Why AI should fit the business, not the other way around

Every service business already has a way of handling an enquiry. The phone rings or an email lands, somebody works out what it is, and a quote or a booking follows. When AI arrives, the easy mistake is to rearrange that sequence to suit the tool instead of asking the tool to fit the sequence.

The business is the fixed point

Software is the easiest thing in a small business to change. A subscription can be cancelled this afternoon and a setting altered in a minute. Everything else moves slowly: the customers who expect a callback, the supplier who only takes orders by email, the person who has always answered the phone.

So the burden of proof should run one way. The way the business works now is the default, and anything asking it to change has to make the case.

That idea sits under everything I write on this site. AI is only worth adopting when it fits a job you already do, and you can tell whether it did the job.

That does not make the business right about everything. Some steps exist only because nobody went back and questioned them, and those deserve to change. The point is about order.

Fix a process because it is wrong, then decide whether a tool helps with the fixed version. Changing a process because a tool needs it changed is the backwards route. It is also the easier one to be talked into.

What the national figures suggest

I have read the recent official figures on adoption, and they show a pattern worth taking seriously. The Office for National Statistics reported in July 2026 on UK businesses with 10 or more employees. Around 35 per cent now say they use AI, up from around 12 per cent in late 2023.

Over the same period, the average number of AI technologies used by each adopting business rose only from about 1.4 to about 1.6. The ONS describes adoption to date as “relatively shallow”.

The Federation of Small Businesses reports a similar shape among smaller firms. It found that 55 per cent now use AI, up from 20 per cent in 2023, while 92 per cent have concerns about its risks.

The government’s own AI Adoption Research, published in January 2026, asked businesses what had held them back. The most common answer, chosen by 71 per cent, was “We have not identified a use for AI in our organisation”.

These figures cannot tell me why adoption is wide and shallow, and I don’t want to pretend they can. My reading, and it is only a reading, is that tools are arriving before the jobs they are meant to do have been named. A business that cannot name the job has nothing to fit a tool to.

Three tests before you change anything

These are the three questions I would put to any AI proposal, including one from my own company. None of them needs technical knowledge. All of them need a pen and 20 minutes with the person who does the work.

Does it remove a step for a person?

Name the step and name the person. “Answering the phone” is too broad to test. “Taking a message when the only engineer is out on a job” is a step, and you can see who does it and how often.

Then look for the steps the tool adds. Someone may now have to check its output, copy a record from one screen to another, or tidy up after it. A tool that saves one person 10 minutes and costs another person 15 has only moved the work. Count both sides before you call it a saving.

Does the person doing the job agree?

The owner usually buys the tool, and somebody else uses it several times a day. If that person thinks it makes the job worse, they will quietly work around it. The owner will see the subscription on the bank statement long before they see the workaround.

So ask them first. Ask which parts of the job they would happily hand over, and which parts they would never let a machine do. Their answer is a better specification than any demonstration, because it comes from the person who knows where the job goes wrong.

Can you tell it worked, and turn it off on a Tuesday afternoon?

Before anything is switched on, write down what “worked” means in terms you can count without the supplier’s dashboard. Enquiries acknowledged the same day, say, or calls returned before close of business. If you cannot describe the count, you will not know whether the tool did the job.

Then check that you can stop it. You should be able to turn it off without losing the records it created and without breaking whatever depends on it. You should not need the one person who set it up. If stopping it would take a project, you have not reduced your risk at all.

Fit in practice: two jobs compared

Take two jobs a service business might hand to AI. These are illustrations, not accounts of any particular business.

The first is acknowledging an enquiry: a short reply saying the message arrived and when someone will be in touch. It removes a step that is easy to forget on a busy day, and the person who handles enquiries has little reason to object.

Every enquiry either got an acknowledgement or it didn’t, so the job is simple to count. It is just as simple to switch off.

The second is writing and sending quotes. It removes a much bigger step, which is why it is tempting. But a wrong price goes straight to a customer, and the person who prices jobs has good reason to be wary. The checking it needs may cost more time than it saves.

Quoting might still fit, with a person approving every quote before it leaves. As a job the system does unsupervised, it fails the tests. The first job passes all three on day one; the second is a later decision, made once the business has seen how the first one went.

What staying in control looks like in software

I should say where my own view comes from. I run Step Into AI (opens in a new tab), an AI platform for UK service businesses, which is live and working with a small group of pilot customers. Its first user is Manchester PC (opens in a new tab), the computer repair and IT support company I also run. Step Into AI calls it “customer zero”.

That gives me a commercial interest in this argument, and you should read what follows with that in mind. Step Into AI gives each automated action one of three settings (opens in a new tab). It can require approval first, run automatically under rules the business sets, or be escalated to a person. The company’s line for this is “AI does the work. You stay in control.”

That is one company’s design choice. It is not evidence that the approach works, and Step Into AI has not published results from its pilot. I mention it because it shows the third test built into the software instead of left to good intentions.

Whatever you buy, from me or anyone else, ask to see where that setting lives. Then ask who is allowed to change it.

What fitting the business costs

The subscription is the visible cost. The larger one is time. Someone has to describe the step properly and agree it with the person who does it. Someone has to check the output closely for the first weeks, write down what to do when it goes wrong, and show whoever covers holidays.

None of that appears on a price list. In a business of five people it all comes out of the same week as everything else. If there is no time to do it, the honest answer is to wait. I would expect a tool adopted without that work to end up as a subscription nobody uses.

Where I think this goes

What follows is a view, not a forecast. I expect the tools to keep getting more capable faster than small businesses can absorb them. If that is right, the limit for most small firms will be the owner’s time to name a job and check the result.

I would want several more years of the ONS series before I believed that with any confidence. In the meantime, the useful first step costs nothing. Write down one step in your business you would hand over tomorrow, and the name of the person who does it now. Then go and ask them.

The same principle applies to people, not just businesses. I wrote about it in how AI can help people with dyslexia find their words.