Website chatbot vs AI agent: what is the difference?

TL;DR

• A website chatbot holds a conversation. An AI agent works toward a goal: it decides which steps are needed and uses tools, such as your calendar or CRM, to take them.
• Many products sold as "AI chatbots" are a language model answering questions from your content. That is useful, but by the definitions Anthropic and OpenAI publish, it is not an agent.
• Agents can do more and can get more wrong. Start from the job you need done, then pick the least autonomous system that does it.

The short answer

A chatbot's job is to reply. You ask, it answers, and the conversation is the product. An AI agent's job is to get something done. It reads the request, works out the steps, calls tools to carry them out, checks the result, and keeps going until the task is finished or it needs a person.

Two of the major model vendors draw the line in similar places. Anthropic's engineering guide Building effective agents separates workflows, "systems where LLMs and tools are orchestrated through predefined code paths," from agents, "systems where LLMs dynamically direct their own processes and tool usage." OpenAI's A practical guide to building agents defines agents as "systems that independently accomplish tasks on your behalf" and says that simple chatbots, which use a model but don't let it control the flow of work, are not agents.

So the difference is not whether a large language model is involved. It is who decides what happens next, and whether the system can act outside the chat window.

Three generations of website chat

Scripted chatbots

The first wave of website bots followed scripts. Someone drew a decision tree ("Are you an existing customer? Yes / No") or defined intents: a fixed list of things a visitor might want, each with sample phrasings. Google's Dialogflow documentation describes an intent as something that "categorizes an end-user's intention for one conversation turn," matched against training phrases that are "example phrases for what end-users might say" (Dialogflow ES docs). Tools like this use machine learning to recognize different wordings, but every answer and every branch of the conversation is still written by hand.

These bots are predictable and cheap to run. They are also brittle. When a visitor phrases a question in a way nobody anticipated, the bot falls back to "Sorry, I didn't get that," and every new topic means building another branch by hand.

LLM chatbots that answer from your content

The second wave replaced scripts with a language model. The bot looks up passages from your pages, help articles and PDFs that relate to the question, then writes an answer from them. This pattern is called retrieval-augmented generation, or RAG.

These bots handle open-ended questions well: "Do you integrate with HubSpot, and is it on the cheaper plan?" is no harder than "What are your hours?" But they still only talk. If a visitor says "book me a demo for Thursday," the best a pure answer bot can do is paste a link to your calendar page.

AI agents

The third wave gives the model a goal and a set of tools. An AI agent on a website might check free slots in a sales calendar, create a meeting, add the lead to a CRM, look up an order, or prepare a quote. Anthropic describes agents as "typically just LLMs using tools based on environmental feedback in a loop": the model takes a step, looks at what happened, and decides the next one.

The model inside can be the same kind of model as in the second wave. What changes is the loop around it and the permissions it has.

Side by side

The same visitor, three ways

Picture a visitor on a B2B software site at 9 p.m. They want to know what a team of 40 would pay and whether they can see a demo. The company and plan below are made up for illustration.

Scripted chatbot

Visitor: do you have pricing for 40 people? and can I see a demo
Bot:     I can help with: [Pricing] [Support] [Talk to sales]
Visitor: talk to sales
Bot:     Please leave your email and our team will get back to you.

LLM chatbot

Visitor: do you have pricing for 40 people? and can I see a demo
Bot:     For 40 people the Team plan fits; it covers up to 50 users.
         The current per-seat price is on our pricing page [link].
         You can book a demo here: [calendar link].

AI agent

Visitor: do you have pricing for 40 people? and can I see a demo
Agent:   For 40 people the Team plan fits; it covers up to 50 users.
         The current per-seat price is on our pricing page [link].
         I can book the demo now. Thursday at 10:00 or 14:00, your time?
Visitor: 14:00 works
Agent:   Done. You're booked for Thursday at 14:00 with a product
         specialist. The invite is on its way to the email you gave me.

