How to implement AI in a small business: a 90-day plan

Implementing AI in a small business works best as one 90-day cycle against one process: month one to choose the process and measure it honestly, month two to run the tool alongside the existing way of working, month three to decide whether to scale, fix or stop. The tool is the easy part. The measuring and deciding are the work.

Why do most small-business AI projects stall?

Not because the technology fails. Because nobody agreed in advance what success would look like, so there is nothing to point at in week twelve.

The national picture backs this up. Around 35% of UK businesses with 10 or more employees reported using at least one AI technology in June 2026, alongside 28% of businesses with fewer than 10 employees and 49% of those with 250 or more (ONS, published 20 July 2026). But the average adopting business runs only about 1.6 AI technologies, which the ONS notes implies relatively limited transformative impact so far. Adoption has spread sideways, not downwards. Plenty of businesses have a subscription; far fewer have changed how a job actually gets done.

The government's own read of the barriers is the same three things it has been for years: capability, cost and awareness (SME Digital Adoption Taskforce: 2026 update, GOV.UK). None of those are solved by buying more licences. They are solved by picking something small and finishing it.

Days 1–30: which process, and what does it cost you today?

The first month produces no AI at all. It produces a target and a baseline.

Pick one process. The candidate you want is high-volume, low-judgement and already written down somewhere. Drafting standard replies, summarising long documents, extracting fields from supplier invoices, first-line triage of enquiries — these have stable right answers that somebody could check. Anything where the correct answer lives only in one person's head is a documentation problem wearing an AI costume. Write the process down first; you may find you no longer need a model.

Measure it before you touch it. Pick a single metric a sceptical colleague would accept — minutes per invoice, hours to first draft, percentage of enquiries closed without a person — and record it for a fortnight. This is the step teams skip, and skipping it is why so many pilots end in "it feels faster."

Name one owner. Not a committee. One person who is accountable for the pilot and has the hours to run it.

Settle the data question early. Under UK GDPR a data protection impact assessment is legally required before processing likely to result in a high risk to people's rights and freedoms, and the ICO's guidance on AI and data protection is clear that AI use will meet that threshold in the vast majority of cases. If your process touches customer, employee or applicant records, a DPIA belongs in month one. So does deciding what staff may and may not paste into a public chatbot — the NCSC's guidelines for secure AI system development are a reasonable starting frame for the security conversation.

Days 31–60: how do you run a pilot that proves something?

Run the tool in parallel with the existing process, not instead of it. Both do the same work for four to six weeks. That costs you some duplicated effort and buys you the only thing that matters: a like-for-like comparison, and a safety net if the output is worse than you hoped.

Three rules make a parallel pilot honest:

1. Write the success criteria down before you start, with a number and a date. "Cut invoice processing from twelve minutes to under six by 15 November" is a criterion. "Improve efficiency" is a wish.
2. Have a human check every output for the first two weeks, and keep a tally of what needed changing. The error pattern tells you more than the error rate — models tend to fail in specific, learnable ways.
3. Keep the scope frozen. The single most common way a 90-day pilot becomes a nine-month project is someone saying "while we're at it, could it also…".

Use whatever your team already has where you can. If you are a Microsoft shop, that usually means the assistant bundled with the tools people already open. Where a bespoke build genuinely is the answer, that is what our AI development services cover, and our note on what AI consulting costs UK SMEs sets out the rate bands, and free public support exists too — the Business Growth Service and Help to Grow signpost SME digital and AI support at no cost, per the taskforce update above.

Days 61–90: scale, fix or stop?

Month three is a decision, and all three outcomes are legitimate.

Scale if the metric moved and the checking burden fell over the pilot. Roll it to the rest of the team, write the process down properly, and pick the next candidate process. Do not run two pilots at once until you have finished one.

Fix if the metric moved but the outputs still need heavy correction. Usually the problem is inputs, not the model: inconsistent templates, missing reference material, a process that was never standardised. Give it one more 30-day cycle with a specific fix, not a vague "more training."

Stop if the metric did not move. This is the outcome nobody plans for and it is the cheapest possible result — you have spent one process and 90 days learning something concrete, rather than paying for licences nobody opens. Write down why it failed. That note is worth more than the pilot.

What does 90 days actually cost?

Mostly attention. In our experience the realistic ask is a few hours a week from the process owner across the quarter, concentrated at the start and end — that is our own planning assumption, not a surveyed figure, and it will be higher if your process has never been documented.

Licence cost varies too much to quote usefully: assistant seats, per-conversation pricing and consumption-based capacity all exist, and vendors reprice them frequently. Get a current quote for the two or three tools on your shortlist rather than budgeting from anyone's blog, including ours. The bigger cost, and the one that surprises people, is the documentation you should have written years ago.

For a worked example of what this looks like in one function, see our piece on AI for customer service in small businesses. For the wider view of where automation pays back across an operation, our AI automation service page covers process selection, and the AI practice overview sets out how we scope this kind of work.

Frequently asked questions

How long does it take to implement AI in a small business?

One well-chosen process takes about 90 days end to end: roughly a month to pick the process and measure it honestly, a month to run the tool in parallel with the existing way of working, and a month to decide whether to scale, fix or stop. That is our own planning band, not a benchmark. The tool takes an afternoon; the measuring and the deciding take the other 89 days.

What should a small business automate with AI first?

The task that is high-volume, low-judgement and already written down somewhere. Drafting standard replies, summarising long documents, extracting fields from invoices and first-line triage all qualify. Anything where the right answer only exists in one person's head is a documentation problem wearing an AI costume — write the process down first, then decide whether it still needs a model.

Do we need a DPIA before using AI in our business?

Often, yes. Under UK GDPR a data protection impact assessment is legally required before processing that is likely to result in a high risk to people's rights and freedoms, and the ICO's guidance on AI and data protection says AI use will trigger that threshold in the vast majority of cases. If your pilot touches customer, employee or applicant data, budget for a DPIA in month one rather than discovering it in month three.

How many UK businesses are actually using AI?

Around 35% of UK businesses with 10 or more employees reported using at least one AI technology in June 2026, against 28% of those with fewer than 10 employees and 49% of those with 250 or more, according to the Office for National Statistics (published 20 July 2026). Depth is the more revealing number: the average adopting business runs about 1.6 AI technologies.

How do we know whether the pilot worked?

By comparing it with a number you captured before you started. Pick one metric that a sceptical colleague would accept — minutes per invoice, hours to first draft, percentage of enquiries resolved without a person — measure it for two weeks before the tool arrives, and measure the same thing the same way after. Without that baseline, "it feels faster" is the only verdict available.

---

If you would rather not spend the first month working out which process to point at, that is exactly where our AI strategy engagements start.

Easy Insight is a UK consultancy for AI, web, apps and data — senior specialists only, no juniors.