Most businesses that say they want an AI agent want an automation, and would be better served by one. The three things get sold interchangeably and they are not interchangeable at all.
| Automation | Chatbot | AI agent | |
|---|---|---|---|
| What it does | Executes a fixed rule | Responds to messages | Pursues a goal across steps |
| Who decides the steps | You, in advance | You, in advance | The system, at run time |
| Handles the unexpected | No — it stops or errors | Poorly — falls back to a script | Yes — sometimes badly |
| Supervision needed | Check the rule once | Check the script | Check every outcome |
| Cost to build | Low | Low to moderate | Moderate to high |
| Best for | Predictable, repeated tasks | Answering known questions | Multi-step work with variation |
When an automation is the right answer
If you can write down every step, use an automation. It will be cheaper, faster, more reliable and easier to debug than anything with a model in it.
Invoice arrives, extract the total, post to the ledger, flag anything over a threshold. That is a rule. Adding intelligence to it adds cost and failure modes without adding capability.
The test: if you can describe the task as a flowchart with no "it depends" boxes, you want an automation.
When a chatbot is the right answer
If the job is answering questions people already ask, and the answers exist somewhere, a chatbot handles it. Opening hours, service descriptions, order status.
Chatbots fail when the question falls outside the anticipated set, which is why they have a reputation for frustration. That is not a technology problem; it is a scoping problem.
When you actually need an agent
Agents earn their cost when the task has genuine variation. Not "it depends" as an excuse for vagueness, but real branching that you cannot enumerate.
An inbound call is the clearest example. The caller might want to book, might have a question, might be a supplier, might be selling something. You cannot write that flowchart, because you do not know what the call is until it happens.
The test: if a competent new employee would need judgement rather than a checklist, an agent is a candidate.
The mistake almost everyone makes
Buying agent capability for automation work. It is more expensive, less predictable, and harder to verify — and it fails in ways a rule never would, because a rule cannot be creative about how it is wrong.
The reverse mistake is rarer but worse: forcing genuinely variable work through a rigid automation, then discovering the exceptions are handled by a person doing manual cleanup nobody counted.
How to decide in ten minutes
Write down the task. Then try to write down every step. If you finish the list, you want an automation. If you find yourself writing "then it depends on what they say," you may want an agent — and the next question is whether you can state what a correct outcome looks like. If you cannot, you are not ready to automate it at all.
More on where to begin in How to Choose Your First AI Automation.