AI ROI for small businesses: five patterns that pay for themselves
Easy Insight Team ·
An AI project pays for itself when the hours or errors it removes are worth more than its full running cost, including checking time. For most small businesses that means narrow, high-volume, text-heavy tasks, often on software they already pay for. Below are five patterns that tend to pass that test, and a simple way to check your own.
How do you work out AI ROI for a small business?
Take the hours the tool saves each month, after the time spent checking its output, and multiply by what those hours cost you. Add any errors or rework it removes. Subtract the full running cost: licences, set-up spread over a year, and support. If the result is positive within a year, it is paying for itself.
The bar is lower than most people expect. Here is the arithmetic for an AI assistant licence, using two published figures:
- Licence: Microsoft 365 Copilot Business is £16.10 per user per month, paid yearly and excluding VAT, on Microsoft's UK pricing page as of October 2026. A first-year promotional price of £13.80 applies to annual plans bought between 1 July and 31 December 2026. It is an add-on to an eligible Microsoft 365 Business plan.
- Time: median gross hourly pay for full-time UK employees was £19.67 in April 2025, excluding overtime (ONS Annual Survey of Hours and Earnings, published October 2025). That is pay before employer National Insurance and pension, so it understates what an hour really costs you.
| Illustrative 10-person team (our arithmetic) | Per user | Team of 10 |
|---|---|---|
| Copilot Business at list price, per year | £193.20 | £1,932 |
| Hours that must be saved to break even, per year | 9.8 | 98 |
| Break-even time saved, per month | about 49 minutes | about 8 hours |
| Value if each person saves a net 15 minutes a day over 220 working days | £1,082 | £10,820 |
The 15 minutes and the 220 days are our assumptions, chosen to be cautious; swap in your own. The point of the table is the shape, not the totals: a per-seat assistant needs under an hour a month per person to cover its licence, so the real question is whether every licensed person saves that, not whether the tool is clever.
What does the UK evidence say so far?
It is encouraging, thin, and almost entirely self-reported.
The government's AI Adoption Research, from DSIT, surveyed 3,500 UK businesses with five or more staff between February and May 2025 (last updated February 2026). It found:
- 16% of businesses used at least one AI technology, and 14% of micro firms.
- Of the businesses using AI, 75% reported improved workforce productivity and 34% reduced operating costs. Only 10% reported no impact so far.
- On revenue the picture is flatter: 77% of adopters saw no change, 12% an increase.
- Spend is small for most. Among those who gave a figure, median 2024 AI spend was about £2,000, against a mean of about £19,000 pulled up by a few big spenders.
- Micro and small firms mostly tested AI by trial and error, and many interviewed businesses had no formal metrics linking AI to revenue or productivity.
The largest published UK trial of an AI assistant is the cross-government Microsoft 365 Copilot experiment (published June 2025). About 20,000 civil servants had licences from 30 September to 31 December 2024. They reported saving 26 minutes a day on average, with drafting documents the biggest single saving at about 24 minutes. The report is candid that these are self-reported estimates, that it could not tell how the saved time was spent, and that the professions least satisfied with Copilot saved the least time.
So the honest summary is: most adopters believe AI is making them more productive, few have measured it, and revenue gains are rare so far. Public, independently measured ROI figures for individual UK small businesses are scarce, which is why the rest of this post is about patterns and method rather than borrowed success stories.
Which AI projects tend to pay for themselves?
These five patterns are our view, drawn from the evidence above and from how the costs work. Each one makes the arithmetic easy to pass.
1. High-volume drafting on software you already pay for
Emails, proposals, minutes, first drafts of policies. Text generation is by far the most common AI use, used by 85% of adopters in the DSIT survey, and drafting was the biggest saving in the government trial. It pays back because there is no build cost, the licence is small, and the work happens many times a week. It fails when the licence goes to people who draft little.
