Where to start with AI in a small UK business

A small business should start with AI on one admin job: one that comes round every week, follows rules the person doing it could write down, and ends in something a person can check before it counts, such as a chased document, an order entered or a reminder sent. Choose the job before you choose any tool, write down what it involves and what stays with your people, and agree how you will judge it before it begins.

Choosing the job is where many businesses stall. When the ONS asked UK firms with 10 or more employees about AI for its Management and Expectations Survey of 2023, the barrier they reported most often was difficulty identifying activities or business use cases (39%), ahead of cost (21%) and the level of AI expertise or skills (16%). The government's AI Adoption Research, from interviews in 2025 with 3,500 private sector businesses of 5 or more employees, found the same: the factor most often given as holding back adoption, now or in the past, was "We have not identified a use for AI in our organisation" (71%).

Why is it hard to know where to start?

A tool does not tell you which part of your week to give it. The ONS's figures for June 2026 show how far use has gone. Around 35% of UK businesses with 10 or more employees reported using at least one AI technology, up from around 12% in September 2023, yet only 10% of those using it said they used it extensively. Adoption so far, in the ONS's words, has been "relatively shallow".

The way through is to start from the work. Every business already knows its own admin: the jobs that pile up before a deadline, the ones that wait for the one person who knows how, the ones done in the evening. One of those is your first job for AI.

Which admin job should we hand to AI first?

Take the jobs your team complains about and test each one against five questions.

1. Does it come round every week? A job done often is worth setting up, and gives you enough cases to judge it on.

2. Could the person who does it write the steps down? If two people do it two different ways, agree one way first.

3. Does it arrive as emails, documents or forms? Work that starts in an inbox and ends in a record is the kind an AI worker reads and writes.

4. Is a mistake caught before it does harm? A good first job ends in a draft, a queue or a held record that a person releases, not in a payment or a decision about someone.

5. Will someone own it? One person in the business has to brief it, answer its questions and look at what it did.

Jobs that pass usually look like these. A wholesaler keys orders that arrive by email into its order system, and the office releases them (order processing). An architects' practice chases missing timesheets before each invoice run, and the managers approve the hours (timesheet chasing). A letting agency logs each repair report, gathers what the contractor needs and keeps the tenant updated, with emergencies and spending left to a person (maintenance request coordination). Each sector page names the admin that fills the week in that trade.

Which jobs should wait?

Work that is different every time and leans on experience is a poor start, and so is anything that is really advice or judgement. A job nobody has time to brief or review should wait until someone does. So should a job whose main system offers no way in for other software, until you know what its maker allows on your licence.

How do I know if a task is right for an AI worker?

Write it down on one page: what starts the job, the steps in order, the systems it touches, what waits for a person, what a finished piece of work looks like and how many come in a month. That page is a role specification. If you can fill it in, the task is a candidate. If you cannot say what finished looks like, it is not ready, and writing the page will have shown you why.

Then agree two or three measures in your own figures, such as documents back before the deadline or orders entered on the day they arrive, and a date to review them. The probation period is that review, made formal.

Is an AI assistant enough, or do we need an AI employee?

An assistant such as ChatGPT or Copilot helps one person do a task faster, and that person still does the job. An AI employee is given the job itself: a written role, access to the systems it needs and a named person who checks its work (see what an AI employee is). For a first job with steps you can write down, an AI employee can take the whole job off a person's desk; for drafting and research, an assistant may be all you need.

What stays with a person?

Decisions stay with your people from the first day: whether to take on a client or a tenant, advice of any kind, anything a regulator expects a named person to sign, every payment, and any conversation with someone who is upset. In the government's follow-up interviews most businesses said having a person check AI outputs was of critical importance, and the role should say exactly which outputs wait for that check (see human in the loop). Where the job touches personal data, the Information Commission's Office says organisations remain responsible for data protection compliance of the agentic AI they develop, deploy or integrate. The role should therefore also say what the worker may open and change.

For a small business the first AI worker would take one repeatable job, written down, and every decision, every payment and any conversation with someone who is upset would stay with your people. If you can name the job but have not yet written it down, talk to our agent and scope that first role together.