What is an AI agent? A plain answer, minus the sales pitch

An AI agent is a chatbot that has been given permission to act. That is the whole difference. A chatbot reads your question and writes an answer. An agent reads your question, decides what steps would answer it, and then goes and does them: opening pages, running searches, filling forms, sending things, checking whether the result worked and trying again if it did not.

Everything else people say about agents is downstream of that one change. It is also why the honest version of this article has to spend as much time on what goes wrong as on what goes right, because handing software the ability to act is a different kind of decision from handing it a question.

The distinction that actually matters

Ask ChatGPT to book you a table and it writes you a very good description of how to book a table. Ask an agent and it opens the booking site.

That gap is not about intelligence. The same model can be behind both. What changes is the scaffolding around it: a list of tools it is allowed to use, a loop that lets it take more than one turn, and somewhere to keep track of what it has already tried.

Four things have to be true before something is meaningfully an agent:

1. It can take actions, not just produce text. Usually by calling tools: a browser, a search index, an API, your calendar.
2. It runs more than one turn. It can look at what happened and decide what to do next, rather than producing one answer and stopping.
3. It holds a goal, not an instruction. "Find me three suppliers who ship to Ireland" rather than "write an email to this supplier".
4. It decides the steps itself. You said what you wanted, not how to get it. This is the part that makes agents useful and the part that makes them unpredictable, and you do not get one without the other.

Take any one of those away and you have a chatbot with extra features, which is what a lot of things currently sold as agents actually are.

The loop, in four steps

Under the marketing, almost every agent runs the same cycle.

It looks at the current state: your request, what it has done so far, what came back from the last thing it tried. It plans a next step, which for a good agent means one step rather than a twelve-point programme it will never finish. It acts, by calling a tool. Then it checks what came back and starts again, until the goal is met or it gives up.

The checking step is where agents mostly succeed or fail. An agent that reads the result properly notices that the page returned a login screen instead of a product listing, and adapts. An agent that does not notice carries on for another nine steps building a confident answer on top of a login screen.

Agentic AI vs generative AI

These get used as if they were rival technologies. They are not. Generative AI is the underlying capability: a model that produces text, images, code. Agentic AI is a way of using that capability, where the model is wired into a loop and given tools.

Every agent is built on generative AI. Not every use of generative AI is an agent. When a vendor's page draws a hard line between the two, it is usually because they sell the second one.

There is a real distinction hiding in there, and it is about risk rather than architecture. Generative AI produces something you look at before it goes anywhere. Agentic AI does things whether or not you looked. That is the line worth drawing, and it is the one to think about before you connect anything to a system that can send email.

What agents actually do well right now

Being specific here matters more than being enthusiastic, so: three shapes of task where agents genuinely earn their keep today.

Research that needs many sources. Give an agent a question that needs twenty pages read and cross-referenced, and it will do a creditable job in a few minutes. This is the clearest current win, because reading is cheap, mistakes are visible, and nothing irreversible happens.

Repetitive work across a set of things. Take these forty invoices, pull out the supplier, the date and the total, put them in a table, flag anything that looks like a duplicate. Boring, well-defined, easy to check.

Code that comes with its own test. Coding agents are further ahead than any other category, for one structural reason: the code either compiles and passes the tests or it does not. The agent gets an honest signal after every step, which is exactly what the checking step needs. Almost no other task provides that.

Notice the pattern. Agents work best where feedback is immediate and unambiguous. They work worst where the result only looks wrong three weeks later.

Where they still fall over

They do not know when they are stuck. A person who has tried the same thing four times concludes the approach is wrong. An agent frequently tries it a fifth time. The most common failure is not a dramatic mistake, it is a quiet loop that consumes your budget and produces nothing.

Small error rates compound. An agent that is 95% reliable per step is about 60% reliable across ten steps, and long tasks have far more than ten. This is why demos are so much better than daily use: a demo is short.

Confidence does not drop when accuracy does. An agent nine steps into a task built on a misreading reports its progress in exactly the same tone as one that is doing fine. There is no wobble in the voice to warn you.

Permissions are all or nothing, mostly. In practice you are usually choosing between an agent that has to ask you about everything, which is slower than doing it yourself, and one that does not ask, which is the thing you were worried about. The middle setting, where it acts freely on reversible things and stops for the rest, is where these tools need to get to and mostly have not.

They are slow and they are not cheap. Every step is a model call. A task with thirty steps costs thirty times a single question, and takes minutes rather than seconds.

How to tell a real agent from a chatbot with a marketing budget

Almost every AI product now describes itself as agentic, which has made the word close to useless as a filter. Four questions get you past it, and you can ask all four on a demo call.

"What tools can it call, specifically?" A real answer is a list: a browser, these three APIs, your calendar. A vague answer about capabilities means there are no tools and it is a chatbot.

"What happens when a step fails?" Ask them to show it. The whole value is in the recovery, so a product that has thought about agents has thought about this, and one that has not will change the subject.

"How many steps does it take before it comes back to me?" If the answer is one, it is not an agent. If the answer is unlimited, ask what stops a loop.

"What can it do without asking me?" The most important question, and the one to get in writing. You are looking for a product where the boundary is configurable and the default is conservative. A product that cannot answer this crisply has not designed the permission model, which means there is not one.

Should a small business be using one?

If you run something small, the useful question is not "are agents the future" but "is there a task here that is boring, repetitive, well-defined, and cheap to check". If yes, an agent will probably do it and it is worth an afternoon finding out. If no, you will spend more time supervising than the task was worth.

What I would not do yet: give an agent standing permission to send things to customers, spend money, or change records that are hard to put back. Not because the technology cannot, but because the failure mode is silent and the cost of finding out late is much higher than whatever the automation saved.

A reasonable order to try things in: start with research, move to summarising and extracting from documents you already have, then to drafting things a person sends. Reserve anything irreversible until an agent has been boringly reliable at the reversible version for a while.

What this costs you when it goes wrong

Worth pricing out, because the failure modes have different bills attached.

A research agent that gets it wrong costs you the time to check, which you were going to spend anyway. This is why research is the right starting point: the downside is bounded and visible.

A drafting agent that gets it wrong costs you an embarrassing email, once, and then you start reading them before they go out. Annoying, recoverable.

An acting agent with real permissions costs you whatever it did. A refund issued in error, a message sent to the wrong list, a record overwritten. The bill is not usually large in any single instance. What makes it expensive is that nothing alerts you, so you find out from a customer, some weeks later, and then have to work out how many other times it happened.

That asymmetry, rather than any view about the technology, is the reason to move through those three in order rather than starting at the end.

The honest summary

Agents are genuinely new and genuinely useful in a narrower band than the marketing suggests. The band is real work, not a demo, and it is growing quickly. The gap between "impressive in a video" and "trustworthy on a Tuesday" is still wide, and every vendor in this market has a reason not to mention it.

If you want to understand the machinery underneath any of this, how large language models actually work is the piece that makes agent behaviour stop being mysterious. If you are deciding which assistant to build on in the first place, start with ChatGPT vs Claude.