On this page
Key takeaways
- An AI agent takes a goal, plans the steps and uses tools to act. It does more than reply.
- Agents are strongest on varied, multi-step work; fixed processes are often better as plain automation.
- Scoped access, approvals and audit trails are what make agents safe in production.
A definition you can use
An AI agent is software that is given a goal, decides the steps needed to reach it and uses tools such as search, databases, email and your CRM to carry those steps out. Where a chatbot produces an answer, an agent produces an outcome.
How an agent works
- It receives a goal: 'Resolve this customer's delivery complaint.'
- It plans: look up the order, check the courier, check the refund policy, reply.
- It acts using tools it has been given access to.
- It checks the result and adjusts if a step failed.
- It pauses for approval when an action crosses a limit you set, and logs everything.
Agent vs chatbot vs automation
| Chatbot | Automation | AI agent | |
|---|---|---|---|
| Main output | An answer | A fixed sequence of actions | An outcome |
| Handles variation | In conversation only | Poorly | Well |
| Uses your tools | Rarely | Yes, as scripted | Yes, as needed |
| Best for | FAQs | Stable, repetitive processes | Varied, multi-step requests |
A worked business example: inputs, tools, actions and limits
An agent is easiest to understand through one concrete job. This example is an order-support agent for a small online shop. It is an illustration of the design, not a client deployment.
| Part | In this example |
|---|---|
| Goal | Resolve a customer's message about an existing order |
| Inputs | The customer's email, the order record, the courier tracking status, the refund policy |
| Permitted tools | Read orders, read tracking, draft and send replies, create a refund request |
| Not permitted | Changing prices, editing customer accounts, deleting anything, refunds without approval above the limit |
| Typical actions | Look up the order, check tracking, reply with an accurate update, propose a refund or resend when a parcel is lost |
| Approval | Refunds above a set amount and any message to a key account go to a named person |
| Stop conditions | No matching order, an angry or legal complaint, low confidence, or the same customer writing a third time: hand over to a person with the context |
| Record | Every step, tool call, decision, approver and outcome written to a log |
The goal makes it an agent rather than a script: the steps differ for a late parcel, a wrong item or a cancellation, and the agent chooses them. The tool list, the approval rule and the stop conditions are what make it safe to run.
Examples by team
- Support: resolve order, refund and account requests end to end.
- Sales: research new leads, score them and draft follow-ups for a rep to approve.
- Operations: onboard a new customer across several systems.
- Internal helpdesk: answer policy questions from your own documents, with sources.
When not to use an agent
If a process is the same every time and the data is already structured, a plain automation is cheaper, faster and easier to test. Agents earn their place where requests vary and judgement is needed.
Controls that make agents production-ready
- A dedicated, least-privilege identity for each agent.
- Approval gates on money, deletions and sensitive messages.
- An audit trail of the plan, the tools called and the result.
- A clear handover to a person when the agent is unsure.
Frequently asked questions
Yes. That is the main reason to use one. Each tool connection is scoped to what the agent needs.