AI is everywhere except in the results

The messy AI rollouts of the last eighteen months keep getting blamed on the technology. The real fault sits in how they were managed, and we wrote those lessons down decades ago.

Here is a number that should stop a boardroom cold. At Uber, around 95% of engineers use AI tools every month, and roughly 70% of committed code is now AI-generated. By most adoption metrics, that is a runaway success. And yet in May, the company's own COO admitted he can't draw a clear line between all that activity and the things that actually matter: whether useful features are shipping and users are any better off. His phrase was that the trade is becoming "harder to justify."

Uber just happens to be saying it out loud. The same firm burned through its entire 2026 AI tooling budget in four months. Microsoft questioned the cost of its coding licences before pulling them from a division. Across the industry, companies that went all-in are now staring at the gap between how much AI they're using and how little they can prove it's worth.

If you've spent any time in change management, none of this is surprising. It's familiar to the point of being uncomfortable.

We have seen this film before

In 1987 the economist Robert Solow delivered the line that defined a decade of frustration: you can see the computer age everywhere but in the productivity statistics. Computers were spreading through American offices, yet US labour productivity growth had slipped from close to 3% a year before the early 1970s to around 1% after it. Organisations were spending enormous sums on a technology that was obviously transformative, and the returns simply weren't showing up.

The mistake everyone made at the time was to read this as a verdict on the technology. It wasn't. When researchers went back through the data years later, the pattern was clear. The companies that bolted computers onto their existing ways of working got the paradox. The companies that redesigned how they worked around the technology got the returns. The gains were real, but they were gated behind organisational change, and organisational change takes deliberate effort and time. The constraint was organisational all along, and the firms willing to do that work were the ones who eventually saw the numbers move.

Swap "computers" for "AI" and you have a fairly complete description of 2026.

What companies actually did instead

Faced with a genuinely powerful new tool, most organisations reached for the oldest play in the book: push the tool out and mandate its use. Some went further and gamified it. Internal token-consumption leaderboards became a thing. Jensen Huang floated the idea that a software engineer on $500,000 ought to be getting through at least $250,000 of tokens a year, as though consumption were the point. Token maxxing, if you want a name for it.

This is procurement logic dressed up as transformation. Buy the licences, push the tool out, watch the usage dashboard climb, and assume value will follow because the activity is high. It is precisely the "overlay it on the existing process" error that produced the Solow paradox, except faster and more expensive, because tokens cost real money and that bill lands at the company level even though it feels free to the individual engineer experimenting at their desk.

The frameworks we already use in change work would have flagged every part of this before the first licence was signed.

ADKAR predicted the faceplant

Prosci's ADKAR model is almost boringly sequential: Awareness, Desire, Knowledge, Ability, Reinforcement. You cannot skip steps. People need to understand why the change is happening and want it before any amount of capability or reinforcement does anything useful.

The standard AI rollout started at step three. Here is the tool, here is some training, your usage is on a leaderboard. No real Awareness beyond "AI is important and we're behind." No Desire, because nobody answered the only question employees actually care about, which is what this means for them. And what passed for Reinforcement wasn't reinforcing a behaviour at all. It was rewarding a proxy. Leaderboards reward token spend, so token spend is what organisations got. The outcomes were never in the equation.

COM-B explains why the code got worse

The behavioural science is just as tidy. The COM-B model says behaviour only happens when Capability, Opportunity and Motivation are all present at once. Rolling out a tool and mandating its use buys you Opportunity and a crude, externally-imposed Motivation. It does nothing for Capability, the actual skill of using AI well, or for reflective Motivation, whether people genuinely believe it helps them.

You can watch the consequences in the research. In a controlled trial run by METR on 2025-era tools, experienced developers working on familiar codebases were 19% slower with AI. The detail that matters most for change practitioners is this: they believed they had been 20% faster. The capability to use the tools well wasn't there, and the judgment to notice the difference wasn't there either. METR has since cautioned that the picture is moving. With newer tools and more practice, it thinks the slowdown may be closing, which is rather the point: capability is something organisations build over time, not a switch they flip. GitClear's data tells a complementary story, with code-quality measures sliding even as volume rose. Activity went up. The work got worse. Almost nobody could feel it happening.

The marketing was an own goal

And then there's the narrative the AI companies themselves chose. For two years the message from the top was apocalyptic. AI would eliminate half of all white-collar jobs. Entire categories of work would be gone. It made for thrilling press and, presumably, useful fundraising.

It also happens to be the single worst change-communication strategy imaginable. You cannot tell a workforce that a technology is coming to replace them and then act surprised when they don't embrace it with open arms. The anti-AI sentiment everyone now frets about is a textbook threat response to the exact story the vendors chose to tell. Entirely rational, and entirely predictable. The fear translated into real harm, too: surveys suggest most AI-attributed job cuts have been made in anticipation of impact rather than because the technology actually delivered it.

The tell is that the same executives are now quietly reversing. The narrative has softened to "delighted to be wrong" and "automation will expand the work people do," conveniently timed with a run of blockbuster IPOs. The doomsday framing is being walked back. The resistance it manufactured will take a great deal longer to undo.

The thread that ties it together

Look across all of it and the same failure repeats. Developers felt faster and were slower. AI feels free to the individual and bankrupts the budget at the company. The doomsday story manufactured a sense of inevitability that produced real resignation and real resistance. Every one of these is a perception-reality gap.

That is the whole game. Change management, stripped of its jargon, is the unglamorous discipline of closing that gap. It means doing the awareness work properly and building real capability you can evidence. It also means measuring actual behaviour and outcomes, even when raw activity is the far easier thing to count. It is slow, human work, and it is exactly the work that got skipped in the rush to look fast.

The irony is that skipping it didn't buy speed. It bought eighteen months of heavy spend with no clear line to value, a dip in the quality of the work, and a workforce that now needs convincing twice.

An AI rollout is a change programme wearing an IT project's clothes. The organisations that win the next phase will be the ones that treat it that way, willing to do the human work the technology was never going to do for them. Speed of deployment is turning out to be a poor predictor of who gets there.

At Lima Delta we spend most of our time on exactly this gap, helping enterprises roll out change in a way the people on the receiving end will actually take up. If your AI programme has the adoption numbers but not the results, that's usually where the answer is.