The words get used interchangeably and the difference is real. A scripted chatbot recognises set phrases and returns set replies, following a decision tree somebody drew in advance. An AI agent interprets what was actually said, works from a body of information rather than a script, and can take an action: look up a time, create a record, escalate to a person.
The practical test is what happens to a question nobody anticipated. A chatbot dead-ends, usually into "I didn't understand that, here are some options". An agent either answers it from what it knows or hands it to someone who can.
Definition
AI agent
Software that interprets a request in natural language, works from a defined body of information, and can take actions in other systems on the basis of it, as opposed to a chatbot, which matches inputs against a predefined script and returns predefined replies.
Where the scripted version is still the right answer
Scripted flows are not obsolete, and pretending otherwise leads to models being put where a rule would do. A fixed script is better when the path is genuinely fixed and being wrong is expensive: taking a payment, confirming a legal acknowledgement, a compliance step with specific wording. You want those identical every time and inspectable by a person who does not have to understand a model.
The rule of thumb: if you can write every branch down and they are all the branches, write them down. Interpretation is for where you cannot.
What makes an agent actually work
Three things, and a deployment missing any of them tends to be the one that embarrasses somebody.
1. Grounding
The agent answers from your information (your pages, policies and documents) rather than from a general model's impression of businesses like yours. Without grounding you get answers that are plausible, fluent and about a business that does not exist. With it, the answer is traceable to where it came from.
The important half of grounding is the negative case: what happens when the material does not cover the question. The correct behaviour is to say so and hand over. Producing a confident sentence anyway is the failure everyone means when they say a model "made it up".
2. Boundaries
Written down before the build: what it must never answer, what it must always escalate, what it says when it does not know. In a clinical setting that means no advice, no triage, no interpretation of symptoms, enforced in the build, not appended as a disclaimer under the chat window.
Boundaries are also what make an agent testable. "Does it work?" is unanswerable. "Does it refuse this list of twenty things and escalate this other list of fifteen?" is a test you can run.
3. Escalation that keeps the context
When it hands over, the person receives the conversation. Being asked to start again is what makes a handoff feel worse than no automation at all. The customer has now spent time and arrived where they would have been anyway, minus the patience.
The question about hallucination
It is the right question. The answer is not that modern models do not do it. It is that the system around the model decides how much it matters. A grounded agent with a narrow scope, explicit boundaries and a low threshold for escalation has very little room to invent, because most of what it might invent is outside what it is allowed to answer at all.
Test for it specifically before launch, using the hard cases rather than the easy ones: the ambiguous question, the angry one, the one that skirts a boundary, the one about something you do not offer.
Choosing between them
- The questions are unpredictable and phrased differently every time: agent.
- The path is fixed, short, and being wrong is expensive: script.
- It needs to do something in another system, not just answer: agent.
- It needs to be identical every time and inspectable by a non-technical person: script.
- The honest answer for most businesses is both: a script for the two or three fixed paths, an agent for everything that does not fit one.
The thing that actually decides it
Neither category tells you whether a deployment will be any good. What decides that is whether someone did the unglamorous work: collecting the questions people actually ask, writing the boundaries before the build, testing against the hard cases, and reading the transcripts for a month afterwards.
The log of questions it could not answer is usually the most valuable output of the first month. It is a list of what is missing from your own material, written by your customers.
