What is an AI agent, really?

The phrase "AI agent" has been used so loosely over the last two years that it has almost stopped meaning anything. Vendors apply it to everything from a chatbot with three canned responses to genuinely sophisticated software. If you are trying to decide whether your business needs one, that vagueness is a real problem.

So let us be precise.

The simplest useful definition

An AI agent is software that is given a goal rather than a set of instructions, and that can take actions in the real world to achieve it.

That second half is the part that matters. A language model on its own can write you an email. An agent can read the enquiry, check your calendar, decide which slot to offer, book it, send the confirmation and update your customer record. The model does the thinking. The agent does the doing.

How it differs from a chatbot

A traditional chatbot follows a decision tree somebody drew. If the visitor says X, reply Y. It cannot handle anything outside the tree, which is why talking to one is so frequently infuriating.

An agent works differently. It understands the request, works out what steps are needed, carries them out, checks whether the goal has been met, and adjusts if it has not. Nobody drew a tree. The behaviour emerges from the goal plus the tools it has been given.

The practical difference for a customer is enormous. A chatbot handles the three scenarios someone anticipated. An agent handles the long tail of things nobody thought to script.

The three ingredients

Every working agent has the same three parts.

A model. The language model that does the reasoning. This is the part everyone talks about, and frankly the least interesting decision to make. The models available now are all capable enough for the vast majority of business tasks.

Tools. The specific things the agent is permitted to do: look up a booking, send an email, query stock levels, raise an invoice. An agent with no tools is just a chat window. The tools are where the value lives.

Boundaries. The rules about what it must not do, when it must ask a human, and what information it may access. This is the part cheap implementations skip, and it is the part that determines whether the thing is safe to put in front of customers.

What they are genuinely good at

Agents excel at work that is repetitive but not identical. Handling an enquiry, processing a document, triaging a request, scheduling something around constraints. Tasks where the shape is consistent but the detail varies every time, and where a rigid rules-based system would need a hundred special cases.

They are also very good at being available. An agent does not have a bad morning, does not go on holiday, and answers the 200th call of the day exactly as well as the first.

Where they fall short

They are not good at tasks requiring genuine judgement about people. They should not be making decisions about someone's employment, their credit, or their care.

They are not reliable at arithmetic unless you give them a calculator, which is a slightly absurd but very real limitation.

And they will occasionally state something wrong with complete confidence. Good implementations mitigate this by grounding every answer in your actual documents and data, and by designing the agent to say "I do not know" rather than guess. But it never quite goes to zero, which is why a human handover route is not optional.

Do you need one?

A reasonable test: is there a task in your business that happens more than twenty times a week, follows a recognisable pattern, and currently requires a person to be available? If yes, an agent is probably worth costing. If the task happens twice a month, automate something else first.

The businesses getting the most out of agents right now are not the ones with the most ambitious plans. They are the ones who picked one irritating, high-volume, low-risk task and did it properly.