What is the concrete difference between a chatbot, an automation and an agent?
All three get sold together and they are not the same thing.
| Chatbot | Classic automation | AI agent | |
|---|---|---|---|
| What it does | Returns predefined information | Runs a fixed sequence of steps | Decides what to do, then does it |
| How it decides | Decision tree or keyword match | Rules written up front | Business criteria applied per case |
| On the unexpected | Breaks or repeats the menu | Fails and halts the process | Evaluates and escalates if needed |
| System access | None, or read-only | Whatever was programmed | Read and write, through tools |
| Use it when | FAQs are stable | Processes are identical, high volume | Volume and nuance both exist |
| Do not use it when | The user needs something resolved | Every case differs | Criteria are not stable |
The distinction that matters is not technical: it is whether the process needs judgement. If every case is identical, a classic automation is cheaper, faster and easier to audit — and recommending it is the honest answer. Only when cases vary, and that variation is why they currently go to a person, does the agent case appear.
Why did earlier chatbots disappoint?
Because they did not understand — they matched. Behind most of them sat a decision tree and a list of keywords. As long as the user asked the anticipated question, it worked. The moment they went off-script, which is nearly always, it returned the menu.
Two things changed. Current models grasp intent even when the question is badly phrased or mixes two topics. More importantly, a model can now use tools: query a database, write a record, call an API. That is what turns a system that converses into a system that resolves.
What three pieces does an agent need to be useful?
Without all three, it is an expensive chatbot.
- Access to the data. If it cannot see stock, order status or customer history, all it can do is improvise polite answers.
- Tools to act with. Book, record, generate, notify. An agent that only informs leaves the work exactly where it was.
- Business criteria, written down and adjustable. When to offer a discount, when to escalate, what never to do alone. This is not inferred from the model — it is written, reviewed, and changed when the business changes.
The third is the one almost nobody gets right, and it decides whether the project is worth anything. It is also the part that requires understanding the operation rather than knowing how to code.
What can an agent do, by function?
- Sales. Qualifies an inbound lead, answers product questions, books, and hands to a salesperson at the moment the conversation warrants it rather than 48 hours later.
- Purchasing. Reviews turnover, stockouts and lead time and proposes the order with the reasoning for each line.
- Admin. Reconciles movements, organises documentation, and flags specifically what does not balance — the only case that needs a person.
- Analysis. Query the database in plain language and answer a business question without waiting for someone to build a report.
What happens when the agent gets it wrong?
Very few vendors write this section, and it is the one that matters to whoever signs.
An agent in production needs four things: explicit limits on what it can execute without confirmation; traceability, a record of what it did, with which data, and why; human escalation on a clear rule, not only when the model is unsure but when the case exceeds a threshold of value or risk; and measurement, so you know what share it resolved alone and where it failed.
If a vendor cannot explain all four, what they are about to sell you is a demo.
When is an agent NOT the right answer?
- Low-volume processes. At ten cases a month, configuring and maintaining it costs more than doing it by hand.
- Processes without stable criteria. If the rule changes weekly, you will spend more time adjusting the agent than running the process. Stabilise the process first.
- When errors are expensive and unreviewable. If a mistake means a wrong payment or a legal commitment and nobody reviews before it lands, the agent is not the tool — or not without a confirmation step that may erase the saving.
Saying this does not cost us work. It avoids projects that were going to fail.