A word of care first. You will often hear that "70% of all change programmes fail". That figure is folklore: when Mark Hughes traced its five most-cited sources, he found no valid empirical evidence behind any of them. I raise it because a board that repeats a statistic it cannot source has already made the first mistake of a failing transformation — mistaking a story for evidence. The honest data above are troubling enough without it.

What the evidence does show is a repeating pattern, and it takes three forms.

Adoption by imitation

In Orgvue's survey, 57% of leaders said the main reason they deployed AI was that their competitors had. That is not a strategy; it is a reflex. A project that starts because someone else did it rarely has a defined outcome, an owner, or a reason to stop.

Pilot purgatory

Pilots are cheap to start and expensive to end. Without kill criteria or a route to scale, they accumulate — each one "promising", none of them funded to change how the work is actually done. The 43% stuck in pilot are not failures in the technical sense. They are decisions nobody has taken.

Automating the old process

When a task is handed to a machine without rethinking the work around it, the cost falls and the quality often falls with it. Klarna's chief executive said publicly in 2025 that cost had been too predominant a factor in its AI customer-service push and that service quality had suffered; the company began adding human agents again, while stressing that it was not reversing its use of AI. That nuance is the point. The issue was not the tool. It was the design of the work around it.

In my last essay I argued that a court can automate its administration but never its judgement. The same line runs through every organisation, and most transformations draw it in the wrong place — or not at all.

Too often "human in the loop" means a person somewhere downstream who can, in theory, catch a mistake. That is a safety net, and safety nets are where accountability goes to disappear. Air Canada learned this in 2024, when it argued before a Canadian tribunal that its website chatbot was effectively responsible for its own misleading advice. The tribunal rejected the argument: the chatbot was part of Air Canada's website, and Air Canada was responsible for what it said.

The organisations that get value from AI treat the human role as part of the design. Before anything is deployed, they decide which decisions the machine prepares and which a named person makes; what that person needs to see to exercise real judgement; and who answers when it goes wrong. That is not a brake on transformation. It is what makes a transformation trustworthy enough to scale.

The human in the loop is not a safety net. Safety nets are where accountability goes to disappear.

None of this needs a new committee. It needs the board to ask different questions of the AI line it is about to approve.

The technology will keep improving whether or not boards change how they govern it. That is precisely the danger: better tools make it easier to mistake activity for progress. The transformations that succeed will not be the ones that bought the best models. They will be the ones whose leaders had the patience to change the work, and the discipline to keep a human — a named, accountable, well-informed human — in the loop.

As 2027 approaches, the question for every board is not whether to invest in AI. It is whether it is prepared to invest in everything around it.

The question is not whether to invest in AI. It is whether you are prepared to invest in everything around it.