AI agents, chatbots and automations are not the same thing

Three words get used interchangeably in almost every conversation we have about AI, and they describe three quite different things with different costs, different failure modes and different reasons to exist.

Getting them straight is the cheapest thing you can do before spending money, because most disappointing AI projects are really a case of buying one and needing another.

Automation: no intelligence required

An automation is a rule. When this happens, do that. A form is submitted, so a row is created, a message is sent and someone is notified.

There is no judgement involved and there does not need to be. The steps are the same every time, which is exactly why a computer should do them.

Most of what businesses call "AI projects" are automations, and that is good news. They are cheaper, they are predictable, and when they break they break loudly rather than quietly producing something plausible and wrong. Tools like n8n, Zapier and Make exist for this, and we wrote about choosing between them.

Use one when: the rule can be written down. If you can describe the process as a flowchart without the word "usually" appearing, you want an automation.

Chatbot: a conversation with a boundary

A chatbot answers questions. A good modern one answers them from your own material rather than from whatever the model absorbed during training, which is the difference between a useful support tool and a confident liar.

That is the part worth paying for. The technique is unglamorous: your documents get indexed, the relevant pieces get pulled in when someone asks something, and the model answers from those rather than from memory. It is often called RAG, and it is the single most reliable way to stop a chatbot inventing things.

A chatbot mostly talks. It can look things up and it can tell you what it found. What it generally cannot do is change anything.

Use one when: people keep asking the same questions and the answers already exist somewhere, in documents, a manual, or three years of support email.

Do not use one when: the honest answer to most questions is "it depends on your account", and the bot has no access to accounts. You will have built an expensive way to frustrate people.

AI agent: it can act

An agent can decide what to do and then do it. It has tools, it chooses which to use, and it acts on systems that matter.

Concretely, that is the difference between a bot telling a customer your refund policy and one that looks up the order, checks it against the policy, issues the refund and logs it.

That capability is the reason agents are worth building and the reason they need more care than either of the above. Something that can act can act wrongly. An agent that is confused does not sit there confused, it does something.

So the engineering in a real agent project is mostly not the clever part. It is the boundaries: what it is allowed to touch, what needs a person to approve it, what gets logged, and how it hands off when it is out of its depth. We treat that as the work rather than an add-on, because an agent nobody can inspect is a liability wearing a friendly interface.

Use one when: the task genuinely requires judgement, the judgement is repeatable, and you can define what "wrong" looks like clearly enough to catch it.

The cost ladder

The order matters, because it runs the same direction as both capability and risk.

Automation is cheapest to build, cheapest to run, and fails safely. Nothing happens, and you notice.

Chatbot costs more, runs on per-use pricing that scales with traffic, and fails by saying something wrong. Grounding it in your own content reduces that a lot.

Agent costs the most to build, mostly in guardrails rather than in the model, and fails by doing something wrong. Which is recoverable if you logged it and expensive if you did not.

Nobody sells you an automation when you ask about AI, because it is the least exciting answer. It is also frequently the correct one.

How to tell which you need

Describe the task as though you were training a new employee on their first day.

If your explanation is a list of steps in a fixed order, it is an automation.

If it is mostly "here is where to find the answer", it is a chatbot.

If it involves "use your judgement, and check with someone if it looks unusual", it is an agent, and the checking part is not optional.

If you find yourself saying "well, it depends" more than twice, the process is not ready to be automated at all. That is not a technology problem. Writing it down clearly is work worth doing whether or not you ever build anything.

Start smaller than you think

The projects that work start with one well-defined job, prove it on real data, and expand from there. The ones that stall try to automate a department.

There is also no rule that says you pick one. Most useful systems are an automation doing the predictable work, with a model handling the one step in the middle that genuinely needs reading comprehension.

If you want help working out which of the three you actually need, describe the task as it works today, including the parts still done by hand, and we will tell you honestly, even when the answer is the boring one. That is where every AI and automation project we take on begins.