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

The word chatbot now covers three quite different pieces of software, and "AI agent" is used loosely enough that two vendors can quote for different things under the same name. The distinction matters before you compare quotes, because the three fail in different ways and cost different amounts to run.

Three things sold under one name

What actually makes something an agent

Retrieval: it looks things up before answering

The question is matched against your own material — services, fees, timings, policies — and the relevant passages are handed to the model as the source it must answer from. This is retrieval-augmented generation: the agent states your opening hours because your page says so, not because most businesses keep similar hours.

Tools: it can do something

The agent is connected to specific functions in your systems — check availability, create a booking, add a lead to the CRM, send a WhatsApp confirmation — and decides during the conversation when to use them. Its output is a change in your business, not only a reply.

A defined stopping point

A rule, written before anything else, for what it must not answer or do, and who it hands over to with the conversation attached. An agent without one is a liability, however good its answers are.

Some vendors also call a background task an agent: following up a quote the next day, sending a reminder, summarising yesterday's enquiries. That is better described as AI automation, and it usually runs alongside an agent rather than inside it.

When each one is the right answer

A menu bot, when the choices are few and fixed

If nearly every message is one of a handful of requests — order status, branch timings, a payment link, a callback — buttons are quicker than typing, cannot invent anything, and cost little to run beyond the messaging itself. On WhatsApp especially, a good menu is often the sensible first step, and it can grow into an agent later.

A plain LLM chatbot, almost never for customers

A general model with no grounding answers questions about your prices and policies with whatever sounds plausible. It is useful inside a business for drafting and summarising, where a person checks the result. Facing customers, it is the version most likely to promise something you do not offer.

An AI agent, when questions vary and something should happen

The case for an agent is a steady volume of questions asked in many phrasings — what a service covers, whether it suits a situation, what it costs, when the next slot is — where the useful outcome is a booking, a qualified lead or a handover. Evening enquiries strengthen the case, because that is when nobody else is answering.

When you do not need a bot at all

• Your enquiry volume is low. If a person can answer every message within the hour, they will do it better than any software.
• Your content is thin or contradicts itself. An agent is only as accurate as its sources. If two pages quote different fees, fix that first.
• Every answer needs professional judgement. Where the question is really a diagnosis or legal advice, the bot's job shrinks to collecting details and booking a consultation, which a form does well.
• Nobody will watch the handovers. A bot that passes difficult conversations to an inbox no one checks is worse than no bot, because the customer was promised a person.

The risks, and what controls each one

Made-up or stale answers

Controlled by grounding the agent in the same content your site publishes, instructing it to answer only from what it retrieved, and giving it a plain "let me pass you to someone" when the content does not cover the question.

Being talked into things

People will try to make a bot ignore its instructions or offer a discount, a technique known as prompt injection. The defence is limiting what the agent is able to do in the first place, not trusting it to refuse.

Actions that go wrong

Give the agent the narrowest permissions that do the job: it can create a booking but not cancel someone else's. Anything involving money or a commitment should be confirmed by the customer or a member of staff first.

Personal data

Conversations collect names, phone numbers and often more sensitive detail, and India's Digital Personal Data Protection Act 2023 applies: collect what the purpose needs, say what it is for, keep it secure, and do not keep it longer than necessary.

What to ask a vendor before you sign

1. What exactly will it answer from, and how is that updated when our content changes?
2. What does it do when the answer is not in our content? Show us.
3. Which actions can it take in our systems, and which ask for confirmation first?
4. When does it hand over to a person, and where does that person see the conversation?
5. Where are conversations stored, for how long, who can read them, and which model provider processes them?
6. What will it cost to run each month as conversations grow, and who owns the code and configuration?

The wider version of that test is in how to choose a software development company. MyFloww builds AI agents and AI chatbots for websites, WhatsApp and the software we build for clients, each grounded in the client's own content with the handover written in from the start.