Ask three vendors for an “AI agent” and you may get three different things: a chat window with scripted buttons, an automation that sorts your enquiries, and a system that moves appointments in your calendar on its own. The labels overlap because they sell; the products don’t. The difference between an AI agent and a chatbot decides what you pay, what can go wrong and who keeps it running.
A chatbot that gives a wrong answer is embarrassing. An agent that cancels the wrong booking is a business problem. Below are the four options side by side, with five questions to decide between them. For the wider picture, start with AI agents for business.
The short answer
The difference between a chatbot and an AI agent comes down to two questions: who decides the next step, and can it act?
- A rule-based chatbot follows a script someone wrote.
- A workflow with an AI step follows fixed steps, with AI handling one narrow task inside it, such as sorting an enquiry.
- An AI assistant answers questions in its own words from your content, but doesn’t act.
- An AI agent is given a goal, tools and limits, and chooses its own steps, including actions in systems such as your calendar or CRM.
For most websites, the right choice is a workflow with an AI step, or an assistant answering from well-kept content. An agent earns its place only when the task is open-ended, the input unpredictable, mistakes reversible, success measurable and the data available. Sometimes a clear FAQ page, a short form or a click-to-chat link beats all four.
Four things sold under similar labels
1. Rule-based chatbots: a decision tree in a chat window
A rule-based chatbot shows buttons or matches keywords, then follows a path written in advance: “Are you buying, renting or selling?” leads to three scripted branches. Every answer is written by a person.
It works for routing visitors and answering fixed questions such as opening hours. Anything the tree doesn’t cover ends in a loop or a dead end.
2. Workflows with one or two AI steps
A workflow is a fixed sequence: when this happens, do that, then that. What’s new is a language model in one step, handling the part that used to need a person reading free text.
For example, a visitor submits your enquiry form. An AI step labels it as a sales lead, support request or job application and pulls out details such as product and timing. The workflow then creates a CRM record, alerts the right person and drafts a reply for them to check. The path is fixed and the AI only fills in a blank, so behaviour is predictable and easy to test.
3. AI assistants that answer from your content
An AI assistant writes its own answers, usually after looking up relevant passages from your website or documents, a technique called retrieval-augmented generation (RAG). “Do you deliver overseas?” gets a written answer with a link to the source page.
It reads but doesn’t act. The main risk is a confident wrong answer, and answer quality depends heavily on the content it draws from, as preparing an AI knowledge base explains.
4. AI agents that choose their own steps
An AI agent is given a goal, tools and limits, then decides which steps to take, checking each result before choosing the next. Ask it to move an appointment and it might look up the booking, check the calendar, offer two slots, update the booking and send a confirmation.
Tools, the connections through which it reads and writes data in real systems, are what let it act, and where the risk lives. Connecting an AI agent to your website, CRM and booking system explains how they work.
Anthropic’s engineering guide “Building effective agents” (December 2024) draws the same line: workflows follow predefined code paths, while agents let the model direct its own process and tool use. It recommends the simplest solution that works.
AI agent vs chatbot vs workflow, side by side
| Compared | Rule-based chatbot | Workflow with an AI step | AI assistant | AI agent |
|---|---|---|---|---|
| Who decides the next step | Your script | Your process design | You set the scope; the model writes answers | The model, within your limits |
| Can it take actions? | Only what the script allows | Yes, the fixed actions you built | No | Yes, through connected tools |
| Set-up effort | Low: write and test the tree | Moderate: map the process, connect systems | Moderate: prepare and structure content | High: tools, permissions, guardrails, testing |
| Running-cost drivers | Platform subscription | Model usage per run, plus the automation platform | Model usage per conversation, rising with traffic | Model usage per task, multiplied by steps and retries |
| Predictability | Fully predictable | Fixed path; the AI step’s output varies | Wording varies; facts shouldn’t | Path and outcome can both vary |
| Main risk | Visitors stuck off-script | An enquiry sent to the wrong place | Confident wrong answers | Wrong actions in live systems, prompt injection, runaway costs |
| Upkeep | Rewrite the script when things change | Monitor runs, spot-check the AI step | Keep content current, read transcripts | All of that, plus action logs and permissions |
Each column to the right adds flexibility and something new to manage, so move right only when that flexibility solves a real problem. A key risk in the agent column is what the OWASP Top 10 for LLM Applications lists as Excessive Agency: more functionality, permissions or autonomy than the job needs. AI agent risks covers the controls that limit it.
