What is an AI agent,and when may it decide on its own?

A simple definition

A chatbot answers a question. An AI agent works towards a goal. It looks at the situation, decides what step to take, takes it, checks the result and decides what comes next. Sending a quote request to three suppliers, reading their answers, comparing prices and putting the best option in front of you is agent work.

Under the hood an agent is a language model with access to tools: it can read your mailbox, look up a client in the ERP, draft a document or schedule a reminder. What makes it an agent is that it chains those steps itself, instead of a person telling it every move.

What agents do in an SME

In practice agents shine at the work nobody likes but everybody needs: following up. A file agent that watches certificates and deadlines and asks the right person in time. A quote agent that chases suppliers who have not answered. A mail agent that reads incoming requests, links them to the right project and drafts a reply.

Several agents can work together. One reads the mail, another updates the file, a third proposes the next step. Each has a narrow job, which makes them easier to trust and to check.

Autonomy is a dial

The question is not whether an agent may decide on its own, but which decisions. Think of a dial per type of action. At one end the agent only suggests and a person decides. In the middle the agent acts, but asks approval first. At the other end it acts on its own and reports what it did.

Reading and linking a mail can sit far towards autonomous: the risk is low and mistakes are easy to fix. Sending a reminder to a supplier can sit in the middle. Changing a price or placing an order should stay with a person, at least at first.

How to decide

Four questions help. What does a mistake cost? A wrong tag is cheap, a wrong price is not. Can it be undone? A draft can be deleted, a sent mail cannot. How often is the agent right today? If people approve nine out of ten suggestions without changes, that action is a candidate for more autonomy. And who is accountable? Someone in your company must own each type of decision, even when an agent executes it.

We agree these levels with you upfront, per action, and adjust them as the track record grows. That is a business decision, not a technical one.

Safeguards that matter

Every action an agent takes should be logged with the reason and the source it used. Actions should be reversible wherever possible. Agents should only get the access rights they need for their job. And there should be limits: a maximum amount, a list of approved suppliers, working hours for sending mails.

With those in place, an agent is not a black box. It is a colleague whose every step you can follow.

Agents that learn

A good agent gets better from the corrections people make. If someone keeps changing the urgency of a certain type of request, the system can propose that as a rule. A person approves the rule, and from then on every agent applies it. That way the system learns your way of working, without anyone writing code.

Agents versus classic automation

Classic automation follows fixed rules: when an invoice arrives with this supplier number, book it on this account. It is fast and predictable, but it breaks as soon as something does not match the rule. An agent can handle variation: a mail that asks two things at once, a document in an unexpected format, a client who refers to a project by a nickname.

The best systems combine both. Rules where the work is predictable, agents where it is not. An agent does not replace good software, it handles the part that software alone could never handle.

Common mistakes

Giving one agent too many jobs, so nobody can tell why it did what it did. Giving it broad access rights because that is easier to set up. Skipping the approval phase because the demo looked good. And not measuring: without knowing how often the agent is right, you cannot decide when it may do more. Each of these is easy to avoid if you plan for it from the start.

What agents cost to run

Running an agent costs model usage per step, plus the software around it. For typical SME volumes, the model cost is small: a few cents for each mail read or reminder drafted. The real investment is in building the agent into your process, with the right access, rules and approvals. Once that is in place, adding a second agent that uses the same connections is much cheaper than the first.

Measure the value in hours of follow up that no longer depend on someone remembering, and in requests that no longer wait for days.

Where to start

Pick a flow with a lot of follow up and low risk per action. Let the agent suggest for a few weeks and measure how often people agree. Then widen its autonomy step by step. You end up with agents your team trusts, because they earned it.