The agent did not answer better than the chatbot. It finished the job. Whether that matters depends on how many visitors drop off between the link and a booked meeting, which is something you can measure before and after (see how to measure an AI agent on your website).

Where the line gets blurry

Real products sit on a spectrum, not in three boxes. Anthropic groups both workflows and agents under "agentic systems" and recommends "finding the simplest solution possible, and only increasing complexity when needed." A chatbot that books a meeting through one fixed, coded step is closer to a workflow than to an agent. That is often exactly what you want: a fixed booking step is predictable and easy to test.

Marketing labels add confusion. In a June 2025 press release, Gartner predicted that over 40% of agentic AI projects will be canceled by the end of 2027 because of rising costs, unclear business value or weak risk controls. The same release described "agent washing," where vendors rebrand existing assistants, chatbots and robotic process automation as agents, and estimated that only about 130 of the thousands of vendors claiming agentic AI offer the real thing. Treat the word "agent" on a product page as a claim to check, not as a feature.

What can go wrong

Wrong answers

Any system that writes free text can be confidently wrong. In Moffatt v. Air Canada (2024), the airline's website chatbot told a customer they could claim a bereavement fare retroactively, which contradicted the airline's own policy page. Air Canada argued it could not be held liable for what the chatbot said. The British Columbia Civil Resolution Tribunal called that "a remarkable submission," wrote that "It makes no difference whether the information comes from a static page or a chatbot," and ordered the airline to pay $812.02 (decision).

The lesson applies to chatbots and agents alike: you own what your website says, whoever or whatever says it. Keeping answers tied to your own pages is the subject of our guide to grounded answers.

Wrong actions

A system that can act can act wrongly: book the wrong time zone, create a duplicate CRM record, or promise a discount nobody approved. Anthropic warns that "the autonomous nature of agents means higher costs, and the potential for compounding errors." A small misreading in the first step can become a wrong action three steps later.

Cost and speed

Agents make more model calls per task, so they cost more and respond more slowly. In Anthropic's words, "agentic systems often trade latency and cost for better task performance." For a visitor asking about your opening hours, that trade is not worth making.

How to choose for your website

Start from the job, not the label.

Before you give any bot the ability to act, work through this checklist:

• List every action it can take and the worst outcome of each one going wrong.
• Require a confirmation step or a person for costly or irreversible actions such as refunds, discounts and cancellations.
• Give each tool the narrowest permission that works, and make read-only the default.
• Log every tool call next to the conversation that triggered it.
• Decide when it must step aside, and make human handoff easy for the visitor to request.
• Test with real questions from your inbox before launch, including ones it should decline.

Questions to ask a vendor

• Which steps does the model decide, and which are fixed in code?
• Which tools can it call, with what permissions, and can I switch each one off?
• Where do its answers come from, and does it show the page it used?
• What does it do when it doesn't know? Show me, live.
• Where can I see every conversation and every action it took?
• How do your reports count a "resolution" or a "conversion"?

The last question matters more than it looks. Vendors define success in different ways, and some definitions flatter the product.

The other agents: the ones visiting your site

"AI agent" has a second meaning that website owners now run into: software that browses on a person's behalf, reading pages, comparing vendors and filling in forms. These visitors don't open your chat widget. They read your pages directly.

Proposals such as WebMCP aim to let a website describe actions that such agents can call in a structured way. As of October 2026 it is a Draft Community Group Report from the W3C Web Machine Learning Community Group, and the draft itself says "It is not a W3C Standard nor is it on the W3C Standards Track" (WebMCP draft). For a plain-English overview, read what WebMCP is and why it matters.

Both kinds of agent, the one you put on your site and the ones that visit it, work best with the same raw material: accurate, current, clearly written pages.

Where to start

If you are adding AI to your website for the first time, start with a chatbot that answers from your own pages, links the page it used, and hands off to a person when it should. Measure what visitors ask and where they drop off. Add actions one at a time, where the numbers show a real gap between an answer and an outcome.

ASDK's Chat agent, in early access, is built for that first step: it answers visitors from your own content, cites your pages, and captures contacts.