There’s a particular kind of meeting that happens in small businesses across the UK roughly twice a year. Someone — often the owner, sometimes a well-meaning supplier — has heard about an AI tool. The conversation drifts toward whether it might do the same for them. The maths gets waved at. A trial gets started. Six months later, no one can quite remember whether it’s saving any time, but the subscription keeps debiting.
This is the most common failure mode in UK SME AI adoption right now, and it’s entirely avoidable. The trick is doing fifteen minutes of arithmetic before you commit to anything.
What ROI actually means in this context
Return on investment is not a complicated idea. It’s the saving you make minus the cost you incur, expressed as a percentage of the cost. For an AI or automation project at a UK SME, the saving almost always comes from one of three places:
- Time recovered from manual work — usually the largest pot.
- Revenue that wasn’t being captured — missed enquiries, lapsed customers, unconverted leads.
- Cost that was being paid to someone else — aggregator commission, agency fees, third-party services that an in-house tool would replace.
The cost side is more familiar: a one-off build fee, ongoing licences, time spent integrating, training, and the very real opportunity cost of staff learning a new system instead of doing what they were doing before.
The honest answer to “is this AI project worth it?” is just one calculation: the saving has to be larger than the cost, by enough margin and quickly enough that the business is genuinely better off. Anything else is wishful thinking dressed up in slide decks.
The hours-per-week framework
For most UK SME automation projects — the kind that sit between £2,000 and £30,000 in build cost — the dominant ROI lever is time recovered. Here’s how to size it before you talk to anyone.
Step one. Pick the workflow. Be specific. Not “admin”. Not “customer service”. Pick a single end-to-end process: quote-to-invoice, enquiry-to-booking, Friday close-down, MOT reminders. Whatever the candidate workflow is, name it precisely.
Step two. Time-box it for one week. Get whoever does the work to run a stopwatch — phone notes app, paper, spreadsheet, doesn’t matter. Every time they touch the workflow, log how long. At the end of the week, sum the minutes.
Step three. Multiply by 52 and by an hourly cost. The hourly cost should be the all-in cost: salary plus employer NI plus pension plus a roughly 20% overhead allocation. For a £28k salary that lands somewhere around £19 per hour. For an owner-operator pulling £45k of effective income, it’s closer to £30. Use the right number for the role.
Step four. Ask what fraction is genuinely automatable. Not all of it is. Be honest. If a tool can take 70% of the work off the human, multiply by 0.7. If a tool can do 100% but a human still needs to spot-check, knock 10–15% off the saving for that. Anything that promises 100% automation with zero supervision is exaggerating; mark it down accordingly.
The output of those four steps is annual saving in pounds. Now compare it to the annual cost of the proposed solution: build amortised over three years, plus monthly licence fees, plus realistic training time at the same hourly cost.
FENCE0
A worked example
A removals company in Lincolnshire with three vans and one office administrator. The administrator spends Monday mornings going through Friday’s job sheets, retyping them into Sage to raise invoices, and chasing customers for the deposits that didn’t come through. She also spends most of Friday afternoon dealing with the queue of phone enquiries that built up during the week.
A week of stopwatch logging puts the figures at:
- Monday morning invoicing rekey: 3.5 hours
- Friday phone enquiry triage and quote-write-up: 4 hours
- Mid-week chase-ups and corrections: 2.5 hours
That’s 10 hours per week. At a fully-loaded £19 per hour, the all-in cost is £190 per week, or £9,880 per year.
A custom application that takes enquiries from the website form, generates quotes from a simple price grid, captures a deposit through Stripe, pushes the invoice into Sage, and sends payment chasers automatically would realistically take 80% of that workload off the administrator. So the saving is around £7,900 per year.
Compare that to a build cost of around £8,500 fixed plus a Silver retainer at £99 per month (£1,188 per year):
| Year 1 | Year 2 | Year 3 | |
|---|---|---|---|
| Annual saving | £7,900 | £7,900 | £7,900 |
| Cost | £9,688 | £1,188 | £1,188 |
| Net | −£1,788 | +£6,712 | +£6,712 |
Year one is a small loss because the build fee is large compared to one year of saving. But payback lands at month 14, and from year two onwards the platform pays for itself nearly seven times over per year. Over five years, the business is £26,860 better off.
That’s a real ROI calculation. It only takes fifteen minutes if you’ve already got the timesheet data.
The questions worth asking any AI or automation vendor
Once you have a number to compare against, the conversation with a vendor changes. Instead of being persuaded by demos, you have a single question: does the saving I’ve already calculated cover the cost they’re proposing?
A few specific things to ask:
- Where does the time saving actually come from? Make them point at specific workflow steps that get removed or shortened. “It will save you hours” is not an answer.
- What needs to change in your workflow for the saving to land? Most automation has a behavioural cost — staff need to do something differently. If that change is huge, the saving might not materialise.
- What’s the realistic accuracy? “99%” sounds like a lot until you realise it means 1 mistake in every 100 transactions. For invoicing, that’s a problem. For email triage, it’s probably fine.
- What happens when it’s wrong? Every AI workflow has a failure mode. Press until you understand what happens when the system makes a mistake — does a human catch it, or does the wrong thing land in front of the customer?
- What does the migration off look like? If you change your mind in 18 months, can you take your data and your customers somewhere else, or are you locked in?
The three traps that destroy ROI
Three patterns kill the ROI on AI projects more often than anything else.
The first is buying the platform before sizing the problem. If you can’t state the workflow, the hours, and the hourly cost before you’re shown a demo, the demo will write the answer for you. Always do the maths first.
The second is letting scope drift. “While we’re building it, can it also do…” is the most expensive sentence in software. Every additional capability inflates the build cost without inflating the original saving you calculated. If new things get added, they need their own ROI calculation; if they don’t pass, they don’t go in.
The third is forgetting about the change cost. New systems need training, new habits, new error-handling. A system that saves 10 hours per week but takes 6 hours per month to maintain has only saved 9 hours per week. A system that the team refuses to use saves zero hours per week. Budget the human side properly.
What good looks like
A well-scoped AI or automation project for a UK SME has the following shape. Build cost lands between £2,000 and £20,000 depending on complexity. Annual saving is between £5,000 and £25,000. Payback is somewhere between month 6 and month 18. After that, the system pays for itself many times over each year for as long as it’s in service.
If the numbers don’t look like that — payback longer than two years, or net loss in year three — the project probably shouldn’t go ahead in its current form. That doesn’t mean no AI. It means different scope, or different tool, or different workflow.
The good news is that for most UK SMEs there are several candidate workflows that do pass this test. Pick the strongest one, do the maths, build it well, then move on to the next one. That’s how AI actually compounds for a small business — one well-sized, well-built workflow at a time.
If you’d like a hand running the maths on a specific workflow in your business, that’s exactly what a Launchpad ROI assessment is. Thirty minutes, no obligation, and we’ll tell you upfront if the numbers don’t stack up.