AI agents vs chatbots: what's the difference?

A chatbot talks. An agent acts. A chatbot reads a message and writes a reply, and that reply is the end of it. An agent can read your diary, book a slot, draft an invoice and file it, choosing which step comes next as it goes.

Anthropic, the company behind Claude, draws a useful line in its engineering guide, building effective agents. In a workflow, the AI and its tools follow steps written out in advance. In an agent, the AI directs its own process and decides which tools to use. That middle category is the one most sales pitches skip.

A chatbot answers

The box on a website that answers questions about opening hours, prices and delivery. It is useful when the same twenty questions come in every week. It can only say things; it can't change anything in your business.

A workflow follows your steps

Every new enquiry email gets read, sorted into quote, booking or complaint, and a draft reply is written in the house style. The AI does the reading and writing, but the steps never change. You decided them.

An agent decides the steps

Ask an agent to "sort out the bookings with missing details" and it works out how: checks the diary, finds the gaps, drafts the messages, notes what it couldn't fix. More capable, and harder to predict.

What the AI companies mean by agent

OpenAI's developer documentation describes agents as AI that can plan and complete tasks using tools, work with other agents and keep track of context across steps, in its guide to agents. The word that matters is tools: an agent is only as capable, and only as risky, as what it is connected to.

How agents connect to your tools

Many AI tools now connect to other apps through the Model Context Protocol, an open standard its makers compare to a USB-C port for AI, letting assistants such as Claude or ChatGPT reach calendars, files and databases. Its introduction explains the idea. Every connection you add is a new thing the agent can change.

Why the words get blurred

Suppliers now call almost anything with AI in it an agent, including plain chatbots. Ignore the label. Ask what the thing can change in your business, and who decides the steps. Those two answers tell you which of the three you are buying.

Start with the simplest thing that does the job. Anthropic gives builders the same advice: find the simplest solution possible and add complexity only when it is needed. For most owner-run businesses, that means a chatbot or a workflow first, and an agent only for the jobs where the steps genuinely vary. The UK government's AI Playbook goes one step further: choose the right tool for the job, which may not be AI at all.

Choose a chatbot when the job is answering

A café getting the same questions about allergens, opening times and bookings. The answers are written down already; a chatbot just gives them out at midnight when nobody is behind the counter. Allergen answers still need care, so keep them to what is written and send anything unclear to a person.

Choose a workflow when the steps never change

A bookkeeper turning receipts into entries, the same way every time. A fixed workflow is easier to check and easier to fix than an agent, because you know exactly which step went wrong. Most paperwork is this shape, which is why Document Automation is built as a workflow.

Choose an agent when the steps vary

A venue checking its listings across several sites overnight, where each site needs something different. The agent works out what to do on each one, and leaves a report of what it did and what it refused to do.

Be wary of agents sold as magic

If a supplier can't tell you exactly what the agent may touch, what it may never do and how you will see what it did, the agent isn't ready for your business. Ask those three questions before you sign, and ask to see a real action log from a live job.

The same business can need all three, for different jobs. Here is how a choice plays out for three made-up but typical firms.

A dog groomer: a chatbot

Most messages ask about prices by breed, opening times and whether there is a slot this week. A chatbot with the price list and opening hours answers the first two at any hour, and passes slot questions to the owner. It changes nothing, so the risk is a wrong answer, which good facts and a clear brief keep rare. How to write instructions an AI will follow covers that brief.

A small accountancy practice: a workflow

Clients send receipts by email. A workflow reads each one, pulls out the date, supplier and amount, files it in the right client folder and adds a line to a sheet for checking. The steps never change. Example figures: if each receipt takes two minutes by hand and the practice gets 300 a month, that is ten hours of typing turned into a shorter job of checking.

A lettings agency: an agent, on a short lead

The agency lists properties on several sites, and each night someone should check that prices, photos and availability match the agency's own records. The steps differ by site and by problem, so an agent fits. It may read every listing and draft corrections, but it may not publish them. The agent leaves a morning report of what it found, what it drafted and what it couldn't check.

An agent can take a wrong action as well as give a wrong answer, and it can take several before anyone notices. The fixes are limits on its tools, a person approving anything final, and a record of everything it did. Four risks to plan for.

One mistake leads to the next

An agent builds each step on the last. If it misreads the diary at step one, every step after inherits the error. Anthropic recommends extensive testing in sandboxed environments, meaning a safe copy where mistakes don't matter, with guardrails in place, before an agent works on anything real.

It can be tricked by what it reads

An agent that reads emails or web pages can meet text written to give it orders, known as prompt injection. The UK's National Cyber Security Centre says this may never be fully fixed, and advises against letting an AI that reads outside emails use powerful tools, in prompt injection is not SQL injection.

It has more access than the job needs

The security community OWASP lists excessive agency as a top risk in AI apps: too many tools, too many permissions or too much freedom. Its advice includes keeping a person in the loop to approve high-impact actions.

It makes decisions about people

If an agent would decide on its own whether someone gets a tenancy, a job or credit, UK data protection law has rules for that. The ICO's guidance on automated decision-making covers decisions made solely by automated means with a legal or similarly significant effect. For an owner-run business, the simple answer is to keep a person making those calls, as which AI jobs still need a person to sign off explains.

Whichever you choose, start small, on one job, with a pass mark agreed in advance. A chatbot or workflow can go live in days; an agent should earn its freedom over weeks.

Run it in the shadows first

Let it do the job alongside you for two weeks without anything reaching a customer. Compare its work with yours. How to trial AI on one job gives a simple pass or fail bar.

Give an agent drafts before actions

Even when an agent is the right choice, start it on drafting and reporting. Let it act alone on one routine kind of task only after weeks of drafts you didn't need to change.

Wrap it in a harness

Whatever the label, the set-up around the AI decides whether it is safe and useful: the brief, the tool limits, the checks and the habit of turning mistakes into rules. That is what harness engineering is, and it matters more for an agent than for anything else.