What AI Actually Costs
a Small Business
The licence is the cheap part and the part everyone quotes. Here is the rest of the bill, and what a first project realistically comes to.
Quick Take
Per-user licences for mainstream business AI assistants sit in the tens of pounds per person per month. That is rarely what makes or breaks the business case. The costs that decide it are the time of whoever runs the thing, the time spent checking output, and the change in how people work - and those are the ones no proposal itemises. Budget for a small first project in the low thousands, not the low hundreds, and be ready for the answer to be that it is not worth it.
What the Tools Cost
Three shapes of pricing, and it matters which one you are signing.
Per user, per month
The familiar model, and the easiest to budget. Microsoft's published UK pricing for Copilot business plans sat between roughly £14 and £25 per user per month on annual billing when we last checked in July 2026, with other mainstream suites in a similar band.1 Prices in this market move, so treat any figure you read, including this one, as a prompt to check rather than a quote.
Usage-based
You pay for what you consume, which is fine until consumption becomes enthusiastic. The same seat can cost very different amounts in consecutive months, and the invoice arrives after the usage. If you sign a usage-based deal, set an alert at a number you would be uncomfortable exceeding, on day one, before anybody starts.
Bundled into something you already buy
Increasingly the AI feature arrives inside software you already pay for. Worth auditing before you buy anything new: a fair number of businesses we speak to are paying for a capability twice because nobody checked what the existing subscription now includes.
The Costs Nobody Quotes You
In our experience these outweigh the licence by a considerable margin in year one.
- • Somebody has to run it. Choosing the tool, setting it up, answering questions, deciding what to do when it gets something wrong. That is a real person's real hours, usually one of your better people, and it does not appear on any quote.
- • Checking the output. The saving is the time to produce minus the time to verify. On tasks where verification is expensive, the sum comes out negative, and that is the single most common reason a promising pilot quietly stops.
- • Changing how people work. The technology is the easy half. Getting five people to do a familiar task in an unfamiliar way, consistently, after the novelty has worn off, is the half that decides whether any of it lasts.
- • Getting at your own data. If the information lives in an old system, in scanned documents, or in one person's head, that is the actual project. Sometimes it is worth doing on its own merits - see the true cost of technical debt - but it is not an AI cost and it should not be smuggled into an AI budget.
- • The experiments that fail. Budget for two or three tries. Anybody promising the first one will work is guessing, and a first attempt that fails cheaply is a good result rather than a wasted one.
What a First Project Really Comes To
Assume you follow the ninety-day method: pick candidates, choose one, test it with two or three people, decide. The bill has four parts.
- ✔ Licences for the trial. Two or three seats for three months. Tens of pounds a month each. This is the part people focus on and it is close to noise.
- ✔ Internal time. Someone senior enough to judge the output, a few hours a week for six to eight weeks, plus the participants' learning time. Cost it at their actual rate rather than pretending it is free.
- ✔ Any setup work. Connecting to your systems, or preparing documents so the tool can use them. Ranges from nothing to a genuine project depending on where your data lives.
- ✔ Outside help, if you use it. Optional. Our opportunity review is fixed-scope at two to three weeks; day rates sit in the same range as our fractional CTO work and the pricing guide explains what drives the number.
For most SMEs a properly run first project lands in the low thousands all in, dominated by internal time. If a proposal arrives with a much larger number, the question to ask is which of the four parts it is buying, and why yours is different.
If the project is about tooling for a development team, the numbers are better documented and rather higher per head - we cover them separately in measuring the ROI of AI coding tools.
What It Costs to Do Nothing
Usually less than you are being told. The businesses that will be damaged by waiting are the ones whose competitors are using AI to do something a customer can feel: answer faster, quote faster, price better. If nobody in your market is doing that yet, a year of watching costs you very little and saves you the tuition fees other people are currently paying.
Where waiting genuinely hurts is quieter. Your staff adopt tools on their own and nobody knows what has been sent where - a governance problem rather than a competitiveness one, and covered in is our data safe. And the groundwork compounds: businesses whose information is organised and whose processes are written down can adopt anything quickly, whereas businesses whose data is a mess will still be untangling it whenever they start.
That groundwork is worth doing regardless of what you decide about AI, which makes it the safest thing to spend on first.
Five Ways to Keep the Bill Down
- ✔ Audit what your existing subscriptions already include before buying anything new.
- ✔ Trial monthly, not annually. The discount for committing early is smaller than the cost of being wrong.
- ✔ Pick a task where checking the output is quick. It is the difference between a saving and a shuffle.
- ✔ Put a spending alert on any usage-based service before the first person logs in.
- ✔ Agree in advance what result would end the trial, and honour it. The expensive projects are the ones nobody was allowed to stop.
If you are weighing building something custom against buying a product, the economics are the familiar ones and our build vs buy framework applies unchanged. The AI part rarely alters that decision as much as vendors suggest.
Sources
- Microsoft - Microsoft 365 Copilot for business, UK pricing.Checked 28 July 2026. Published per-user prices change; confirm current figures before budgeting.
- Gartner - Over 40% of agentic AI projects will be canceled by end of 2027.Escalating costs named as a leading cause.
Want This Costed for Your Business?
A free half-hour and we will tell you roughly what a sensible first project would run to, or that you should not start one yet.