2. Turning documents into data
Invoices, delivery notes, order forms and application forms all arrive as documents and end up as rows in a system. Extracting the fields is repetitive, the right answer is checkable, and the volume is countable, which makes the before-and-after easy to measure. Our guide to AI invoice processing for SMEs covers where this works and where human approval has to stay.
3. First-line answers to repeat questions
Where the same twenty questions arrive every day, an assistant that answers from your own policies and order data can take a share of them off the inbox. The saving is measurable in tickets, and the risk is controlled by handing anything unusual to a person. Our post on AI for customer service in small businesses prices the off-the-shelf options.
4. Removing a bottleneck rather than a person
Some of the best returns are not hours saved but work that stops waiting: quotes sent the same day, month-end closed sooner, a backlog cleared without hiring. These are harder to value, so value them honestly, for example the cost of the temporary staff you no longer need, and never count a saving twice.
5. Small, cheap and easy to stop
The DSIT median spend of about £2,000 is a clue. Projects that pay back tend to start as one narrow use, tested on real records with a pass mark agreed in advance, and stopped if they miss it. Our guide to running an AI proof of concept sets out that discipline. Scaling up comes after the first use has paid for itself, not before.
Why do so many AI projects never pay back?
Usually for one of four reasons: nobody measured the task before, so nobody can prove a saving; checking the output takes almost as long as doing the work; licences are bought for people who rarely use them; or the project grows in scope before the first use has paid back.
Two cost traps are worth naming. First, per-seat licences: rolling an assistant out to everyone at once is the fastest way to pay for seats nobody uses. Start with the people who draft most and widen it on evidence. Second, metered billing: from 1 December 2026, usage-based billing is on by default for new Copilot Business licences bought through a partner, as our 16 October news roundup explains. Set a spending limit you chose rather than accepting the default.
How should you measure it?
Keep it light, but do it before you start:
- Time the task now. A week of rough logs from the people who do it is enough: how many, how long, how many errors.
- Run the AI version alongside the old way for four to eight weeks, on real work, with a person checking the output.
- Count the checking time as a cost, not a footnote.
- Decide at the end: keep, fix or stop, against the break-even figure you worked out first.
Our 90-day plan for implementing AI in a small business walks through that cycle, and the AI consulting cost guide covers what outside help adds to the cost side. If you would rather choose the first process with someone who has done it before, our AI automation service starts with exactly this measurement, and AI strategy covers the wider roadmap. More on what we do is on our AI consultancy page.
Frequently asked questions
How do you calculate the ROI of AI for a small business?
Take the hours the tool saves each month, after the time spent checking its output, and multiply by what those hours cost you. Add any errors or rework it removes. Subtract the full running cost: licences, set-up spread over a year, and support. If the result is positive within a year, it is paying for itself.
How much time does a Copilot-style AI assistant save?
The best public UK evidence is the government's own trial: about 20,000 civil servants used Microsoft 365 Copilot from October to December 2024 and reported saving 26 minutes a day on average. The figure is self-reported, and the report itself says to treat it with care, so measure your own team before relying on it.
How many hours does Copilot Business need to save to break even?
Not many. At £16.10 per user per month (Microsoft's UK list price, paid yearly, excluding VAT, as of October 2026) and the ONS median full-time hourly pay of £19.67, it breaks even at about 49 minutes saved per user per month. The harder part is making sure each licensed person actually saves that.
Why do small business AI projects fail to pay back?
Usually for one of four reasons: nobody measured the task before, so nobody can prove a saving; checking the output takes almost as long as doing the work; licences are bought for people who rarely use them; or the project grows in scope before the first use has paid back.
Easy Insight is a UK consultancy for AI, web, apps and data — senior specialists only, no juniors.
Next step
Wondering where AI would actually pay in your business?
An AI strategy and readiness review (from £2,500) tells you where AI pays, where it doesn't, and what to leave alone. The price is fixed in writing before we start, and the advice comes from a team that runs AI in its own products.