The difference in one table
| Aspect | Chatbot | AI agent |
|---|---|---|
| Main job | Answer questions | Complete tasks |
| Access | A knowledge base or FAQ | Your systems, through defined tools |
| Output | Text in a conversation | Actions taken and records updated |
| Main risk | A wrong answer | A wrong action |
| Oversight | Review conversations | Approvals, limits and action logs |
| Set-up effort | Lower | Higher, because of integration and testing |
| Typical example | Website FAQ assistant | Invoice intake, enquiry triage, document checking |
When a chatbot is enough
If the goal is to help people find information you already have, such as opening hours, policies or product details, a well-grounded chatbot with citations is often all you need. It is simpler, cheaper and easier to keep safe.
When you need an agent
- The work involves reading something, deciding and then updating a system.
- The task has several steps across more than one tool.
- People spend their day moving information rather than using judgement.
- The result must be a changed record, not just a reply.
Three illustrative examples
These are examples of the pattern, not descriptions of client work.
- Enquiry triage. An agent reads incoming enquiries, classifies them, drafts a reply from your own guidance and creates the CRM record. A person approves anything unusual.
- Supplier invoice intake. An agent extracts the details from invoices, checks them against purchase orders, posts matches to your accounts and flags exceptions for review.
- Internal policy assistant. Closer to a chatbot: it answers staff questions from your documents and cites the source, with no ability to change anything.
The guardrails an agent needs
Non-negotiables
- Least-privilege access: the agent can touch only what the task requires.
- Approval for sensitive or irreversible actions.
- A complete log of what it did and why.
- Tested behaviour when it is unsure or a system is unavailable.
- Monitoring of quality and cost after launch.
Where to start
Pick one process, measure it today and build the smallest agent that handles it with a person in the loop. Our guide to what agent projects cost explains how to scope that, and our AI agent development page shows what we build.