AI agents for small businesses: what they are, what they can do and what's hype

Qualify and log leads: when an enquiry comes in through the website form or WhatsApp, the agent replies, asks two or three qualifying questions (what they need, by when, company size), logs it in the CRM with a summary and alerts a salesperson only if the contact is a good fit. The filter nobody has time to do by hand.

Manage appointments end to end: offer slots based on the real calendar, book, send reminders and move the appointment when the customer asks to change it, without anyone touching the calendar.

Process documents: read invoices, delivery notes or orders that arrive by email in twenty different formats, extract the data and enter it into your system, flagging the ones it doesn't understand for human review.

Prepare summaries and follow-ups: every morning, a summary of new leads, conversations waiting for a reply and quotes unanswered for a week, with a follow-up email drafted and ready to approve.

Irreversible decisions without supervision: sending quotes with prices the agent calculates on its own, making payments, deleting data or committing to delivery dates. The technology can do it; the problem is that when it fails, it fails confidently, and a pricing mistake sent to a hundred customers can't be undone.

Sensitive conversations: a serious complaint, a negotiation, a genuinely angry customer. A good system detects these cases and hands them to a person straight away, with the context already gathered.

The rule we use: the agent executes, a person approves anything with external consequences, until months of logs show that a specific step can be let go. Switching everything to automatic on day one is the recipe for the kind of disaster that ends up in the news.

Three questions expose it quickly. First: 'which of my systems does it connect to, and what specific actions does it take?' If the answer is vague, it's a chatbot with a new name. Second: 'what happens when it isn't sure?' A serious agent has thresholds and escalates to a person; hype 'never fails'. Third: 'can I see a log of every action it took?' Without traceability there can be no trust.

And be wary of results that are too neat ('saves 80% of your team's time'). Real savings exist and they can be large, but they're measured case by case, by counting hours before and after.

Don't start by buying 'an agent': start by choosing a process. The best candidate is repetitive, frequent, rule-based and cheap to correct when it goes wrong. Appointment handling and lead qualification meet all four conditions, which is why they're usually the first project.

Set it up as a pilot for a few weeks, with human approval on every external action, and measure: hours saved, reply speed, appointments that used to be lost. With those numbers you decide whether to expand. If you'd like us to look at which process in your business is the best candidate, the digital audit is free.