The words get used interchangeably in marketing and they describe two different things with two different economics. Getting the distinction straight is worth real money, because it decides whether what you are buying can actually remove work.
A chatbot waits. An agent acts.
A chatbot is reactive. Someone opens it, types, and gets a reply. It has no memory of your business unless you paste it in, no access to your systems unless it has been wired up, and nothing happens when nobody is typing. The value ceiling is set by how many people are willing to open it and how good they are at asking.
An agent has a job. It runs on a trigger rather than on your attention. An email arrives, a file lands, a clock ticks. It reads what it needs, does the work, writes the result somewhere real, and stops. Nobody opened anything.
The interface difference is trivial. The economic difference is total: a chatbot multiplies what a person can do, an agent removes a task from the person entirely.
Why chatbots feel productive and change so little
Because the loop feels fast. You get an answer in three seconds instead of finding it in four minutes, and that feels like a large saving. It is a real saving, and it is small, because you still had to decide to ask, and you still have to do something with the answer.
Multiply a small saving by a person's willingness to keep initiating, and you get a modest improvement that never shows up in headcount, cycle time, or anything a finance director can see. This is why so many organisations are a year into an AI rollout with high satisfaction scores and no measurable change.
Why agents are harder, and worth it
An agent needs three things a chatbot does not:
Access. It must reach the mailbox, the folder, the database, the API. This is usually the real project, and it is the part that gets underestimated in every plan.
Boundaries. Because it acts without supervision, someone has to decide what it must never do. The useful rule is that it must be timid: anything it cannot be confident about, it leaves alone and lists for a person. A tool that quietly makes a hundred silent decisions is worse than no tool, because nobody can check it.
An exception owner. Every agent produces cases it will not touch. If that queue has no name against it, the work has moved rather than gone.
Give it those three and an agent takes a job off the list permanently. That is a different category of result from a faster answer.
Where each one is genuinely right
Use a chatbot when the work is unpredictable and judgement-heavy, when a human is going to review the output anyway, or when the value is in helping a person think rather than in removing a step. Drafting, brainstorming, explaining a codebase, first-pass research.
Use an agent when the work is repetitive, rule-shaped, high-volume, and recoverable if it is occasionally wrong. Filing, matching, checking, extracting, routing, reconciling, first-line replies to questions that have real answers.
Most teams need both, and buy only the first, because the first is the one that comes with a monthly licence and a login.
What does agentic AI mean?
"Agentic AI" describes AI systems that pursue a goal across several steps with some independence: they decide which step comes next, call tools such as a search, a database or an email system, check the result, and continue until the job is done or a limit is reached. The word is newer than the idea. In practice, an agentic system is an agent in the sense used above: it has a trigger, access, boundaries and a place to write its result.
A few concrete examples of AI agents, from simple to more complex:
| Example | Trigger | What it does | Where a person stays involved |
|---|---|---|---|
| Inbox triage agent | New email arrives | Classifies it, drafts a reply, files it | Approves replies before sending |
| Lead qualification agent | Form submitted or call answered | Asks qualifying questions, updates the CRM, books a meeting | Owns the sales conversation |
| Invoice processing agent | Supplier invoice lands in a folder | Extracts fields, matches to the order, flags differences | Approves payment |
| Due diligence reading agent | Documents uploaded to a data room | Reads them and returns findings with citations | Checks each finding and decides |
| Voice receptionist | Phone call | Answers questions, books appointments, transfers when needed | Takes transferred calls |
For more depth on two of these, see AI agents for lead qualification and sales follow-up and AI phone agent handoff to a human.
The question to ask a vendor
Not "does it use AI". Ask: "When nobody is logged in, what does it do?"
If the honest answer is "nothing", you are buying a chatbot. That may be exactly right for you, and you should price it as a productivity tool for individuals rather than as automation.
If the answer describes a trigger, a job and a place the result gets written, you are buying an agent, and you should be asking the next questions: what happens when it is unsure, who owns the exceptions, and what does it cost per run.
You can see both kinds running on real work, recorded, in the asynchronous course. If you would rather skip to having one built, that starts with a paid pilot.