Most of what gets sold as AI is a process problem with a subscription attached

Twenty-nine per cent for emails. Four per cent for a process

Asked which parts of the business they use AI in, small employers said communications 29% of the time and marketing or advertising 27%. For automating a process: 4%. That is the National Federation of Independent Business, surveying 521 owners in April 2025, and the 4% counts everyone who is even planning to.

The same answer arrives from a completely different instrument on another continent. Eurostat asked European firms of 10 to 49 people whether they use AI to automate a workflow or a decision. 4.1%.

Two instruments, two continents, four percentage points apart. What owners are buying is a writing aid. What they were sold was work coming off the business.

The obstacle is not the price and not the skills

The Office for National Statistics asked UK firms what was preventing or delaying them from adopting AI. Cost was named by 21%, and a shortage of AI skills by 16%. The most common answer, at 39%, was difficulty identifying which activities to use it on.

That is a diagnostic problem wearing a technology costume. Nobody sells a solution to it, because there is no licence fee attached.

A third of adopters cannot name what it is for

Italy’s statistics office asks firms that report using AI to select the business purposes it serves. The share selecting none of them rose from 15.5% to 33.4% in a single year, and 83% of those firms have between 10 and 49 people.

That is a national statistics office measuring a tool that was bought and never attached to anything. JPMorgan Chase Institute went at it from the payments side instead, tracking 4.6 million small business accounts, and found that 72.5% of the firms paying for AI in 2025 paid for exactly one service. One subscription, and then nothing else.

Automation applied to an unmeasured process

There is no study measuring what happens when you automate a process nobody timed first. I had three independent researchers look, and all three reported the same absence.

The absence is the finding. The industry sells automation to businesses that have not standardised the thing being automated, and has never tested whether that works. What AI reliably does is make an existing process faster. If the process was wrong, you now have a faster wrong process and a monthly bill.

What to do before anybody quotes you

Four questions, and they cost nothing. How many hours did you need to save, and how many did you save. Which process did it go into, and by how much did the time on it drop. Was there a baseline before you started. And would removing a step have done the same job.

Then one check that needs no cooperation from anyone: open the card statement and count the AI subscriptions. The owner’s answer and the statement are frequently different documents.

If the answers point to a real job, the next steps are ranking it, as in what to automate first, and testing it honestly, as in the AI pilot scorecard.