Vendors use “chatbot” and “AI agent” loosely, and sometimes for the same product. The distinction matters because it changes what the system is allowed to do in your business, what it costs to build, and what happens when it gets something wrong.
There are really three things to tell apart, not two.
Three definitions
A chatbot is a conversation. Someone types a question and gets a reply, from a script, from your help pages, or from a language model. The output is text. Whatever the customer needs to happen next, they or your team still have to do.
A workflow is a fixed sequence of steps that may include AI. A form arrives, a model extracts the details, rules check them, a draft is prepared, a record is created. The steps are decided in advance by whoever built it. Anthropic’s engineering guide Building effective agents defines workflows as systems where language models and tools “are orchestrated through predefined code paths.”
An agent is given a goal and tools, and decides for itself which steps to take and in what order. The same guide describes agents as systems where models “dynamically direct their own processes and tool usage.” That flexibility is the point, and also the risk. The guide notes that agents trade cost and speed for performance and can compound errors, and it recommends using the simplest approach that works, which sometimes means no agent at all.
Most of what small and mid-sized businesses actually need is the second one. It gets called an agent because the word sells.
One enquiry, three ways
A fictional example. A request for a quote lands on a joinery company’s website at 21:00.
The chatbot answers the visitor’s questions about materials and lead times from the company’s FAQ, then suggests they fill in the form. Useful, and the request still waits until morning.
The workflow reads the form, pulls out the dimensions, wood type and location, checks them against the services the company offers, and prepares a draft reply with a price range from the approved price list. It creates a CRM record and puts the draft in front of the sales lead in the morning. One click to send, or edit first.
The agent is told “qualify this lead and book a call if it fits.” It decides to email the visitor a follow-up question about the room, reads the answer, checks the calendar and books a slot. More capable, and much harder to predict. Every action it might take needs a permission you’ve deliberately granted.
How to choose
Ask these in order.
- Is the output an answer? Then a chatbot may be enough. Give it a knowledge source you control and a clear way to hand over to a person.
- Are the steps the same every time? Then build a workflow. It’s easier to test, cheaper to run, and you can see exactly what it did.
- Do the steps genuinely vary case by case, across several tools? Then an agent might earn its complexity. Limit what it can touch and set stopping points.
- What happens if it’s wrong? This decides where a person approves. A wrong answer about opening hours is an annoyance. A wrong price sent to a customer, or a booking that clashes, costs money and trust.
For decisions about people, the GDPR sets a legal floor. Article 22 gives people the right not to be subject to a decision based solely on automated processing when it has legal or similarly significant effects on them. In practice, keep a person responsible for anything like that.
Rules that apply to all three
Tell people when they’re talking to AI. Under the EU AI Act, systems that interact directly with people must be designed so those people are informed they’re dealing with AI, unless it’s obvious. The European Commission’s overview says the transparency rules come into effect in August 2026.
Beyond that, a few things I’d insist on whichever type you pick:
- It answers from information you’ve approved, not from whatever the model happens to know.
- It has the fewest permissions the job needs.
- Every action is logged, so someone can see what happened and why.
- Missing or conflicting information makes it stop and ask, rather than guess.
- You can switch it off without breaking anything else.
Questions to ask a vendor
When someone pitches you an “AI agent”, ask what actions it can take on its own, in which systems, and with what permissions. Ask where it stops for a human, what it records, and what it does when the information it needs isn’t there. If the answers are vague, you’re probably being sold a chatbot with a bigger price.
For more on picking a first task, see what AI automation can do for a small business.
Our AI workflows are built around that review step: the system prepares the work, a named person approves it. If you have a repetitive process in mind, describe it to us and we’ll tell you plainly whether it needs an agent, a workflow or just a well-fed chatbot.