How to Choose Your First AI Project in a Small Business

The best first AI project in a small business is a task that happens often, follows rules you could write down, and does little damage when it goes wrong. It is rarely the most exciting idea on the list.

Owners often arrive with a long list. Some items are sensible, such as sorting the shared inbox. Others are ambitious, such as an assistant that runs sales. The temptation is to start with the ambitious one because it promises the biggest return. In my experience that is the quickest way to end up with nothing to show after three months.

Why does the first project matter so much?

A first project sets the tone for everything after it. If it works, staff trust the second one and ask for a third. If it stalls, "we tried AI and it didn't work" becomes the accepted story, and it can take a year to shift.

It also teaches you things no planning document will. You learn how your data really looks, how your team reacts to a system that makes suggestions, and how much checking it needs. Those lessons are cheap on a small project and expensive on a large one.

What makes a good first project?

I score each candidate on four questions.

1. How often does it happen? A task that happens daily or weekly pays back quickly and gives the system plenty of examples to prove itself on. A task that happens twice a year will not.
2. Could you write the rules down? If a capable new starter could do it from a one-page instruction sheet, AI can probably help. If it depends on years of instinct and context, leave it for later.
3. What happens if it is wrong? Sorting an inbox is forgiving, because a misfiled email gets noticed and moved. Sending a legal notice or releasing a payment is not. Start where a mistake is cheap and visible.
4. Is the input digital and reasonably clean? Typed emails, spreadsheets and PDFs are workable. Handwritten notes in a filing cabinet are a different project, and the scanning has to come first.

Give each question a score from one to five and add them up. The scores do not need to be precise. They just stop the loudest idea in the room winning by default.

What does that look like in practice?

Here is an invented example, for illustration. A twenty-person building services firm lists five ideas: an AI estimator, a tool that sorts the shared enquiries inbox, an assistant that answers staff HR questions, automatic invoice entry, and a system that drafts tender responses.

The estimator scores badly. The rules are tied up in the senior estimator's head, and a wrong estimate can cost real money. The tender responses score badly on rules and risk. The enquiries inbox scores well: it happens every day, the categories are easy to list, and a misrouted email costs a few minutes. Invoice entry also scores well, with one caveat. Payment must stay a human step.

So the firm starts with the inbox. Three weeks in, they have real numbers and a team that has seen the system make a few silly mistakes and a lot of useful ones. Invoice entry is next, and the estimator may never be the right idea at all.

Tasks that usually score well as a first project:

• sorting and routing a shared inbox
• pulling figures from supplier invoices into a spreadsheet
• chasing quotes or suppliers that have gone quiet
• answering the same internal policy questions again and again

Tasks that usually score badly:

• final decisions about people, such as hiring or dismissal
• anything that sends money without a human check
• any process nobody in the business can describe step by step

Should a person stay in the loop?

Yes, at the start. Let the system draft, sort or extract, and have someone approve the output. After a few weeks you will know where it is reliable and where it is not. You can then loosen the checks where the results support it and keep them where they do not. Take a human check out only once a few weeks of results show it is not catching anything.

What should you avoid?

Three things come up repeatedly.

• Picking the project to impress. A first project chosen for the board or for a competitor's benefit tends to be too big.
• Skipping the owner. Every project needs one person who understands the current process and wants it improved. Without them, nobody notices when the system quietly drifts.
• No baseline. If you do not time the current process before you start, you cannot say afterwards whether anything improved.

What should you do this week?

Write down five tasks that eat time in your business. Score each one out of twenty on the four questions above. Pick the highest score, not the most interesting one. If two tie, choose the one whose owner is keen to help, because a willing owner matters more than a clever model.

Then time that task twenty times. Those twenty timings are your baseline, and they are the most useful thing you will produce before any AI is involved.