Where artificial intelligence actually earns its place in a smaller business

Most owners I speak to have been sold artificial intelligence at least twice before anybody explained what it would actually do for them. The pitch usually arrives with a demo, a chart, and a number that sounds impressive without saying much. Then it stops, and you are left holding a decision you have no way to evaluate.

So here is a way to evaluate it that does not require you to understand how any of it works underneath.

The three questions that decide it

Point AI at a task and ask three things about that task. How often does it happen? How much does the answer vary? And could a competent new hire learn to do it from a written page of instructions?

Work that happens hundreds of times a week, where the answer follows a pattern, and where the rules could be written down, is where these tools do their best work. Answering the same twelve customer questions. Pulling a total off an invoice. Sorting an inbox into categories. Reading a hundred CVs and putting the plausible ones at the top.

Work that happens twice a month, where every case is genuinely different, and where the right answer depends on knowing your business well, is where the same tools disappoint. Negotiating with a supplier who has let you down. Deciding whether to take on a difficult client. Working out why the numbers feel wrong even though they add up.

The frustrating part is that the second kind of work is the interesting kind, so it is the kind people want to automate. It is also the kind that will waste your money.

What this looks like in practice

A distributor with four people in customer service found that just under seventy per cent of incoming messages were one of five questions. Where is my order, can I change the delivery date, do you have this in stock, what is your account minimum, and can I have a copy of the invoice. Every one of those has an answer that already exists in a system somewhere.

That is a good candidate. Not because it is clever, but because it is dull, constant, and answerable from data the business already holds.

Compare that with a request to build something that predicts which customers are about to leave. It sounds far more valuable. It probably is, eventually. But it needs several years of clean history, a working definition of what leaving even means for your business, and somebody to act on the prediction when it arrives. Most businesses have none of those three.

Your data is the constraint, not the technology

Anything that predicts, forecasts or spots the unusual has to learn from your history. Not a general model of businesses like yours. Yours.

Which means before anybody quotes you for prediction work, somebody should look at what you actually have. Three questions again.

• How far back does it go? Two years of records is usually the floor for anything seasonal, because you need to see the same December twice.
• Is it consistent? If your team changed how they logged job types in 2024, you effectively have two shorter datasets rather than one long one.
• Is it in one place, or spread across a system, three spreadsheets and somebody’s email?

Plenty of projects that get sold as artificial intelligence turn out to be data cleanup with a model bolted on at the end. That is not a scandal. It is just worth knowing which one you are buying, because the cleanup is where the time goes.

The integration matters more than the model

This is the part that gets least attention and causes most of the disappointment.

A tool that answers customer questions brilliantly, in a window nobody on your team has open, will not get used. A model that flags at risk accounts perfectly, in a weekly PDF that lands in an inbox on a Friday afternoon, will not change anything. The value shows up when the output arrives inside the thing people already have open all day. The CRM. The helpdesk. The messaging app the warehouse actually uses.

When you are comparing proposals, ask where the output lands and who sees it. If the answer is a separate dashboard with its own login, adjust your expectations downward.

Things that go wrong, and are avoidable

Confident wrong answers are the obvious one. Language models produce fluent text whether or not they know the answer, and fluent text reads as authoritative. Anywhere the cost of being wrong is high, the tool needs to hand over to a person rather than guess. That handover rule should be written into the brief, not discovered later.

Ongoing cost is the quieter one. These systems charge by usage, so a tool that costs very little during a pilot with three people can cost considerably more once the whole company is using it. Ask for the cost at full volume, not the cost at pilot volume.

And then there is drift. Your products change, your policies change, your prices change. Anything trained on how things were will slowly become a system that confidently describes a business you no longer run. Someone has to own keeping it current. If nobody is named, that job will not happen.

A reasonable first step

Pick one process. Not a department, not a strategy. One process that happens constantly and annoys everybody.

Measure it for two weeks before you change anything. How many times did it happen, how long did each one take, how often did somebody have to redo it. You will need that baseline later, because without it every claim about improvement afterwards is a matter of opinion.

Then change that one thing, and leave everything else alone until you can point at what it did. It is a slower way to start and a much faster way to end up with something that works.

The businesses getting real value out of this right now are rarely the ones with the most ambitious plans. They are the ones who picked something boring, measured it honestly, and let it run.