AI agents for a twenty-person firm: what's real

What holds up

Drafting and summarising. Answering questions over your own documents ("what did we quote them last time?"). Sorting and routing enquiries. Reading invoices and delivery notes into your system with a person glancing over the result. Boring jobs, real hours saved, and the failure mode (a duff draft, a mis-read line) gets caught because a human is still in the loop.

What doesn't, yet

Anything that must be right with nobody checking. Pricing. Payroll. Sending anything to a customer unreviewed. Agents fail differently from people: rarely, but confidently, with no sense of having done so. The rule of thumb that holds up: give an agent the work you'd give a bright temp on their first morning, with the same checking, and no work you wouldn't.

What to pay

Much of the value is already in tools you may be paying for, plus an afternoon of setup and a written procedure. Some of it is a small built piece around your own systems. Be suspicious of five-figure "agent transformation" quotes that can't name the task: if a supplier can't tell you which job the agent does, in numbers (minutes saved, error rate, who checks it), you're being sold a demo. Our straight answers about AI page stays current on what we'd actually use; for whether any of it fits your firm, that's an advice conversation, and "none of it yet" is a common and respectable answer.