AI agents vs chatbots vs automation: what's the difference?

A customer emails at 11 p.m.: "Where's my order? It was supposed to arrive Tuesday." A chatbot replies with your shipping policy. A workflow automation spots the word "order", looks up the tracking number and sends the standard status email, which says "in transit", as it has for nine days. An AI agent looks up the order, sees the parcel hasn't moved since it left your warehouse, checks the courier's site, finds a failed delivery attempt to a mistyped address, drafts an apology with a fix, and asks you before it sends. That's the AI agent vs chatbot question in one scene, with automation in the middle. This guide explains all three, compares them side by side and helps you pick.

Quick definitions

Chatbot: software that holds a conversation. It answers; it doesn't usually act.
Workflow automation: fixed steps that run between apps when something happens, built in tools like Zapier or Make.
RPA (robotic process automation): software robots that copy what a person does on screen, clicks and keystrokes, in tools like UiPath.
AI agent: software given a goal that decides its own steps, takes them with tools, checks the results and adjusts. (Full explanation: what is an AI agent?)

What is a chatbot?

Chatbots are older than most people think. ELIZA, written by Joseph Weizenbaum at MIT and described in 1966, matched keywords in what you typed and turned them into replies. Its most famous script played a psychotherapist, often by turning your statements back into questions.

For decades afterwards, business chatbots worked much the same way: keyword rules and decision trees. Press 1 for billing. Type "refund" and get the refund paragraph. They were predictable and cheap, and they frustrated anyone whose question wasn't in the tree.

Large language models changed the conversation part. A modern chatbot understands questions phrased a hundred ways and answers in fluent, natural language, often from your own documents. But by default it still only talks. Nothing outside the conversation changes. OpenAI's guide for agent builders makes the line explicit: apps that use an LLM without letting it control the work, "think simple chatbots," are not agents.

Talking still has consequences. In 2024 a Canadian tribunal held Air Canada responsible after its website chatbot told a customer he could claim a bereavement discount after travelling, which the airline's policy didn't allow. The airline argued the chatbot was "a separate legal entity"; the tribunal answered that it was "still just a part of Air Canada's website." Whatever your bot says, you said.

What is workflow automation? (Zapier, Make)

Workflow automation connects apps with fixed rules. In Zapier, a Zap is "an automated workflow that connects your apps and services together." It starts with a trigger ("an event that starts a Zap", such as a new form submission) followed by one or more actions ("an event a Zap performs after it is triggered", such as adding a row to a sheet and posting in Slack). Zapier says it connects more than 9,000 apps.

Make works on the same idea with a visual canvas. A scenario is "a series of modules that indicate how data should be transferred and transformed between apps/services." Make advertises more than 3,000 ready-made app integrations.

The strength of classic automation is that it's deterministic: the same trigger produces the same steps every time. That makes it cheap, fast and easy to trust for high-volume, identical work. The weakness is the flip side. It does exactly what you designed, even when the situation calls for something else, like our order email cheerfully saying "in transit" on day nine.

Both tools now offer AI agents too. Zapier describes agents that plan, adapt and work across tools toward a goal, with human review as needed. Make's agents let you "set goals, provide tools, define limits – and see the reasoning." The categories are blurring, which is exactly why the definitions above matter.

What is RPA? (UiPath)

Robotic process automation is automation for the screen. UiPath, one of the best-known vendors, says RPA "mimics human actions in digital systems": software robots log in, click, copy and paste in applications just as a person would. That makes it useful for older systems with no API, where Zapier-style connectors can't reach.

UiPath says RPA "excels at automating high-volume, repetitive, rule-based tasks." It's built to follow the steps it was given; when a case doesn't fit the rules, it needs a person or a new rule. UiPath now positions RPA as part of what it calls agentic automation: AI agents decide what to do, and RPA robots carry out the steps across enterprise systems.

AI agent vs chatbot: the core difference

Strip away the marketing and two questions separate them:

The same job three ways: a chatbot tells you how, automation runs fixed steps, an agent works it out and asks first.

Who decides the next step? In a chatbot, the next step is always "reply". In automation and RPA, a person decided every step in advance. In an agent, the model decides at run time, within limits you set. Anthropic draws exactly this line: workflows follow "predefined code paths", while agents "dynamically direct their own processes and tool usage."
Can it change things outside the conversation? A chatbot produces words. An agent produces actions: an email sent, a record updated, a server restarted.

That's also how to spot what Gartner calls agent washing: chatbots and assistants rebranded as agents without real agentic capabilities. Gartner estimated in 2025 that only about 130 of the thousands of vendors selling agentic AI were the real thing. If a product can't take a step it wasn't scripted to take, it's not an agent, whatever the label says.

