Practical AI for businesses that are not AI companies

The useful question is not "how do we use AI?" but "which hour of our week is spent turning one document into another?" That is where current AI is genuinely strong and where the risk is easy to contain.

Good first projects

Document extraction: invoices, delivery notes, application forms and receipts turned into structured records a person confirms rather than types.

Support drafting: a first reply written from your own past answers and policy documents, reviewed by staff before it is sent.

Internal search: staff asking plain questions of contracts, manuals and price lists, with the source paragraph shown next to every answer.

Summarising: long meeting notes, reports and email threads reduced to decisions and actions.

Projects to avoid at the start

Anything that makes an irreversible decision without a human — refunds, pricing changes, hiring screens, medical or legal judgements. Anything that requires clean historical data you do not actually have.

The three rules we work by

Keep a human in the loop wherever the output leaves the building. Review is cheap; a wrong invoice sent to a client is not.

Always show the source. An answer without a citation cannot be checked, and an unverifiable answer is worse than none.

Measure before and after. If the task took forty minutes and now takes eight, you have a result you can defend. If nothing was measured, the project becomes an opinion.

What it costs

A focused first automation is usually weeks, not months, and it runs on existing cloud services rather than your own hardware. The main cost is the discovery work: writing down precisely how the task is done today, including the exceptions nobody documented.