The distinction that matters
Both are built on the same kind of language model. The difference is not intelligence but authority and persistence. A chatbot is a conversation. An agent is a worker with a goal, a set of tools, and permission to use them.
Ask a chatbot “where is my order?” and a good one tells you. Ask an agent “my order is late, sort it out” and it can check the shipment, apply the goodwill credit your policy allows, and write confirming what it did.
| Chatbot | AI agent | |
|---|---|---|
| Answers questions | Yes | Yes |
| Takes action in other systems | No | Yes, within its permissions |
| Multi-step tasks | No | Yes |
| Typical build | 4 to 8 weeks | 3 to 6 months |
| Failure mode | Says something wrong | Does something wrong |
A chatbot that gets an answer wrong is embarrassing. An agent that takes a wrong action has to be prevented, logged and reversible, and building that is most of the difference in cost.
Describe the task as a sentence and look at the verb. “Tell me”, “explain”, “find” — chatbot. “Update”, “submit”, “reconcile” — agent. If the steps never vary, you may not need AI at all — a conventional workflow will be cheaper.
Side by side, in more detail
| Chatbot | AI agent | |
|---|---|---|
| Answers questions | Yes | Yes |
| Uses your documents | Yes, with retrieval | Yes, with retrieval |
| Looks up live records | Sometimes, read-only | Yes |
| Takes action in other systems | No | Yes, within permissions |
| Multi-step tasks | No | Yes |
| Typical build | 4 to 8 weeks | 3 to 6 months |
| Running cost per task | One model call | Several model calls |
| Failure mode | Says something wrong | Does something wrong |
| Audit requirement | Light | Substantial |
The last three rows change a budget. A chatbot that gets an answer wrong is embarrassing. An agent that posts the wrong bill has to be prevented, logged and reversible. Building that is most of the extra time.
Three buyer situations
Staff keep asking the same policy questions. That is a chatbot grounded with RAG. Do not buy an agent. There is no action to take except “show the paragraph”.
Customers ask where their order is. A chatbot that can look up the order — read-only — is enough for most firms. An agent that also issues a credit or rebooks a shipment is a later phase, after you have seen which questions actually arrive.
Finance retypes invoices that already exist as PDFs. The verb is “post”, not “tell me”. That is agent-shaped, but still start with extract-and-draft. Full posting without a click is how you earn a story you do not want.
The path that usually works
Week one to eight: a grounded chatbot on the content you already have, used by a real team, with citations. You learn what people ask and how rotten the files are. Month three onwards: add one read-only tool. Then one write that creates a draft. Then, if the numbers say so, a write that completes the job for the boring cases.
That path costs less than commissioning an autonomous agent on day one, and it is how most of the systems we actually ship come about. If the steps never vary and the input is already a form, skip both and use workflow automation.
Questions that sort a vendor meeting
What is the system allowed to do without a person? Where do the facts come from on Tuesday if the file changed on Monday? How many model calls does a typical task make? What do we log? Who on our side owns the exceptions? If “agent” and “chatbot” are used as synonyms in the reply, you are not being sold an architecture. You are being sold a word.
Ask them to describe the first eight weeks as if write-access did not exist yet. A serious partner can. A slide that starts with autonomy is a slide that has not priced permissions.
A conversation you can reuse on the next call
“We have a mailbox of supplier PDFs. We want drafts in accounts, not posted bills. The facts come from the PDF and the purchase order. If they do not match, open a task and stop. Log every tool call. After thirty days we will look at hours in that mailbox and the exception rate. If you are quoting an autonomous agent that emails suppliers, that is a different job — say so and price it separately.”
If they cannot repeat that back, they are not building what you asked for. If they can, you have a statement of work. The architecture behind it is on what AI agents are. The grounding, if answers must cite files, is RAG.
What we ship first when someone asks for “an agent”
Almost always a grounded chatbot or an extract-and-draft. We say that in the first meeting. If the verb in their sentence is “tell me” or “draft”, autonomy is a later phase they may never need. If the verb is “post” or “email the supplier”, we still start with a draft and a click. The extra months for an agent are permissions, logs, caps, and reversibility — not a smarter model. Paying those months before you have seen the questions is how pilots die.
If after the chatbot phase the log says people only ever ask three things, you may not need tools at all. If the log says they always then go and update a record, you have earned one read-only tool, then one draft write. That is the path on this page. It is slower to say in a pitch. It is faster to have in production. Keep the comparison table in the room: chatbot versus agent is authority and persistence, not a branding choice.
If the steps never vary and the input is already a form, skip both and use workflow automation. Paying agent money for a fixed path is the usual waste. Paying chatbot money for a write to Tally is the usual accident. The verb in the sentence still decides.
Running cost, said plainly
A chatbot is usually one model call per question, plus retrieval if you grounded it. An agent is several calls per task, plus every tool. That is fine on a finance mailbox of fifty invoices a day. It is visible if you let the whole company “sort things out” in a loop. Cap steps. Cache retrieval. Do not put an agent on a job a query would finish.
For architecture and cost drivers, see what AI agents are. For the wider decision, see AI for business.
One sentence to put on the purchase order
“Phase one is a grounded assistant or extract-and-draft with a person on every write; tools and autonomy are a separate phase with their own price.” If they will not split the PO, they are bundling a chatbot you need with an agent you do not. Split it yourself. You can always buy phase two. You cannot easily unbuy a loop that already emails customers.