If you run a growing UK business without a CTO, AI pressure arrives in a familiar way. A peer mentions ChatGPT. A board paper asks for an “AI strategy”. A supplier demos something shiny. Meanwhile the real work — orders, stock, quotes, chasing — still runs on spreadsheets, email and tribal knowledge.
You do not need a strategy deck. You need a clearer answer: where would AI genuinely help this business, what would it cost to try, and what should we ignore for now?
What “useful AI” looks like in an SME
In owner-managed firms, the wins are rarely science-fiction. They are usually small, repeated pieces of work that steal hours and create errors:
- Triaging enquiries and drafting a first reply for a human to check
- Turning meeting notes or site visit notes into a structured first draft
- Pulling status across tools — order, stock, delivery — without re-keying
- Summarising long email threads into decisions and actions
- Flagging exceptions in day-to-day ops data before a customer complains
Notice the pattern. The business still owns judgement. AI drafts, sorts, summarises or spots. People decide.
If a proposal starts with “replace the team” rather than “remove the grind”, treat it with caution.
Where AI is usually the wrong tool
AI is a poor fix when the underlying problem is process or data:
- Nobody agrees what “correct stock” means
- The same order lives in three places and none of them match
- A process only exists in one person’s head
- You cannot export clean data from the system that matters
- Compliance or customer risk needs a deterministic audit trail, not a probable answer
In those cases, better process, clearer ownership, or a systems choice will pay off before any model does. AI on top of messy foundations mostly produces confident mess.
A simple filter before you spend money
For each idea, write four lines in plain English:
- The work today — who does it, how often, what goes wrong
- The proposed help — draft, triage, summarise, extract, flag
- The human check — who reviews before anything leaves the business
- The stop rule — how you will know in two to four weeks whether to continue, change, or drop it
If you cannot fill those in, you are not ready for a pilot. You are still exploring a slogan.
Score ideas roughly on benefit, risk and effort. Keep one first pilot. A twelve-item wishlist is how projects stall.
How to run a first pilot without betting the company
A good SME pilot is boring on purpose:
- One workflow, one team, a fixed timebox
- Real data — not a polished vendor demo on sample stock
- A named owner inside the business
- Success criteria you can explain to a non-technical director
- An exit: turn it off without rewriting how the company trades
Measure hours saved, errors avoided, or speed to a decision — not “AI adoption”. If the only metric is that people used the tool, you have learned almost nothing.
Questions worth asking any AI vendor or adviser
Ask these before you buy licences, retainers or “transformation” programmes:
- Which specific workflow are we improving — in our language, not yours?
- What data does it need, where does that data live, and who can see it?
- What happens when the answer is wrong — and who is accountable?
- Can we run a proof on our own material before a long contract?
- What will we still need people to check every time?
- If we stop in three months, what do we keep and what disappears?
Vague answers on data, accountability or exit paths are information. Prefer advisers who will also tell you where AI is the wrong tool.
Red flags
- Urgency without a named workflow (“you need AI before competitors do”)
- A platform recommendation before anyone has watched how your team works
- Customisation promised as a substitute for a clear use case
- No human review step for customer-facing or financial output
- Success defined as “rollout complete” rather than a business result
Failed agency experiences leave a mark for good reason. The antidote is not cynicism. It is a sharper brief and an independent sense-check.
You can do useful thinking before you buy anything
You do not need a consultancy engagement to start. Map the five most repetitive admin tasks in the business. Note where work is re-typed, chased or stuck waiting for one person. Circle the ones that are frequent, rules-based enough to draft or triage, and safe enough to review.
That list alone often beats a generic “AI strategy”.
If you want help turning that list into a ranked register and a first pilot with honest effort estimates — including a clear “not yet” if that is the truth — an AI Opportunity Review is one structured way to do it. It is a means, not the point. The point is a decision you can defend to yourself and your team.
Next step
Book a Discovery Call if AI pressure is already on the agenda and you want a grounded next move — review, a smaller experiment, or nothing for now.