How to run a practical AI audit on your own business

Most AI spending in small service businesses is wasted for one reason: the tool was chosen before the problem was defined. The order matters more than the technology. What follows takes about two hours and produces a shortlist you can defend.

Step 1 — Map where the time and the enquiries go

Pick the two or three processes that feel most expensive. For each, write the steps in order, and against each step note two things: who touches it and what it waits for.

The waiting is where the findings are. Work that is genuinely slow is usually visible and already resented. Work that sits in a queue for two days because it needs one person's attention is invisible and far more expensive.

Step 2 — Find the bottlenecks, not the annoyances

A step is a bottleneck if at least one of these is true:

• It happens many times a week and takes the same shape each time.
• Work stops at it while it waits for one specific person.
• Things get lost at it — enquiries, details, follow-ups.
• Its quality varies depending on who does it or how busy the day is.

Something that is merely irritating but happens twice a month is not a bottleneck. Be ruthless here; this is the step where most self-run audits go wrong.

Step 3 — Score each candidate on impact, effort and risk

Three scores, one to five, no more precision than that.

• Impact: what it is worth if this is fixed — in recovered enquiries, recovered hours, or errors avoided. Put a number on it, even a rough one.
• Effort: what it would take to change, including the changes to how people work, not just the build.
• Risk: what happens if it goes wrong in front of a customer. A wrong answer to a pricing question is recoverable. A wrong answer about safety is not.

Order by impact divided by effort, then strike out anything with a risk of four or five that you cannot supervise. What remains is your shortlist.

Step 4 — Apply the four tests before AI goes anywhere near it

A candidate should pass all four. If it fails one, it is not an AI problem yet.

• Is the process stable? If it changes every month, you will be rebuilding constantly. Fix the process first.
• Does the information exist? If the knowledge lives only in one person's head, the first job is writing it down — which is valuable on its own and often enough.
• Is a wrong answer survivable? If not, the system must escalate rather than answer, and you should design for that from the start.
• Is there a simpler fix? A form, a phone setting, a shared calendar, one changed handover. If a simpler fix would do most of the work, do that first and re-measure.

That last test kills more candidates than the other three combined, and it should. The cheapest AI project is the one you correctly decided not to run.

Step 5 — Write a 30-day plan, not a roadmap

Take the top item only. Write what you will change, what you expect to happen, and the one number you will look at in 30 days to judge it. Set the number now — deciding afterwards whether it worked is how projects get declared successes.

A useful plan fits on one page and has one owner. If it needs a programme, a steering group and a phased rollout, you have picked something too large for a first move.

What this method will not tell you

It will not tell you whether a specific integration is technically possible with your software — that needs a feasibility check against the actual systems, and the honest answer is often no. It also will not tell you whether a supplier is any good. For that, ask them to show you the thing working before you pay for it.

If you would rather run this with someone who has done it before, that is exactly what our AI Opportunity Audit is — and if the answer is that a simpler fix will do, that is what the plan will say.