AI agent versus chatbot, and which one your job actually needs

The one difference that produces all the others

Both are things you type at, and both usually sit behind the same message box, which is why the words get used as though they were interchangeable. The difference has nothing to do with how clever the model is. It is whether the system has hands.

A chatbot receives a message and returns text. The text may be excellent, grounded in your own documents and written in your house style, but nothing anywhere else is different afterwards. An agent receives a message and can decide to call a tool: look this customer up by phone number, check delivery slots for that postcode, create a draft invoice. It reads what comes back, decides whether it needs another tool, and then answers. Something outside the conversation has moved.

That single capability changes the buying question. With a chatbot you are asking whether the answers are correct. With an agent you are asking that and three more: what is it allowed to touch, who checks what it did, and how would you find out if it did the wrong thing at three in the morning.

What each one is genuinely good at

Sorted by job rather than by technology, the split is unusually clean.

Notice how much of the useful work in an ordinary business falls into the second group. That is why agents get bought. Notice also how much of that work involves writing to a system somebody depends on. That is why they get bought badly.

• A chatbot suits questions whose answers already exist in writing: policies, product details, what is covered and what is not, how a process works, what your hours are during a holiday.
• A chatbot suits volume at low stakes, where a wrong answer is caught by a person and nothing has to be undone in a system.
• An agent suits anything where the answer depends on a live value: where is my order, is that slot free, how much is left on the account, has the payment landed.
• An agent suits work that ends in a change rather than a sentence: booking, rescheduling, updating a record, raising a ticket with the right fields already filled in.
• An agent suits multi step work where the next step depends on what the last one returned, which a fixed script cannot express without somebody enumerating every path in advance.

How each one fails, and why one is louder

A chatbot fails by being wrong in text. Somebody reads a confident answer built on a document that was superseded last year and acts on it. That is a real cost, but a recoverable one, and the repair is in the content rather than in the model.

An agent fails by being wrong inside your systems. The same confident wrongness now has a record attached: a booking made against the wrong customer, a duplicate order created because a retry ran twice, a field overwritten with a value that came back from the wrong lookup. Recovery means finding every action it took, which is only possible if every action was logged with its arguments and its result.

The failure that is hardest to catch is neither of those. It is an agent doing something reasonable that nobody asked for. It was asked to check an order and it also updated the delivery address, because updating addresses was in its tool list and the message happened to mention a new one. Nothing errored. Nobody was told. It will be found by a customer.

The supervision bill nobody puts in the proposal

A chatbot needs its content maintained. Somebody owns the documents, keeps one canonical answer per question, and removes superseded versions instead of leaving them beside the current one. That is real work, and it is work a non technical person can do well.

An agent needs all of that plus a review habit. Somebody reads a sample of runs each week, watches for actions that got reversed, and notices when a system it depends on quietly gains a field or renames a product. Skip that and the agent drifts, because nothing errors when the world changes shape underneath it.

Ask who does this before the build rather than after. Where an agent gets quietly abandoned six months in, the build is rarely the reason. Nobody had the job.

Choosing between them for one specific job

Take a single job you are considering and answer four questions in order. The first no decides it.

Plenty of businesses asking for an agent need a chatbot with two lookups attached, which is a smaller build, a shorter test cycle and a system they can actually operate. The full version is worth having later, once the review habit exists and the tool list has been argued over properly by people who understand the consequences.

• Does answering require a value that changes during the day? If no, a chatbot over good content will do it more cheaply and more predictably.
• Does a successful outcome involve something changing in a system? If no, you still do not need an agent.
• If something does change, can a person undo it in a few minutes? If no, you want an agent that proposes and a person who confirms, not one that acts alone.
• Is there a named person who will review a sample of its actions every week, indefinitely? If no, build the version that drafts and let a human press send.

The two side by side

Only the first row is about what the software can do. Every row under it is about what happens after it is live, which is where the difference in price and in effort actually sits.

