Most SME founders we talk to aren't stuck on whether AI works. They've seen the demos, read the case studies, maybe even run ChatGPT through a few tasks themselves. What stops them is narrower and more practical: where do we actually start, and what happens if we get it wrong?
That second question does more damage than people realise. Because the first deployment tends to get treated, consciously or not, as a referendum. If it goes well, AI is "a thing we do now." If it goes badly — a wrong answer reaches a customer, a mistake costs money, a stakeholder says "I told you so" — the whole initiative can stall for a year, regardless of whether the underlying technology was ever the problem.
That pressure pushes founders toward one of two bad choices. Either they pick something so low-stakes it proves nothing (a one-off internal experiment nobody relies on, quietly abandoned a month later), or they reach for something ambitious and customer-facing — a support chatbot, an automated quote tool — because that's the version of AI everyone's heard about. The second path is where most of the expensive disappointments happen, because customer-facing processes are usually the worst place to learn what your AI deployment is actually good at.
There's a better way to choose, and it isn't about finding the most impressive use case. It's about finding the one that pays back fastest with the least exposure while you're still learning how the tool behaves in your business. Three criteria do most of the work: volume, tolerance for error, and whether a wrong answer is recoverable.
Volume: is there enough of it to matter?
AI earns its keep on repetition. A task you do twice a year isn't worth automating, no matter how tedious it is — you'll spend more time setting up and checking the system than you'd ever save. A task that happens dozens of times a week, with a consistent shape each time, is where the economics start to work.
This is where founders often underestimate their own processes. Invoice coding. Meeting notes and action items. Supplier onboarding paperwork. Lead triage from a contact form. These don't feel like "AI projects" because they're unglamorous, but they typically happen often enough, and in a similar enough pattern each time, that a well-built system can take over the bulk of the work and let a person check rather than produce.
If you're not sure whether a process has enough volume, a rough test: could you write down the steps once and expect them to apply, with minor variation, the next fifty times this task comes up? If yes, it's a candidate. If every instance is genuinely different, AI will struggle and so will the person reviewing it.
Tolerance for error: how far off can it be before someone notices?
Not all mistakes cost the same. A draft summary that misses a nuance costs a few minutes of editing. A miscoded invoice that goes unchecked costs real money and possibly a compliance headache. A wrong answer given directly to a customer costs trust, and trust is slower to rebuild than almost anything else in a business.
The processes worth starting with are the ones where an error is cheap to catch and cheap to fix — where a human is already in the loop reviewing the output before it goes anywhere, so a mistake gets caught before it becomes a problem rather than after. This is less about picking a "safe" task and more about picking a task where your existing checks still apply. If someone already reviews invoice coding before it posts, AI doing the first pass doesn't remove that check — it just changes what the reviewer is checking.
Recoverability: can you undo it?
This is the criterion people skip, and it's usually the one that matters most. Ask: if the AI gets this wrong, what does it take to put it right?
An internal document that's miscategorised gets re-filed in five minutes. A meeting summary with an error gets corrected before anyone acts on it. These are recoverable. A customer-facing quote that's wrong, a support answer that's confidently incorrect, an automated email that goes out to a client list — these are much harder to walk back, because the damage happens the moment the output leaves your system, not when someone eventually notices.
This is the real argument against starting customer-facing. It's not that customer-facing AI doesn't work — plenty of it does, once it's been tuned on real volume and real edge cases. It's that you haven't done that tuning yet on your first deployment, and the cost of learning in public, in front of your customers, is higher than the cost of learning quietly, internally, first.
Putting the three together
Run a candidate process through all three and you get a reasonably honest answer about whether it's a good starting point. High volume, high tolerance for error, high recoverability — that's your first deployment. Low on any one of those, and you should either wait or build in more human checking before you lean on it.
What this usually points to isn't the flashiest use case in the business. It's something unglamorous: document processing, internal reporting, first-pass categorisation, draft generation that a person finishes. None of these make for an exciting announcement. All of them teach you, at low cost, how the system behaves — where it's reliable, where it needs tighter instructions, how much review it actually needs versus how much you assumed it would need.
That knowledge is what makes the second deployment faster and the third one more ambitious. Businesses that try to start with the ambitious one usually end up relearning these lessons anyway, just with a customer watching and a bigger bill for the rework.
What to do with this
If you're trying to find your own starting point, list the repetitive processes in your business — the ones that happen often enough to be worth the setup. For each, ask the three questions: how often does this happen, how bad is it if the output is wrong, and how easily can we undo it if it is. Pick the one that scores well on all three, not the one that would look best in a slide deck.
If you want a second opinion on where that starting point is in your business, get in touch. We spend most of our early conversations on exactly this question, because getting it right is what makes everything after it easier.