Most of the AI advice aimed at business owners is either breathless or vague. Here is what I have seen actually pay off in businesses of roughly 20 to 150 people, and a simple rule for where to stop.

Where it reliably pays off

  • Drafting from structured information. A brief from a client request, a quote from a price list, a status update from a project record. The AI drafts, a person checks and sends. This is the single biggest time saver I see.
  • Reading documents. Pulling the fields out of invoices, purchase orders, shipping documents and application forms. It is fast and, with a check step, accurate enough.
  • Triage. Sorting incoming enquiries, support requests or applications into categories and routing them to the right person.
  • First-line answers. A chatbot that answers from your own documents and hands over to a person when it is unsure. The handover is the important part.
  • Summarising. Long email threads, meeting recordings and reports into something a busy person will actually read.

Where it usually does not

  • Replacing a process that was never defined. AI will cheerfully automate the confusion.
  • Tasks that are already quick for a person and happen rarely.
  • Anything where the cost of an occasional wrong answer is high and nobody is checking.

The rule: AI prepares, a rule decides, a person approves

For any consequential step, I design it in three parts. The AI prepares the work: the match, the draft, the classification. A fixed rule, written in plain code outside the AI, decides whether a person needs to look: amount over a limit, confidence below a threshold, a new customer, an exception of any kind. And a named person approves anything the rule sends to them.

The rule is the piece most people skip. It is what stops the AI, or someone talking to it, from approving something it should not. I built a working accounts-payable prototype to show this: the AI matches invoices to purchase orders, but it cannot approve a payment above its delegated limit no matter what it is told.

How to start

Map your core processes first. Mark the decision points. Then pick one task from the first list above that sits between two of those decision points, automate that, and measure it. One working, measured automation does more for adoption than any amount of strategy.