Three products are routinely sold under one word, and the difference between them decides your cost, your reliability and how long the thing survives contact with reality.
The actual distinction
| Automation | Chatbot | AI agent | |
|---|---|---|---|
| Decides its own steps | No | No | Yes |
| Handles unforeseen cases | No | Poorly | Sometimes |
| Same input, same output | Always | Usually | Not guaranteed |
| Cost to run | Lowest | Low | Highest |
| Debugging when wrong | Straightforward | Straightforward | Hard |
Read that table in one direction and the guidance falls out: prefer the leftmost option that can do the job.
When an automation is the right answer
If you can write the process as a flowchart and the flowchart is finished, you want an automation. A form arrives, so create the record, notify the owner, send the acknowledgement. There is no judgement in that. Adding a language model to it makes it slower, more expensive and less predictable in exchange for nothing.
This covers far more real business processes than the current conversation suggests.
When a chatbot is the right answer
When people need answers, in their own words, about things you already know. Opening hours, order status, what is included, whether you cover their area.
The critical design decision is grounding: answers must come from your systems rather than from the model's memory. A chatbot answering from general knowledge will state your delivery radius confidently and incorrectly.
When an agent earns its cost
When the next step depends on what was found in the last one. Qualifying an enquiry is a genuine example — read the message, look the company up, decide it is out of scope, close it politely. Writing branches for every combination is not practical; deciding case by case is what an agent is for.
The cost is that flexibility and unpredictability are the same property. A system that can handle a case you did not anticipate can also handle it wrongly, which is why agents need logging, guardrails, and a human in the loop for anything irreversible.
A test that works
Ask: can I write down every rule?
If yes, automation. If yes but people will ask in a hundred different ways, chatbot. If no, because the right action depends on what turns up, agent.
Then ask a second question: what does it cost when this is wrong? The higher that number, the further left you should sit, and the more checking you should build in.
Why this goes wrong so often
Because "agent" sells. It is the current word, and a proposal built around it sounds more ambitious than one built around a workflow tool and four API calls — even when the second describes what will actually be delivered.
The practical defence is to ask any vendor which of the three they are proposing and why the simpler option was ruled out. A good answer is specific about the judgement the task requires. A vague one usually means nobody asked.
Questions
- Can one system be more than one of these?
- Most real systems are. A typical deployment is mostly ordinary automation with one or two AI steps inside it where judgement is genuinely needed. That is a sign of good design, not a compromise.
- Is an agent always more expensive to run?
- Per task, generally yes — it reasons before acting, and reasoning costs. Whether it is more expensive overall depends on what it replaces. The comparison to make is against the labour it removes, not against a simpler automation that could not do the job.
- How do we start without over-committing?
- Pick one process, define the outcome you want moved, and build the simplest thing that could move it. A narrow system that works is a better foundation than a broad one that mostly works.
Want this working on your numbers rather than in an article? We scope it on one call.
Book a Consultation Call