Side by side: chatbot vs automation vs RPA vs AI agent

	Chatbot	Workflow automation (Zapier, Make)	RPA (UiPath)	AI agent
What it does	Answers in a conversation	Runs fixed steps between apps when a trigger fires	Repeats a person's clicks and keystrokes in apps	Works toward a goal, choosing and taking steps with tools
Who decides the steps	Rules or a model, for words only	You, in advance	You, in advance	The model, as it goes, within your limits
When something unexpected happens	Answers within what it knows	Follows the path anyway, or stops	Stops, or hands the case to a person	Notices, re-plans, or asks you
Takes actions	Usually not	Yes, the same ones every time	Yes, the same ones every time	Yes, chosen for the situation
Reaches	Text and voice	Apps with connectors or APIs	Anything with a screen, including old systems	APIs, browsers, files, screens
Predictability	High (rules) to medium (LLM)	Very high	High	Lower: same goal, sometimes a different path
Cost per run	Low	Low	Low per run, more to set up	Higher: many model calls per job
Best for	FAQs and first-line answers	High-volume, identical, rule-based jobs	Repetitive work in systems without APIs	Varied work that needs judgment and messy inputs
Main risk	Wrong answers	Doing the wrong thing perfectly	Breaking when the process doesn't fit the rules	Wrong actions, if it isn't limited

When to use which

Use a chatbot when the job is answering: product questions, opening hours, "do you deliver to my area?". Keep it to what your knowledge covers and make handing over to a person easy. If that's your need, see how to add AI chat to your website and our roundup of the best AI chatbots for websites.

Use workflow automation when the steps are always the same and the inputs are clean: copy every new order into the accounting app, post every new lead in the sales channel. It's cheaper and more predictable than any agent. Microsoft's agent documentation says it bluntly: "If you can write a function to handle the task, do that instead of using an AI agent."

Use RPA when the steps are fixed but the system has no API and someone is currently doing it by hand, screen by screen.

Use an AI agent when the work varies from case to case. OpenAI's guide names three signs: decisions that need judgment and exceptions, rulebooks that have grown too tangled to maintain, and heavy reliance on unstructured information like emails and documents. Our late-order email is all three. For worked business examples, see AI agents for small business.

A useful rule of thumb: automate what's identical, give agents what's similar, and keep people on what's unique.

They work best together

In practice the strongest setups mix all of them, each on what it's good at:

Automation triggers an agent. A new order arrives and a webhook hands it to an agent, which checks it, spots the odd one and asks you.
An agent uses automation. Once it decides what to do, it calls an existing workflow to do it the approved way.
A chatbot hands over. The website bot answers what it can, and passes everything else to a person or an internal agent with the conversation attached.

Here's how that looks in crew bot. The website support chat is deliberately a chatbot: one line on your site, a bot per site that answers from what your site and team say, and hands the visitor to your team when it can't help or they ask for a person. Because it talks to strangers, it has no tools at all, so whatever a visitor types can at worst produce an odd reply, never an action. Behind it, the crew is a set of agents: you hand a job to the Lead, it works your email, apps and servers, and it asks before anything that writes. Routines run jobs on a schedule and webhooks let your other systems hand it work, the automation part. Different jobs, different tools, one place.

Questions to ask before you buy

Does it act, or only answer? Ask for a demo where it changes something in another system.
Who decides the steps? If every step was configured in advance, it's automation, which may be exactly what you need.
What happens when the case doesn't fit? Watch it handle an odd one.
What needs my approval? An agent should ask before actions that are hard to undo. See human in the loop AI.
What does a job cost, and is there a cap? Agents make many model calls; a per-job budget matters.

FAQ

Is ChatGPT a chatbot or an AI agent?

Both, depending on how you use it. In an ordinary conversation ChatGPT is a chatbot: it answers and nothing outside the chat changes. Its agent features, which work on a virtual computer with a browser and take actions toward a goal, make it an agent for those tasks.

Is Zapier an AI agent?

Classic Zaps are workflow automation, not agents: a trigger and a fixed set of actions you designed. Zapier also now offers AI agents that plan and adapt toward a goal, and you can combine the two, for example a fixed workflow with one step handed to an agent.

Will AI agents replace RPA and workflow automation?

Unlikely for the jobs automation does best. For identical, high-volume steps, fixed automation is cheaper, faster and more predictable than an agent. What's changing is the edges: agents handle the exceptions and messy inputs that used to fall to people, and vendors like UiPath, Zapier and Make are adding agents next to their automation.

What's the difference between an AI agent and an AI assistant?

The words overlap, but an assistant usually helps you while you work (drafting, summarising, answering) and waits for your next request. An agent is handed a goal and works through the steps itself, using tools, coming back to you when it's done or needs a decision.

Can a chatbot become an AI agent?

Yes, by giving it tools, a goal and a loop. But for a bot that talks to the public, that's often a bad idea: anyone can type anything into it. A safer pattern is a chatbot with no tools facing customers, handing anything that needs action to a person or to an agent that works under your approval.