Where should a small business start with AI?

Ask each person who does the work, not only the owner, to describe the last real job they handled from first contact to final invoice. Write down every step, who did it and roughly how long it took. You are not looking for ideas yet. You are building an honest picture of a normal week.

Look for four patterns: work that repeats, work that is easy to get wrong, work that waits for a busy person, and work that depends on information stored somewhere awkward. Those are the places where automation pays back fastest.

You do not need precision. For each task, estimate the hours it takes per week, multiply by what those hours are worth, and note what it costs you when it goes wrong: a missed enquiry, a late invoice, a double booking. A task that takes two hours a week and never goes wrong is a different proposition from one that takes two hours and loses you a customer each month.

Writing the numbers down matters because it stops you automating what is annoying rather than what is expensive.

Score each item on two questions: how much would fixing it save or earn, and how hard would it be to fix. The best first project is high on the first and low on the second. A rough sort into “do first”, “do later” and “leave alone” is enough.

Be willing to leave things alone. Some tasks are annoying but rare, and the cost of automating them will never be repaid.

Not every problem needs AI. A form that creates a record, a rule that sends a reminder, or two tools connected so one updates the other are cheaper, more predictable and easier to maintain, and they solve a surprising share of small-business admin.

AI earns its place when the work involves language or judgement: reading a free-text enquiry and working out what the person wants, drafting a reply, sorting a messy inbox, or pulling the facts out of a document. A sensible order is to automate the obvious first, then apply AI to the judgement-heavy middle.

Pick the top item and try it on a narrow slice: one service, one team, one inbox. Agree in advance how you will know it worked, such as time saved per week, enquiries answered within an hour or errors avoided, and a date to review it.

A pilot that fits inside a few weeks is easy to judge and cheap to stop. If it works, widen it. If it does not, you have learned something for a small price.

Buying a tool before you have named the problem. Automating a process that was already broken, which only makes the mess faster. Giving software access to more than it needs. Letting an AI act on something costly or public without a person approving it. And measuring success by how impressive the demo looked rather than by what changed in the week.

What exactly will it do, and what will it not do? What does it have access to, and can I limit that? What needs a person to approve it? How will I see what it did and why? How do we start small? What does it cost to try, and what does it cost to stop? A good answer to each is more reassuring than any feature list.