What is an AI agent? How it works, examples and when you need one

An AI agent is software that is given a goal, works out which steps are needed to reach it and carries out those steps in other systems, such as your email, CRM or ERP. Where a chatbot only answers, an agent acts: it reads a request, looks up data, updates a record and runs anything that matters past a person first.

The short definition is in the glossary entry AI agent. This article covers how an agent works, how it differs from a chatbot, a workflow and Copilot, what AI agents look like in a business and when you're better off without one.

Our starting point: begin with the work, not the agent. Many tasks that come in as an agent question are solved faster and more cheaply with a regular workflow. When that's the case, we say so.

What exactly is an AI agent?

An AI agent combines a language model, instructions, access to systems and limits. The language model understands the request and picks the next step, the integrations let it do something, and the limits decide what it may do on its own.

It only makes sense to call something an agent when three things come together:

• A goal rather than an instruction: not "summarise this email", but "make sure this request is in the CRM with the right details".
• Choosing its own steps: the agent decides which actions are needed and in what order, instead of following a fixed script.
• Access to tools: the agent can look something up, call a system, read a file or create a record.

If one of the three is missing, you have an assistant, a chatbot or a workflow with an AI step. That isn't worse. It's often exactly what you need.

How does an AI agent work?

An agent works in a loop: read, choose a step, carry it out, look at the result and choose the next step. It repeats that until the goal is reached, it gets stuck or a person needs to decide.

Take a customer email asking to push a delivery back a week. The agent reads the email and identifies the customer and order number. It looks up the order in the ERP and checks whether the new date works. It changes the order or prepares a proposal, drafts a reply and sends that to a colleague for approval. Every step is logged.

The tools an agent works with are integrations with your systems: an API, a connector or an MCP server. For each tool you decide what the agent may read and what it may change. How to set that up is covered in connecting an AI agent to existing systems.

What is the difference between an AI agent and a chatbot?

A chatbot answers a question and waits for the next one. An AI agent carries out tasks in your systems and keeps going until the work is done.

A chatbot on your website tells a customer how to register a return. An agent registers the return, creates the return label and sets up the refund for approval. The chatbot talks, the agent finishes the job. That's also why an agent needs more: access to systems, limits per action and logging.

AI agent, workflow or RPA: what's the difference?

A workflow follows fixed steps you design in advance. RPA mimics actions on a screen, such as clicking and retyping. An AI agent chooses the steps itself, based on what it reads.

• Workflow: if A, then B. Predictable, cheap and easy to check. A workflow can include an AI step, for example to read an invoice, without becoming an agent.
• RPA: handy for older systems without an integration, but fragile when the screen changes. See RPA vs AI.
• AI agent: useful when the order of steps differs from case to case and can't be fixed in advance, such as free-text email or requests that need to be understood first.

In practice you combine them: a workflow for the fixed steps and an agent for the part that needs reading and judgement. How to choose per step is covered in AI agent or regular automation and, for Microsoft environments, in Power Automate or an AI agent.

Is Copilot an AI agent?

Microsoft 365 Copilot is first and foremost an assistant: it helps you in Word, Outlook and Teams when you ask it to. With Copilot Studio and Agent Builder you can build agents that follow instructions, use your own knowledge and take actions with tools (Microsoft).

Copilot Studio is Microsoft's low-code environment for building agents, connecting them to your data and publishing them in Teams, Copilot or on your website (Microsoft). What you can do with it and where it stops is covered in Copilot Studio and Copilot agents. For Copilot itself, see Microsoft 365 Copilot for business.

What types of AI agents are there?

For a business there are roughly three types: personal agents, single-task agents and agents that handle a whole process. They differ mainly in how many systems they touch and how much they may do on their own.

• Personal agents: an employee has an agent in ChatGPT, Claude or Copilot research or prepare something. Few integrations, and a person reviews everything.
• Single-task agents: an agent that triages incoming support questions or prepares invoices. Well defined, easy to test and often the best first step.
• Process agents: an agent that finishes work across several systems, such as reading a request, checking data, changing an order and replying to the customer. More value, but it needs more limits and logging.

When several agents work together, each doing part of the job, that's sometimes called agentic AI. It sounds appealing, but every extra agent is one more place where something can go wrong. Start with one.

Examples of AI agents in a business

The best examples are dull: work that comes back every day, involves language or documents and has a clear end point. Three examples we've worked out as AI employees:

• Triaging support questions: Support Triage reads new questions, sets the topic and urgency and routes them to the right team with the context attached. Judgement stays with a person where it's needed.
• Checking purchase invoices: the Invoice Checker matches invoices to the order, flags discrepancies and only puts the exceptions in front of someone.
• Preparing quotes: the Quote Assistant drafts a first version from the request and your own pricing agreements, within the margins you've set.

More examples by department are on AI employees and in AI agent use cases for SMEs.

When don't you need an AI agent?

More often than you'd think. If a task always follows the same steps and the input is clean, a regular integration or workflow is faster, cheaper and easier to run.

• The rules are fixed: an amount over a limit goes to a second approver. You don't need an agent for that.
• The volume is low: a five-minute task done ten times a month won't pay back the build and upkeep.
• A mistake can't be undone or quickly checked: then a person should decide, not an agent.
• The data isn't digital or lives in people's heads: then there's other work to do first.

More situations where AI isn't the answer are in when AI is not a good idea.

What are the risks of AI agents?

An agent's strength is also its risk: it takes steps without asking. So you set limits, approvals and logging before it's allowed to do anything.

• Errors that carry through: a wrong assumption in step two affects step six. Build in fixed moments where a person checks.
• Too many permissions: give an agent access only to the systems and actions the task needs.
• Misleading input: an email or document can contain instructions that try to talk the agent round. See prompt injection.
• No record of what happened: log what the agent saw, what it decided and what it did. Even when nothing went wrong.

On top of that, the GDPR and the AI Act apply as soon as an agent handles personal data or helps make decisions about people. What you need to arrange is covered in the AI Act and GDPR: what do you need to do?.

Build an AI agent yourself or have one built?

You can make a first agent yourself with Copilot Studio, ChatGPT, Claude or n8n. Having one built makes sense when the agent needs to write to your ERP or CRM, handles sensitive data or has to run in production with logging and support.

What each of those tools can do and where you hit the limit is covered in building your own AI agent. If you want an agent that finishes work in your own systems, with approvals and logging, you can have an AI agent built.

What does an AI agent cost?

It mainly depends on three things: how many systems the agent has to work in, how many exceptions the process has and where a person needs to be involved in decisions. On top of that you pay for language model usage and for running it after go-live.

Set that against what the work costs now. You can do the cost check on what does an AI agent cost?. How to calculate the return is covered in the business case for AI.

Where do you start?

With one task that happens often, involves language or documents and has an outcome a person can check quickly. Let the agent only make proposals at first, measure how often they're right and only then give it more room.

Not sure which work will pay off most, or whether an agent is the right answer there? Then the AI Opportunity Audit first maps what the current work costs and where the opportunities are. Even if the outcome is that you don't need an agent.