The difference at a glance
| Rule-based chatbot | AI chatbot | AI agent | |
|---|---|---|---|
| How it decides | Buttons and a decision tree someone drew | A language model writes each reply | A language model chooses the next step |
| What it knows | Only the answers written into the tree | Your website, files and documents | Your knowledge plus live data from your systems |
| What it can do | Show the next set of options | Answer, in the customer's own words | Answer, and act: look up, book, open a request, hand off |
| Example | “Press 1 for opening hours” | “Our returns window is 30 days” | “Your order 10482 shipped today. Shall I book the delivery slot?” |
| Main risk | Can't handle anything off-script | Can invent an answer | Can take a wrong action |
What makes something an agent
Anthropic, which makes the Claude models, draws the line by who decides the steps: in a workflow, code sets the path; agents are systems where the model directs its own process and its use of tools. Anthropic: building effective agents
IBM describes the same difference from the other side: a non-agentic chatbot has no tools, little memory and mostly handles one request at a time, while an agent breaks a goal into steps and uses tools to complete it. IBM: what are AI agents?
In practice, a customer-facing agent has four things a chatbot doesn't:
- Tools: things it can do, such as check stock, book a time or open a case.
- Its own choice of steps: it decides whether to answer, look something up, ask a question or act.
- Memory of the conversation and the customer: what they already said, their name and number.
- A way out: it knows when to stop and hand the conversation to a person.
When a chatbot is enough, and when you need an agent
Start with the simplest thing that works. Anthropic's own advice is that for many uses, a single model call with good retrieval is enough, and agents are worth their extra cost and slowness only when the task needs flexibility.
| Customers mostly want to | What fits |
|---|---|
| Find out your hours, prices and policies | An AI chatbot trained on your website and files |
| Book an appointment or a viewing | An agent with your calendar |
| Check stock or where their order is | An agent with read-only access to your data |
| Ask for a refund, a repair or a quote | An agent that opens a case for your team |
| Talk to a person | Either, as long as it hands off cleanly |
Keeping an agent safe
Once software can act, a wrong guess becomes a wrong booking or a wrong promise. The answer isn't to avoid agents but to limit what they can do alone:
- Give it the least access that works: read-only for your data, and only the tables it needs.
- Check important actions in code, not just in the instructions: a booking only for a time the customer chose, data only the customer may see.
- Have a person approve what can't be taken back, such as a refund decision or a reply in a dispute.
- Record what it did, so you can see every lookup and action afterwards.
- Let it hand off. An agent that knows when to stop is safer than one that always tries.
Our guide to stopping a chatbot from making things up goes further into which rules belong in code.
Questions to ask before you buy
- Which actions can it take, and which need a person's approval?
- Are its actions checked by the system, or only described in its instructions?
- Does it read your data read-only, and can you limit which parts?
- What does it do when it doesn't know: guess, ask, or hand off?
- Can you see every action it took, and why?
- Does it work on the channels your customers use, such as WhatsApp?
- How is it billed: per message, per conversation or per action?
Where HeyKiko fits
HeyKiko's Kiko is an agent. It answers from your knowledge on your website, WhatsApp and Telegram, and it can act: book appointments into Google Calendar, look up live data read-only, recommend products, open cases whose replies a person approves, and hand the chat to your team.
What it doesn't do: take payments, issue refunds, or send email on its own. Bookings are checked against the time the customer chose, and what an anonymous customer can read from your database is checked on every lookup.
There's an agent for your side too: Ask Kiko reads your records and suggests changes that apply only when you approve them. You can try both free for 14 days.
Sources
Checked on 5 November 2026. We make HeyKiko, which this guide uses as its example.