Do you need an AI agent? Five questions that decide
Answer these for one specific task, not for “AI on the website” in general.
- Is the task open-ended? If you can write the steps down in advance, even with a few branches, it’s a workflow. An agent adds value only when the right sequence depends on what it finds along the way.
- Is the input unpredictable? If visitors choose from a few options, buttons or a form will do. Free text in countless phrasings and languages needs AI somewhere, perhaps only in one step.
- Can mistakes be caught and undone? A mistagged enquiry takes seconds to fix. A cancelled booking, a refund or a message sent in your name may not. Irreversible actions need a person to approve them, or they stay out of scope.
- Is there a clear measure of success? You need a way to check each task was done right: the booking exists at the right time, the enquiry reached the right person.
- Is the data available? It must be digital, current, reachable through a connection your systems support, and yours to use.
How to read your answers
- Yes to all five: an agent may be justified. Start with one narrow task, have a person approve every action at first, and widen its freedom only when logs show it’s reliable.
- Fixed steps, free-text input: a workflow with an AI step, the most common result.
- The job is answering, not doing: an assistant grounded in your content, or simply a better help page.
- Predictable input: a rule-based chatbot or, more often, a good form.
- No to question 3, 4 or 5: not an agent yet. Close that gap first: digital transformation for small and growing businesses shows how to map a process and choose what to automate.
Why a workflow with an AI step covers most needs
Most website processes are repeatable: an enquiry arrives, gets sorted, reaches the right person and gets a reply. What varies is how people describe what they need, not the steps. So put AI where the variation is and keep everything else fixed:
- Costs are bounded, because each run calls the model a known number of times.
- Testing is straightforward. Run a hundred past enquiries through it and check the labels before it touches a live one.
- Failures are visible. You know which step went wrong, and logs show where people still step in: the places where more autonomy might later pay off.
Good first candidates: routing enquiries with a short summary (AI lead qualification goes deeper), drafting replies a person checks, translating messages from customers who write in other languages, and flagging urgent complaints.
Put the human in the loop where it counts
“Human in the loop” means a person reviews or approves what the system produces before it takes effect. You need one where a mistake is costly or permanent: anything sent in your name, anything involving money, and any change to a booking or record a customer relies on. Low-risk steps such as tagging can run alone, with spot checks.
When a page, a form or a click-to-chat link beats all of them
Before adding any of the four, check whether a simpler fix solves the problem. (Adding AI to your website weighs each AI feature the same way.)
- A well-built FAQ page. If the same dozen questions fill your inbox, answer them on the site. Search engines can index the answers, and any assistant you add later needs that content anyway.
- A short, well-designed form. When you need specific details to reply, the right fields collect them faster than a chat asking one question at a time. See the web form design guide.
- A click-to-chat link. The major messaging apps let a business share a link that opens a conversation in an app the visitor already uses. A person replies, and a plain link adds no widget script to slow your pages (see third-party scripts and site speed).
- A visible phone number or booking link. For high-value, relationship-led sales, many buyers want a person from the start.
These win when enquiry volumes are modest, each enquiry is valuable and your team replies promptly. A website chatbot standing between a buyer and the person they wanted to reach can cost more than it saves.
Questions to ask before you buy
Whatever a vendor calls its product, these questions show what it really is:
Frequently asked questions
What is the difference between a chatbot and an AI agent?
A chatbot holds a conversation, following a script or answering from content. An AI agent works towards a goal, choosing its own steps and using connected tools to act, such as updating a booking.
Can a chatbot be turned into an AI agent later?
Often, if it’s an AI assistant rather than a scripted bot: it becomes an agent once it’s given tools that act and the freedom to decide when to use them. Start read-only and add one action at a time, with a person approving each at first. A purely scripted chatbot has no language model to build on, so it’s usually replaced instead.
Are AI agents more expensive to run than chatbots?
Usually, per task. A rule-based chatbot typically costs a flat subscription, while an agent may call a model several times per task, plus retries and monitoring. Ask vendors to estimate costs at your real volumes, and set a spending cap.
Start with what your website already does
Any of these is only as good as what sits underneath: current content, forms that reach the right person, tracked enquiries and connected systems. Often, getting those right solves the problem a chatbot was meant to fix.
A free website audit is a sensible first step: it looks at how easily visitors can get in touch today, which helps show whether a missing page or a clumsy form is the real gap before you pay for an assistant. If the gap turns out to be the website itself, talk to us about what your website needs.