Two in five UK small businesses that pay for AI have had at least one bill bigger than they expected. That's one of the findings in Pax8's latest survey of small and mid-sized firms, which included 350 UK businesses with between 5 and 499 staff.
The other numbers explain why. Only 12% of UK firms in the survey have a dedicated AI budget. 19% can't separate AI spend from the rest of their tech spend. And only 16% are confident they can work out what AI is giving them back.
So most firms are paying for something they can't see, against a budget they don't have, for a return they can't measure. A surprise bill is the obvious result.
Why AI bills creep up
A lot of AI is now sold by use, not by seat. You pay for what it reads and writes. That's fair, but it means the bill depends on how much work you give it.
The same survey found that 45% of UK firms building with AI are connecting it to other business systems, and 39% are building or testing agents that do multi-step tasks on their own. An agent that checks an order, looks up stock and drafts an email does a lot more work than someone asking a chatbot a question. More work means a bigger bill.
That's not a reason to avoid it. It's a reason to know what each piece of work costs.
Price it by the job
Stop thinking about AI as a monthly fee. Think about it as a cost per job.
Pick the jobs you want AI to do. Be specific:
- Check a new order against stock before it's released to the warehouse.
- Match a supplier invoice to the purchase order and goods-in record before it goes into Xero.
- Answer "where's my order?" from the courier tracking.
For each one, find out two numbers. What does it cost now, in someone's time? And what does it cost when AI does it?
The first number is usually easy to estimate. The second you can measure once it's running. If the AI cost per job is well under the people cost, and the result is right, keep it. If it isn't, you've found out cheaply.
You can't price a job you haven't named
This is where most firms come unstuck. "Use AI to help with admin" isn't a job. You can't measure it, so you can't control it.
Naming the jobs also tells you what the AI needs to touch. Checking an order needs to read stock. Matching an invoice needs to read purchase orders and write to Xero. That's the same list I've argued you should write down before you connect AI to anything. It does two things at once: it limits what the AI can do, and it gives you something to measure.
Put AI on the boring jobs first
The jobs that are easiest to price are the dull, repeated ones. Same steps, same systems, many times a week. They're also where AI tends to pay off first, because you can count them.
A one-off clever task is hard to price and hard to judge. A job you do 200 times a month is easy. You know what it costs now, and you'll know quickly whether AI is cheaper.
What to do this week
- List the three jobs you'd most like AI to take off someone's desk.
- For each, write down how often it happens and roughly how long it takes now.
- Check what your current AI tools charge, and whether it's per seat or per use.
- Give AI spend its own line in the budget, even if it's small.
That's enough to stop the surprise bill and to see, in a month, whether it's paying its way.
At OnDraft, this is how we approach it with Bud, which is in beta. It starts by understanding how the business works: the systems, the stock, the jobs. Then it can be given specific jobs, with a clear record of what it did. If you want to talk through which jobs in your business would be worth handing over, get in touch.
Sources
- Pax8 Q3 2026 SMB AI Pulse Report, UK findings, as reported by TechRadar Pro, 14 Oct 2026: techradar.com/pro/most-small-businesses-arent-reviewing-their-own-ai-usage-properly
- Compare the Cloud: comparethecloud.net/news/only-23-of-ai-engaged-uk-small-businesses-formally-review-the-ai-they-build-pax8-survey-finds