What AI can and cannot do for a small business right now

Every owner we meet has been told two things about AI in the same month. That it will run their company for them, and that it is a toy that makes things up. Both are wrong in useful ways.

What it does well today

AI is good at the work you already know how to do but do not have time for. Reading a bank statement and matching each line to an invoice. Drafting the third polite reminder to a customer who has not paid. Turning a messy enquiry into a tidy record with the right fields filled in. Noticing that a supplier’s price on this invoice is not the price on the purchase order.

The pattern is the same in each case. The rules are clear, the volume is high, the judgement is small, and the cost of a human doing it is boredom and error. Handed to a trained agent with the right access to your systems, this work gets done every day, on time, and consistently.

What it does badly, or should not do at all

It should not decide anything that moves money or reaches a customer on its own. Not because it always gets it wrong, but because when it is wrong the cost lands on you, and you were not in the room. The right design is confirm-first: the agent prepares, you approve, it acts.

It is also poor at work with no ground truth. If nobody in your company can say what the correct answer is, an agent cannot be trained to produce it. Get the process right with people first, then automate it.

How to tell the difference

Ask three questions about any task. Could a careful new hire do it after a week with a written procedure? Does it happen often enough to matter? Would a mistake be caught before it hurt anyone? Three yeses and it is a candidate. Fewer, and it is either a people problem or a decision, and both deserve a person.

Where to start

Start with the books. Reconciliation, follow-ups on unpaid invoices, and the monthly close are the most common, most rule-bound, and most hated jobs in almost every business we have seen. They also produce the fastest, most visible saving, which is what earns the trust you need to automate the next thing.