What AI agents actually are — and what most articles get wrong

I've been in dozens of conversations about AI agents over the past year. With potential clients, with people at industry events, with journalists writing about the space. And there's a pattern I keep noticing: most of the framing is either too narrow or too expansive.

Too narrow: "an AI agent is basically a chatbot that can do more things." Too expansive: "AI agents are autonomous systems that can handle entire departments." Neither is very useful if you're trying to make a decision about whether to use one.

So here's my attempt at a grounded explanation. This is how I think about agents when I'm deciding whether to build one for a client.

Start with the word "agent"

In software, an agent is something that takes actions in pursuit of a goal. That's it. A thermostat is, in a loose sense, an agent — it perceives the environment (temperature), compares it to a goal (your set temperature), and takes action (turns on the heating).

What makes AI agents different from older software agents is that they use a language model as the "brain" — the thing that perceives the situation and decides what to do. This gives them a qualitatively different capability: they can handle unstructured input, apply something like contextual reasoning, and deal with variation in ways that rule-based systems can't.

A traditional automation rule says: "if invoice format A, extract field X." An AI agent can read an invoice it's never seen before, understand that the number in the top-right corner is the invoice total even though it's labelled differently, and extract it correctly.

That's genuinely useful. It's also genuinely limited in specific ways.

The thing most explainers miss: agents are not autonomous in a vacuum

When people say an AI agent is "autonomous," they usually mean it can complete a task without someone having to do it manually. That's true in a narrow sense. But agents operate within systems — they have access to specific tools, specific data, and specific permissions. They can't do things outside those boundaries.

An agent I built for a client last year can process incoming supplier invoices, extract the data, validate it against expected values, and push it to their ERP. It does that without anyone touching it. That's genuinely autonomous for that specific task.

But it doesn't decide to add new suppliers. It doesn't notice that a supplier is consistently sending wrong VAT numbers and escalate that as a relationship issue. It doesn't read the news and update its behaviour accordingly. The autonomy is bounded and specific.

When evaluating AI agents, the question isn't "how autonomous is it?" — it's "what specific things is it autonomous about, and what are its failure modes?"

How agents differ from chatbots

Chatbots are response machines. You give them input, they give you output. A customer service chatbot reads your question and generates a reply. That's the loop.

Agents take actions in the world. They don't just generate text — they call APIs, read from databases, create records, send notifications, update systems. The language model is still doing the core reasoning, but it's connected to tools that let it act.

A chatbot that says "I've processed your return" might be lying — it just said the words. An agent that processes returns actually executes the steps: verifies eligibility, creates a return record, triggers a refund, sends confirmation. There's a meaningful difference.

That said, many systems branded as "AI agents" are closer to chatbots than they admit. The label has become a marketing term. I'd encourage anyone evaluating these tools to ask specifically: what actions does it actually take, in which systems, and how are those actions confirmed?

What makes an AI agent appropriate for a business process

Not every process benefits from an AI agent. In my experience, the cases where agents genuinely earn their complexity are ones with these characteristics:

• Unstructured or variable inputs. If inputs are always identical and structured, a traditional rule-based automation is simpler and more reliable. Agents earn their keep when input formats vary or when contextual interpretation is needed.
• High volume, repetitive decisions. If the same class of decision happens dozens or hundreds of times a day, the overhead of building and maintaining an agent is justified.
• Clear success criteria. You need to know what "correct" looks like — both for testing and for ongoing monitoring. Agents that produce output you can't validate are risky.
• Tolerance for occasional errors. AI agents make mistakes. This isn't a dealbreaker, but you need a process for catching and correcting them. If the process has zero tolerance for error, you need a human in the loop.

One more thing: the quality of your data matters enormously

I have had conversations where a client wants to automate a process, and the first thing I discover is that the underlying data is inconsistent, incomplete, or stored in three different places with slightly different formats. An AI agent can handle some of that variation — but it can't compensate for fundamentally bad data.

If your CRM has duplicates, your email folders are a mess, or your documents don't have consistent naming conventions, those problems will slow down or compromise any automation you try to build. This isn't a dealbreaker — it just means there's scoping work to do before the agent work starts.

The short version

AI agents are software systems that use language models to perceive situations and take actions in connected systems. They're more capable than rule-based automation for variable, unstructured inputs. They're more limited than the "autonomous AI" framing suggests. They work best when the task is high-volume, the inputs are varied, and there's a clear way to verify their output.

That's the mental model I use. It's not comprehensive — the field is moving fast — but it's practical enough to help with real decisions.