Where AI agents genuinely pay off in a small business

Most small businesses are being sold agents as a way to replace people. That is the wrong frame and it produces disappointing projects. The useful question is narrower: which specific repeated task, with a clear boundary and a tolerable failure, currently eats hours that nobody enjoys?

The shape of work that suits an agent

Three properties, and the work needs all three rather than any one.

High volume. Something happening tens of times a day. Automating a weekly task saves a few minutes and costs you the attention of maintaining it, which is a bad trade.

Low stakes per instance. A wrong answer should be an inconvenience somebody notices and corrects, not a loss you cannot reverse. That rules out anything moving money or making commitments to customers on its own.

A clear boundary. The agent can see exactly the data it needs and nothing else, and there is an obvious point where it stops and a person takes over.

Three that tend to work

First-line enquiry handling on WhatsApp. Opening hours, whether something is in stock, where you are, how delivery works. High volume, low stakes, and the answers come from data you already maintain. The agent handles the repetitive majority and hands anything unusual to a person.

Triage and routing. Not answering, just sorting. Deciding which enquiries are urgent, which are sales, which are support, and putting each in front of the right person with a one-line summary. Being wrong costs a redirect.

Daily reporting in plain language. Turning yesterday's numbers into a short readable summary that flags what moved. The underlying figures come from your systems and remain checkable, so the agent is doing the writing rather than the counting.

Three that tend not to

Anything that commits you. Quoting a price, agreeing a delivery date, approving a refund. Being right most of the time is not good enough when the exceptions are binding.

Low-volume, high-context judgement. The decisions that make a business work are usually rare, heavily contextual and dependent on things nobody wrote down. Poor automation candidates almost by definition.

Anything needing accountability. If a regulator, an auditor or a customer may later ask who decided this and why, the answer needs to be a person.

The handover is the product

The most common failure is not a wrong answer. It is an agent with no exit, holding a confused customer in a loop while a person who could have solved it in ten seconds never finds out.

Design the handover before the conversation. Decide what triggers it, where the conversation lands, and how much context arrives with it. An agent that escalates early and cleanly is worth more than one that resolves a higher share of cases and traps the rest.

What it costs to run

Per-message costs are usually the smallest line. The real ongoing costs are keeping the agent's information current, reviewing what it actually said, and owning the failures. That last one is not optional: an agent speaking in your name is your responsibility, and somebody has to read a sample of the transcripts.

If nobody in the business has time to review the output, that is a strong signal not to deploy one yet.

How to try one without regretting it

Pick a single task from the first list. Give the agent read-only access to what it needs and nothing more. Run it alongside your existing process rather than replacing it, so you can compare. Read the transcripts weekly for a month. Then decide whether it earned its place, and be willing to conclude that it did not.

That is a slower path than the demonstrations suggest, and it is the one that leaves you with something you still want in six months.