Something odd is happening with AI right now. Adoption is nearly everywhere, and measurable results are nearly nowhere.
McKinsey's surveys keep putting numbers on it. Most organisations now use AI somewhere in the business, but only a minority have managed to scale the more capable work, the agents and the automation that actually runs a process end to end, into how the company operates each day. MIT's research last year was blunter. Of the companies that ran generative AI pilots, roughly 95% saw no measurable effect on profit or loss.
Ninety-five out of a hundred. Not because the technology flopped in the demo. It usually performed. The pilot impressed someone, earned a nod in a meeting, and then quietly never reached production.
So what keeps going wrong?
The model was never the hard part
When a project stalls, the instinct is to blame the AI. Wrong model, wrong prompt, needs the newer version. Almost always, the real problem sits upstream of the model.
A pilot is a tidy demo. Someone picks clean inputs, runs it once, and shows the best case. Production is the opposite: messy inputs, real volume, people who didn't build the thing trying to use it, and a hundred edge cases nobody scripted. What breaks isn't the model's intelligence. It's the surrounding process that was never rebuilt to make use of it.
McKinsey found something worth sitting with here. The companies reporting genuine financial returns were about twice as likely to have redesigned their actual workflows before they settled on a modelling approach. They fixed how the work moved first. The AI came second.
What this means if you run a smaller business
If you're a company in the UK or Canada and someone has told you you're "behind on AI," ignore the pressure to buy the newest thing on the shelf. Speed of adoption isn't the game. Plenty of firms adopted fast and got nothing for it.
A better place to start is a single question: which one process, if it ran faster or more accurately, would move a number you actually care about? Revenue, time to invoice, support backlog, stock you're writing off. Pick that one. Then check whether your data can support it, rebuild the steps around the AI instead of bolting it on the side, and measure the before and after honestly, including the cases where it gets things wrong.
That's less exciting than a company-wide AI programme. It also tends to work, which the company-wide programmes mostly don't.
The honest version
I'll be straight about it: a lot of AI spend in the last two years bought demos, not outcomes. That isn't a reason to sit out. It's a reason to be picky. One useful, boring, well-built workflow that saves a team three days a month is worth more than ten pilots that wowed a room and then died.
The companies that pull ahead over the next couple of years won't be the ones with the most experiments running. They'll be the ones who took one or two and did the unglamorous work of making them real, measured, and trusted. That gap is wide, and it's worth being on the right side of it.