AI guide
What is an AI agent?
An AI agent decides a sequence of steps and calls tools to carry them out. A useful business agent has a narrow job, a short list of tools, and a person responsible for anything that is hard to undo.
Agent, chatbot, assistant
A chatbot holds a conversation, usually with a customer. An assistant helps a signed-in employee beside a system they already use. An agent is defined by action: it may look up, draft, and, if allowed, change something. The labels get mixed in sales material. In a design they should not be mixed, because a customer conversation and a tool that edits records have different risks.
The loop
The model is given a goal, the tools it may call, and the results of those calls. It chooses the next step until it stops or it is stopped. The tools are ordinary functions you wrote. Authorization sits in those functions. A sentence in the prompt that says “do not refund more than the policy allows” is not a control. The refund function either checks the rule or it does not exist.
Logs matter. You should be able to see which tool ran, with which inputs, and what it returned. Without that, you cannot explain a bad morning.
A job small enough to trust
Prepare a renewal brief from three records and wait. Classify inbound mail and draft a reply for a queue. Chase missing fields on an application and stop when a person must decide. Those are agents. “Handle the customer” is not. The agent development page is the delivery description. Start with reads. Add one write with approval. Expand only when the log is something you would show an auditor.
An agent is a task with tools
Call it an agent when
- It takes a defined task
- It may use named tools
- It leaves a log
Do not call it an agent when
- It only chats
- It may do anything
- Nobody can see the steps
Bound it before you build it
Write access
Reading is not the same power as changing a record.
The stop
Low confidence should wait for a person.
The audit
If you cannot explain a step, the agent is not ready.
Bound the word before you use it
- 01
One task
Write the job in a sentence.
- 02
The tools
Read, draft, or write.
- 03
A stop
The case where it must wait.
- 04
A log
Who will read the steps.
Marks the agent explainer has a boundary
- 01
A stop is defined
The article says an agent needs a condition that ends the run.
- 02
Tools are limited
It can only take actions that were allowed.
- 03
A person approves
Money, customer messages, and deletes are not left unattended.
- 04
Not a renamed chatbot
The difference from a bot that only replies is explicit.
Related reading
Questions we hear
Do agents replace staff?
They take preparation and lookup off a person’s plate. Someone still owns the outcome. If a vendor cannot name that person, do not buy the agent.
Are multi-agent systems better?
Sometimes they are a way to avoid specifying the job. One agent with three tools is easier to test than four agents talking to each other. Add a second only when the first has a measured task.
Is this safe with customer data?
Only if each tool enforces the same permissions the user has, and the transcript retention is deliberate. The secure software page describes the surrounding habits. The agent does not get a special pass.