Where to start with AI if you run a UK business

Only 15% of UK businesses had adopted a single AI technology as of 2022, according to research from the University of Manchester and the University of Edinburgh, with adoption concentrated in London, the South East and larger firms. If you run a smaller company outside those places and have not started, you are in the majority, not behind.

That research also names why: skills and data, not software. Which is the useful part, because it means the first move is not choosing a tool.

Most advice starts with products. That is backwards, and it is why so many first attempts end with an unused subscription. Start with a process.

The wrong first question

“Which AI tool should we get?” produces a licence, a login, and three months later a discovery that nobody knows what to do with it.

The right first question is “which piece of repeated work costs us the most and needs judgement the least?” That is answerable in an afternoon and it tells you what tool, if any, you need.

Step 1: list the repeated work

Sit with whoever does the work and list every task that happens more than once a week and takes more than fifteen minutes. Not projects. Tasks.

Typical answers from small UK firms: writing quotes, chasing invoices, summarising calls, answering the same twenty customer questions, formatting reports, pulling numbers from PDFs into a spreadsheet, first-pass CV sorting, writing job adverts.

Aim for ten to fifteen. This takes about an hour and it is the only research you need.

Step 2: score each one on two axes

For each task, two questions with a number from one to five.

How much does it cost? Frequency times duration times who does it. A two-hour task done weekly by someone senior costs far more than a daily ten-minute one done by anybody.

How much judgement does it need? Five means a wrong answer has real consequences and only an experienced person can tell. One means the answer is in a document somebody could look up.

Start with high cost and low judgement. That is the whole method.

Something scoring five on judgement is a bad first project no matter how expensive, because you will not be able to tell whether the output is right, and an unverifiable win is not a win.

Step 3: pick one, and only one

The single most common failure is starting three things.

One process, one owner, one deadline. If the first works you will have learned enough to pick the second properly. If it fails you will know why, which a three-way pilot never tells you.

Step 4: define what success is before you start

Write down, before you touch a tool:

• The number today. Hours per week, or cost per month, measured not guessed.
• The number that would make this worth keeping. Be specific.
• The date you decide. Two weeks is usually right.
• What would make you stop. Quality falling below a stated bar, or the checking taking longer than the doing.

Without this you get “it feels quicker”, which is how unused subscriptions survive for a year.

Step 5: run it for two weeks, by hand

Do not buy anything yet if you can avoid it.

Most first projects can be tested on a free or low-cost tier of a general assistant, by pasting the work in and looking at what comes back. That is a crude test and it answers the only question that matters at this stage: does the output need so much correction that it is not worth it?

Two weeks of doing it manually alongside the existing process tells you more than any vendor demo. It also produces the thing you need for a real decision: your own numbers rather than a case study about somebody else.

Before any customer or employee data goes into a tool, check two things: whether your contract with the vendor allows it, and whether the project needs a data protection impact assessment. The second takes twenty minutes: do you need a DPIA?

Step 6: decide on the date, honestly

Three outcomes, all fine:

It worked. Now make it routine: who does it, what they check, what happens when it is wrong. This is where most of the value is and where most people stop paying attention.

It half worked. Usually the process was too broad. Narrow it to the sub-task that worked and re-run.

It did not work. Write down why and move to the next item on your list. A fortnight and no licence spend is a cheap answer to a real question.

What this deliberately avoids

Buying a strategy. A consultant’s AI strategy for a 20-person business is usually a list of the same tools you would have found yourself.

Starting with the hardest problem. The instinct is to point AI at the thing that hurts most. That thing is usually high-judgement, which makes it the hardest place to tell whether it worked.

Committing to annual plans early. Monthly costs more per month and far less in total when the answer turns out to be no.

Automating a broken process. If the task is only necessary because something upstream is wrong, fix that instead. This is the most valuable thing on this page and the least followed.

What next

When something works and you want to make it permanent, the compliance side becomes real rather than theoretical:

• Do you need a DPIA before using AI on customer data?
• UK GDPR and AI: when you can let software decide, if the process makes decisions about people rather than just handling their information
• UK AI funding: what your business can actually get, though the honest answer for most adoption projects is that grants are not for this

If none of the tasks on your list scored high cost and low judgement, that is a real result. Some businesses do not have an obvious first AI project, and the time not spent looking for one is worth more than a subscription.