Why AI adoption fails

Most companies I talk to have started doing something with AI. Usually there's a task force, everyone has a license for some chat tool, and there's a strategy deck. But if you look at how the work actually gets done a year later, not much has changed. The same people are in the same meetings making the same documents. The documents just take a bit less time to write.

The standard explanation is that the technology isn't ready yet. The models make mistakes, the data is a mess, and connecting everything is hard. That's all true, but I don't think it's the main reason. The tools are already good enough to change a lot of office work. What's usually missing is someone inside the company who actually wants the change to happen. [1]

Start with the people doing the work. Suppose you work in finance and you figure out how to make the monthly report in two hours instead of two days. The best thing that can happen is that nobody notices and you get some time back. More likely your manager fills that time with other work. And in the worst case someone realizes your job now takes two hours and decides they need fewer people like you.

So the sensible thing is to keep quiet about it. You use the tool and you don't mention it. Ethan Mollick calls people like this secret cyborgs, and I suspect most companies have more of them than they think. [2] They're often the people who have gotten furthest with AI, and the company learns nothing from them, because telling would be against their own interest.

I don't think you can blame them. We're asking employees to help build the thing that might replace them, and in most companies there's nothing in it for them if they do. The fear isn't paranoid either. They read the same headlines as everyone else, about CEOs who want teams to prove AI can't do a job before they're allowed to hire, and about layoffs announced right next to an AI strategy.

This is where a lot of change management goes wrong. It treats resistance as a communication problem, so the answer is more explanation: another town hall, another workshop, another newsletter. But people usually understand the plan perfectly well. They just don't like the deal they're being offered, and explaining a bad deal more clearly doesn't make it a better one.

You'd expect things to be different one level up, since management is who benefits. Sometimes they are. In my experience it depends mostly on whether the company is still growing.

When a company is growing, the question everyone is working on is how to get more. Anything that produces more output is welcome, because there's always more demand to meet, and the people who find better ways to work tend to get rewarded for it.

When growth stops, the question slowly changes to how to keep what you have. The total is roughly fixed, so one person's gain is someone else's loss, and a lot of energy goes into protecting budgets, headcount and scope. That's what people mean when they say a company has become political. The people usually haven't changed much. What they're rewarded for has.

In a company like that, AI is risky for a manager. If your standing depends on the size of your team, a tool that makes the team smaller doesn't help you. If you start an AI project and it fails, everyone knows it was yours. If you don't start one, nothing happens to you at all. People get judged on the targets they miss, and almost never on the improvements they didn't try, so doing nothing is the safe choice.

Kodak is the famous example of this. One of their own engineers built a digital camera in 1975, and the company didn't do much with it for a long time. It's easy to say they were blind, but I doubt that's what happened. Film was where the money came from and where everyone's career had been built, and nobody inside had much reason to push something that would shrink it.

The way AI is being rolled out now reminds me of what happened when factories got electricity. The economist Paul David pointed out that productivity barely improved for decades after electric motors arrived. [3] Most factories simply replaced their central steam engine with one big electric motor and kept the building and the work the same. The gains came later, when factories were redesigned around a small motor at each machine, which changed how the whole floor could be laid out. That took a new generation of managers.

Most companies are at the one-big-motor stage with AI. Everyone got a chatbot, and the processes, roles and approvals stayed the same. So it shouldn't be surprising that the numbers haven't moved much.

You can make the opposite mistake too. Klarna talked loudly about replacing a large part of its customer service with AI, and later said it would hire more people again because quality had suffered. [4] That story sometimes gets used as evidence that AI doesn't work. I'd read it differently. They swapped people for a tool without really changing the work around it, which is the one-big-motor mistake again, just pointed in the other direction.

If the problem is incentives, then the fix has to be incentives too, and buying more licenses won't do it. The most direct thing is to let people share in what they save. If someone automates half their job, they should get something out of it, like more interesting work, a bonus or a promotion, so that finding a shortcut makes you look good instead of making you a target.

It also helps to be honest about jobs. Telling people nobody will lose their job because of AI is rarely believable, and they can tell. It works better to say what will actually happen and how you'll help people move into other work, and then to do that.

On the management side, AI has to be part of what leaders are measured on, the same way revenue and cost are. Otherwise it stays a side project that nobody gets blamed for neglecting. And it's usually easier to start in the parts of a company that are still growing, like a new product or a new market, than in a department that is mostly trying to survive the next reorganization.

The last thing is to pick one process and redesign it from scratch as if AI had always existed, instead of adding AI to the process you already have. That's the equivalent of putting a motor on each machine, and it's where most of the gains will probably come from.

When companies ask me why their AI rollout isn't working, the question I've found most useful is a simple one: who in the company is better off if it works? If the honest answer is nobody in particular, then it probably isn't a technology problem, and a better model won't solve it.

I wrote a follow-up about what this means for the work itself, and for the jobs around it: How work has to change around AI.

---

Notes

[1] Last year an MIT report got a lot of attention for claiming that the large majority of corporate generative AI pilots showed no measurable return. People disputed the exact number, but it matched what a lot of people had seen happen in their own companies.

[2] Mollick uses the term for employees who use AI at work without telling their employer.

[3] Paul David, "The Dynamo and the Computer," 1990.

[4] This was widely reported in 2025.