AI agents vs chatbots: one answers, one finishes the job

A chatbot answers. An agent acts.

A chatbot takes a question and returns text. However good that text is, the work still lands on a person: someone reads the answer, opens the CRM, updates the record, sends the email.

An agent takes the same question, then reads your systems, decides what needs doing, does it, and writes the result back. The conversation is a side effect, not the product.

Everything else in this AI agent vs chatbot comparison follows from that one distinction.

Retrieves and phrases. It searches a knowledge base or a model's training and returns a well-written response.

Follows a decision tree. Most support chatbots are branching flows with language on top. The branches were written by a human in advance.

Hands off when stuck. The standard escape hatch is "let me connect you to someone", which is an admission that the work is still human work.

Stays inside the chat window. This is the important one. A chatbot's output is a message. It does not have credentials to your practice management system, your ledger, or your ATS, and it is not expected to.

Put the AI agent vs chatbot question another way: an agent is defined by having tools and permission to use them. In practice that means four things a chatbot does not do.

• It reads state. Before answering "where is my order", it queries the order system rather than guessing from a help article.
• It takes multi-step actions. Check the conflict database, draft the engagement letter, request the signature, then file the result. Each step depends on the last.
• It writes back. The record in your system is updated, so nobody re-keys anything afterwards.
• It escalates on judgment, not on failure. A well-built agent stops when a decision needs a human, and routes that decision with the context attached.

That last point is the one buyers underrate. The value is not full autonomy. It is that the 80% which is mechanical gets finished, and the 20% that needs judgment arrives on your desk already prepared.

The row that matters commercially is the second-to-last one. If a tool answers beautifully and leaves the same amount of work on your team, it has not changed your capacity.

Plenty of the time. Do not over-buy.

Your questions are genuinely repetitive and read-only. Opening hours, return policy, does this plan include X. A chatbot deflects these cheaply and you should let it.

You have no systems worth integrating. If the work lives in one person's inbox and a spreadsheet, an agent has nothing to read or write and you are buying complexity you cannot use.

Volume is low. Under a few hundred interactions a month, the integration work rarely pays back inside a year.

We turn down work on this basis fairly often. If the honest answer is that a help centre and a decent chatbot would fix it, an agent deployment is the wrong purchase.

The expensive mistake is buying a chatbot against an agent business case.

The pitch is usually deflection rate. A chatbot deflects 40% of tickets, so you model 40% less support labour. What actually happens is that the 40% deflected were the fastest tickets, the ones taking two minutes. The queue that remains is the slow, multi-system work, and headcount does not move.

Measure the right thing instead. Not "what share of conversations ended without a human", but what share of the work finished without a human touching a system. That number is usually close to zero for a chatbot, and it is the number that shows up in capacity.