What “AI agent” actually means (and what it doesn’t)

“Agent” has quietly become the most stretched word in software. Right now it gets applied to anything with an LLM API key behind it — a chatbot widget, a form that auto-fills, a script that summarizes an email. Some of those are genuinely agentic. Most aren't. The difference matters, because it's usually the difference between what you're actually paying for and what you think you're paying for.

The actual test

A chatbot answers a question. An agent does something about it. That's the whole test. If the system can only produce text back to a human who then has to go take the action themselves, you have a well-dressed FAQ, not an agent — however good the writing sounds.

A real agent can look something up, decide what to do with what it found, and take the next step without a human relaying it by hand: check an order status in your system, update a record, escalate to a person when it's genuinely unsure, or hand off cleanly with the right context attached instead of a dead end.

A concrete example

Say a customer messages asking where their order is. A chatbot answers with generic shipping policy text because that's all it has access to. An agent checks the actual order in your system, sees it's delayed, and either replies with the real status or flags it to a human with the order ID already pulled up — no copy-pasting between five tabs to find that same answer.

Same input, same LLM underneath in both cases. The difference is entirely in what it's wired into.

What to actually ask before buying one

• What systems does it actually read from and write to — or does it just generate text?
• What happens when it doesn't know the answer? Does it guess, or does it hand off?
• Can you see what it did, after the fact, not just what it said?
• Is there a human in the loop for anything that changes real data or costs real money?

None of this makes chatbots useless — a well-scoped one is often the right, cheaper answer. But it should be sold as what it is. When we scope an AI System, the first question we ask is whether the workflow actually needs an agent that can act, or whether a good answer is enough. That decision changes the build, the price, and what you should expect from it — and it's the honest starting point, not an upsell.