Chapter 01 · 5 min read
What an AI worker is, and when not to build one
The difference between a chatbot, a workflow and a worker, and the four questions that tell you whether you need one.
Most businesses meet AI through a chat window: you type, it answers. An AI worker is something else. It has a job, it can act in your systems, and it is judged on whether the job got done, not on how pleasant the conversation was.
Chatbot, workflow, worker
The three are often mixed up, and the difference decides how much you build.
| What it does | Who decides the steps | Example | |
|---|---|---|---|
| Chatbot | Answers questions from what it was told | Nobody: one question, one answer | "What are your opening hours?" |
| Workflow | Runs fixed steps, with the model doing one part | Your code | Every new invoice: read it, extract the totals, file them |
| AI worker (agent) | Works towards a goal, choosing which tools to use and in what order | The model, within the limits you set | "Is my phone ready, and can I pick it up tomorrow?" (it checks the job, checks the calendar, books, confirms) |
A worker is the most capable of the three, and also the most expensive to run and the hardest to test. Start with the simplest one that does the job.
The four questions
Before building a worker, answer these. If any answer is no, a single call or a workflow is the better choice.
- Is the task genuinely multi-step and hard to specify in advance? "Turn this support ticket into a resolved case" is. "Extract the order number from this email" is not.
- Is the outcome worth the cost and the extra seconds? A worker makes several model calls per task.
- Can Claude actually do it? Try the task by hand in a chat first. If it cannot do it with a person feeding it the information, it will not do it alone.
- Can a mistake be caught and undone? Refunds, deletions and messages to customers need checks, approvals or a way back.
What a worker is made of
Every worker in this guide has the same five parts. The chapters follow them in order.
- A model that reads, reasons and decides.
- A job description (the system prompt): its role, its limits and when to hand over.
- Tools: functions it can call to read and change things in your systems.
- A loop that runs the tools it asks for and feeds the results back until the job is done.
- Guardrails: approvals, logging and tests, so you can trust what it does when you are not watching.
A note on expectations
A good worker does not replace a team. It takes the repetitive part of a job (the lookups, the first reply, the booking, the summary), does it consistently at any hour, and passes the unusual cases to a person with the context already gathered. That is where the time is saved, and it is also what makes customers trust it.