Criterion	Chatbot	AI agent
What it does with a message	Returns text, and nothing outside the conversation is different afterwards	Can decide to call a tool, so something outside the conversation moves
Questions it suits	Answers that already exist in writing: policies, product details, hours	Answers that depend on a live value: where is my order, is that slot free
What success looks like	A sentence	A change: a booking, a record updated, a ticket raised with the fields filled
How it fails	Wrong in text, caught by a person, repaired in the content	Wrong inside your systems, with a record attached to undo
Its hardest failure to catch	A confident answer built on a document superseded last year	A reasonable action nobody asked for, which nothing errors on
What has to be maintained	The content: one canonical answer per question	The content, plus somebody reading a sample of runs every week
The buying question	Are the answers correct	That, plus what it may touch, who checks it, and how you would find out

Jobs where neither one belongs

If the work has one correct path and no judgement in it, a plain workflow automation beats both. A form submission that has to create a record, send a confirmation and add a task needs no conversation and no model, and putting one in the middle only adds a way for it to be wrong.

If an answer has to be right every single time for legal or safety reasons, publish it and route the question to a person. A model that is right almost always is the wrong instrument for a question where almost always is a liability someone signs for.

And if the underlying content contradicts itself, neither will save you. A chatbot picks one version and states it confidently, and an agent acts on it. Deciding which version is correct is a judgement the business has to make first, and no amount of model quality substitutes for making it.

The supervision bill, in hours rather than adjectives

The supervision bill has a shape as well as a size. An agent's supervision is a stop list written before it goes live: money, pricing, anything published in your name and any complaint that has turned serious all wait for human approval, and a case outside the tool list hands it to a person with the conversation attached. A chatbot's supervision is a document owner and a rule that there is one canonical answer per question. Neither is optional, neither appears in a proposal, and that is why the fourth question in the list above decides the price.

Timing separates them too. A chatbot over documents you already hold can be live in days and is then a content job for as long as it exists. An agent takes longer at the start and less afterwards: the audit takes the first week, one agent with two read tools and a single write tool is live inside a fortnight, and further write tools are added one at a time, because each needs its own confirmation and its own trace before it is trusted with anything.

Whichever you choose, hold the pieces. The prompts, the tool definitions, the document set and the run traces sit in accounts under your own logins, so you own the system and the traces exist when somebody asks why it answered that. Building the chatbot half in-house is realistic for most businesses, because it is a documents problem rather than an engineering one, and the discipline of one canonical answer per question is worth more than whoever you would pay to impose it.

The published version of the agent side is Big Texas Land Buyers, where inbound and outbound voice agents run more than 500 calls a day with automatic CRM categorisation, and Center for Sight in New York, booking appointments largely hands-free. Ibrahim owns build and workflows at Wobble, which works from Karachi, bills month to month and carries 25 engagements across six countries.

Common questions

What is the difference between an AI agent and a chatbot?

A chatbot returns text and changes nothing outside the conversation. An agent can call tools, which means it reads from and writes to real systems and chooses which tool to use and in what order. That is the whole distinction, and everything else, from cost to testing to who has to supervise it, follows from whether the system can act.

Is an AI agent better than a chatbot?

Better at different jobs. If the question can be answered from something already written down, a chatbot is cheaper, faster to build and easier to predict. If answering requires a value that changes during the day, or the outcome involves something being booked, updated or created, only an agent can finish the job.

Can a chatbot book appointments?

Not by itself. Booking means writing to a calendar or a booking system, which requires a tool, and a system with tools is an agent regardless of what the vendor calls it in the proposal. A chatbot can explain how to book and hand over a link, which is often enough and is considerably less to maintain.

Which is more expensive to run?

An agent, and usually by more than the build price suggests. It carries the same content maintenance a chatbot needs, plus integration work when a connected system changes, plus somebody reviewing a sample of its actions every week. The recurring human cost is the part most often missing from a comparison.

How do I stop an agent doing something I did not want?

By making the unwanted action impossible rather than discouraged. An instruction in a prompt is a request, and the actual control is a tool that refuses: separate credentials for reading and writing, hard limits on value, and a person confirming anything that moves money, deletes a record, changes access or messages a group of customers.

Do we have to choose one, or can we start with a chatbot and add tools later?

Starting with answers and adding lookups afterwards is a sensible sequence, and it is how most workable systems arrive. You learn what people actually ask before deciding what the system should be able to do about it, and each tool gets added deliberately rather than as part of a package nobody itemised.