What are AI agents, and how do they work?

How agents differ from chat assistants and automation, how they work step by step, and the controls that keep them safe.

A simple definition

An AI agent is a system that uses an AI model to work towards a goal. Instead of answering a single question, it decides what to do next, uses tools such as search, email, a calendar or a business system, checks the result and continues until the task is complete or it needs a person.

Assistant, automation or agent?

These three terms are often mixed up. The difference matters because each suits different work.

ApproachHow it worksBest for
Chat assistantResponds to your prompt. You decide what to do with the answer.Drafting, summarising and explaining
AutomationFollows fixed rules that you define, the same way every time.Predictable, repeated steps
AI agentPlans steps towards a goal, chooses tools and may take actions.Varied work that needs judgement within clear limits

How an agent works

Most agents follow a loop:

  1. Receive a goal, such as "sort today’s support requests and draft replies".
  2. Decide the next step based on the goal and what it knows so far.
  3. Use a tool, for example reading an email or looking up a customer record.
  4. Check the result and decide whether to continue, try another approach or ask a person.

The tools an agent can use, and what it is allowed to do with them, are set by the people who build and run it. That is where safety is decided.

Everyday examples

  • Sorting an inbox, labelling urgent messages and drafting replies for a person to approve
  • Turning a website enquiry into a customer record and a follow-up task
  • Researching suppliers and summarising the options in a consistent format
  • Collecting the details of an IT problem before a technician picks it up

The risks, and the controls that manage them

Because agents act, mistakes can have real consequences. Sensible controls include:

  • Least access. Give the agent only the systems and permissions the task needs.
  • Human approval. Require a person to approve anything that sends, pays, deletes or changes important records.
  • A full record. Log every action so you can see what happened and why.
  • Testing with realistic cases, including unusual and difficult ones, before wider use.
  • Limits on cost and running time so a confused agent cannot run indefinitely.
  • Awareness of prompt injection. Text hidden in an email or web page can try to give an agent new instructions. Agents should treat what they read as information, never as commands.

When not to use an agent

If a task follows the same steps every time, ordinary automation is usually cheaper, faster and more predictable. Agents earn their place when the work varies and a person can still check the important decisions.

If you are weighing up an agent for your business, our guide to giving an AI agent access without losing control covers permissions in detail, and our AI agents service starts with a small, supervised pilot.

Common questions

Are AI agents safe to use in a small business?

They can be, when access is limited, important actions need approval and every action is logged. Start with a supervised pilot on a low-risk task, and widen access only when the results justify it.

Do I need special software to run an AI agent?

Not always. Many business platforms, including Microsoft 365 and Google Workspace, now include agent features, and agents can also be built to connect to the tools you already use. The right choice depends on the task, your data and your budget.

How much does an AI agent cost to run?

Costs usually combine the AI model usage, any platform subscription and the time people spend reviewing the work. A pilot should measure all three before you decide to expand.