Agents · Briefing

What an AI agent can and cannot run in a small business

An AI agent is software that is given a goal, some tools and permission to act, then left to get on with it. The promise is real. So are the limits, and the limits matter more than the promise when it is your business.

What an agent actually is

A chatbot answers when you ask. An agent does a sequence of things on its own: it reads an email, looks up the customer, checks a price, drafts a reply and may send it. The difference is not intelligence. It is permission. An agent can touch your systems, which is what makes it useful and what makes it risky.

Where agents do well

Agents are good at work that repeats, has clear inputs and can be checked quickly.

• Sorting incoming enquiries and routing them to the right person.
• Drafting replies from your own past answers.
• Matching records between two systems and flagging the ones that do not agree.
• Preparing a first pass of a quote from a standard price list.
• Chasing the paperwork that is missing from a file.

What these have in common is that a mistake is small, visible and easy to correct.

Where they struggle

They struggle where the rules are unwritten, where a mistake is expensive or hard to undo, and where the right answer depends on knowing the customer. A long standing client who is always late but always pays is not in the data. A supplier who needs a phone call, not an email, is not in the data either.

They also struggle quietly. An agent that is wrong rarely announces it. It produces something that looks finished.

The two tests

The new employee test. If a sensible new starter could do the job from a one page instruction, an agent probably can. If it takes a year of feel, it cannot yet.

The undo test. Ask what happens if it gets one wrong. If the answer is "we fix it tomorrow", fine. If the answer is "we lose a customer" or "we cannot take it back", keep a person in the loop for that step.

Give an agent the repetitive work. Keep the judgment.

What good supervision looks like

• One named person owns the agent's work, the same way they would own a junior's.
• Every action is logged in plain language, so you can see what it did and why.
• Anything that spends money, contacts a customer for the first time or deletes something waits for a person to approve it.
• There is an obvious way to switch it off.

If a supplier cannot show you these four things, they have sold you a risk, not a tool.

How to start

Start with one process, one owner and a way to see what the agent did. Run it alongside the old way for a few weeks. Widen only when it has earned it. Most of the value comes from the first small job, and most of the damage comes from giving a new system too much too soon.

Questions worth asking anyone who sells you one

• What can it do without asking, and what does it always ask about?
• Where can I see what it did yesterday?
• Who is responsible when it is wrong?
• If we stop working with you, does it keep running, and who owns it?

The honest answers are usually short. Be wary of long ones.