Small businesses get a strange version of the AI conversation. The case studies are about enterprises with data teams, and the tools are sold as though every company has thousands of conversations a day. Neither describes a firm of twelve people with a shared inbox.
So, concretely: here is what works at that size, and what doesn't.
Four things that genuinely work
Answering the same questions, at any hour. Most businesses have twenty questions that make up the bulk of enquiries — opening hours, delivery areas, lead times, whether you do X. An agent grounded in your own material handles those immediately, including at eleven at night when the enquiry would otherwise sit until morning. Enquiries that get an instant answer convert better than ones that wait, and for a small team the overnight gap is where most of the loss happens.
Reading documents so nobody has to type them. Invoices, delivery notes, application forms, CVs. If a person is reading a PDF and typing its contents into another screen, that is work an agent does accurately and quickly. It is unglamorous and it is usually the highest-value thing we build for a smaller company.
Making internal knowledge findable. Everything the business knows is in old emails, a shared drive, and one person's head. An internal agent pointed at that material answers questions from it — which matters most when the person who knows everything is on leave.
Qualifying enquiries before they reach you. Rather than a form, a short conversation that establishes what the person needs, their rough budget and timescale, then routes it. You get a sorted enquiry with context instead of three lines and a phone number.
Three things that don't
Making judgement calls. Pricing exceptions, credit decisions, anything with legal weight. Not because the model can't produce an answer — because it will produce one confidently when it shouldn't, and the cost of a wrong one lands on you. These escalate to a person by design.
Working with data you don't have. An agent cannot answer stock questions if stock lives in a spreadsheet that is updated weekly. Frequently the actual project is the system underneath, and AI on top of bad data just produces wrong answers faster.
Replacing a team. For a small business the realistic return is hours returned and faster responses, not headcount removed. Anyone promising otherwise is selling something.
Where to start
Pick the single question or task that happens most often, and do only that one. Prove it works, measure what it saved, then widen.
The opposite approach — a general assistant that does everything — is how most small-business AI projects fail. It is impossible to test, so nobody trusts it, so nobody uses it.
That first narrow agent is usually live in a few weeks. See AI agents and AI solutions for how we scope one, or get in touch with the question your team answers most often.
