AI for a two-person business: where would you actually start?

At half past eight, the work is finished but the replies are not. There are three enquiries to answer, a quote to write up and a customer asking what happens next. If you run a small business, that is a more useful place to start with AI than a list of features.

Pick one piece of that admin. Try using AI to prepare it. Keep the decision and the send button with you until you know what the tool gets right and where it needs correcting.

Choose a task you can check quickly

A first task should happen often enough to be worth improving, with an answer you can recognise. Drafting a reply from a few notes is a reasonable candidate. Deciding whether a job is safe to undertake is a poor first experiment.

For a two-person business, possible starting points include turning rough notes into a customer update, sorting enquiries by the service requested, or extracting a list of questions from a long email. Choose whichever currently interrupts your day most.

You do not need a custom application for the first trial. An existing tool may be enough to find out whether the task suits AI. Spend the effort describing what a good result looks like before considering a bigger build.

Give it the facts and the limits

Suppose you fit kitchens. A customer wants to move an installation date, and you have two possible slots. A useful instruction gives the tool those dates and the tone of the reply. It also tells it not to promise anything else.

Try this with a made-up enquiry first:

Draft a short reply offering 14 or 21 October for the visit. Ask which works for the customer. Do not invent a time, a price or a reason for the change. Leave anything missing as a question for me.

The dates are illustrative. The useful part is the boundary: the assistant has enough information to write, and no permission to fill gaps with guesses.

The NCSC’s guidance on AI and cyber security explains that AI can present incorrect information convincingly. Check the facts even when the writing sounds confident. Before using real customer information, understand the product’s data settings, access and retention terms. A trial with anonymised or invented details avoids needing that information just to test the idea.

Count the corrections

Try a small batch of tasks you would normally handle yourself. Note how long the original work takes, how long the AI-assisted version takes including review, and what you had to fix.

If drafting takes a minute but checking takes ten, you have not solved the problem. Repeated errors may mean the instructions need work, the source information is incomplete, or the task is a bad fit.

Keep the unsuccessful examples. They are more useful when deciding what to build than the one reply that came out perfectly.

Connect it when copying becomes the problem

If the trial works but copying notes between tools becomes a nuisance, an integration may make sense. It could pull the right customer record into a draft and save the result for review. That is the point to discuss permissions, approval steps and what happens when the connection fails.

Our guide to custom AI versus a better process helps with that decision. If quotes are taking up the evening, start with the quoting workflow instead.

A useful first result is modest: one recurring job that takes less effort, with a clear way to check it. You can decide what comes next after that works.