What is an AI agent, and how is it different from a chatbot?
An AI agent is software that pursues a goal by observing its environment, using tools and deciding step by step. Unlike a chatbot it doesn't only produce text: it sends email, creates records, reads files. It helps most with work whose rules are clear but whose steps are tedious. Because it acts, its permissions must stay narrow, critical steps need human approval, and every step must be logged.
- An agent doesn't just talk — it uses tools
- The narrower the permissions, the smaller the risk
- Human approval is mandatory for irreversible steps
- Every action should be logged and reversible
What is an agent?
The idea of an agent is not new: an intelligent agent perceives its environment and acts on it to reach a goal. With large language models this became practical: the model takes a goal, decides which tool to call, sees the result and plans the next step.
How it differs from a chatbot
| Chatbot | Agent | |
|---|---|---|
| Output | Text | Text plus action |
| Tools | None or limited | APIs, files, databases, email |
| Steps | One question, one answer | Many steps, plans its own route |
| Risk | Wrong information | Wrong information and a wrong action |
Where it helps in a business
- Reading an order out of an email, recording it and asking when something is missing.
- Classifying quote requests and routing them to the right person.
- Matching invoices to accounting lines and handing uncertain ones to a human.
- Summarising long reports and flagging anomalies.
For work whose rules are fixed, an agent is unnecessary: classic automation is cheaper, faster and more predictable. We explain that split in our automation article.
Risks and limits
When agents fail, the result isn't only a wrong sentence but a wrong action. OWASP's Top 10 for LLM Applications puts prompt injection first: an instruction hidden in a document or web page the model reads can pull an agent off its goal. So:
- Give the agent only the permissions its job needs; keep deletion and money movement out.
- Require human approval for irreversible steps.
- Treat text from outside as data, never as instructions.
- Log every step so the work can be audited afterwards.
Frameworks and regulation
NIST's AI Risk Management Framework is a good starting point. In Europe, the AI Act entered into force on 1 August 2024 and its obligations apply in stages; the European Commission page keeps the timetable current. If you serve European users, follow it.
When you put an agent to work, ask: what happens if it gets this step wrong? If the answer is "something irreversible", a human belongs in that step.
Sources
- Wikipedia, Intelligent agent
- NIST, AI Risk Management Framework
- OWASP, Top 10 for LLM Applications
- European Commission, AI Act
Frequently asked questions
Will an agent replace an employee?
In practice it takes the tedious steps and leaves the decision to a person, who also sets it up and supervises it.
Does our data go into the model?
It depends on the provider's contract. Business plans can commit to not training on your data — get that in writing.
Who is responsible when an agent gets it wrong?
Whoever builds and runs the system, which is why permission limits and approval steps are part of the design.
Is it expensive for a small business?
Not when scope stays narrow. Automating one tedious task is often a few days of work.
