In this article
- What is an AI agent, exactly?
- How does an AI agent work?
- What is the difference between an AI agent and a chatbot?
- AI agent, workflow or RPA: what is the difference?
- Is Copilot an AI agent?
- What types of AI agents are there?
- What are examples of AI agents in a business?
- What can an AI agent take over in each department?
- What does an AI agent look like when it actually runs?
- When do you not need an AI agent?
- What are the risks of AI agents?
- What may an AI agent do under the GDPR and the EU AI Act?
- Should you build an AI agent yourself or have one built?
- How much does an AI agent cost?
- Which terms come up around AI agents?
- Where do you start?
- Frequently asked questions
What is an AI agent? It is software that gets a goal, decides for itself which steps are needed and then carries out those steps in other systems, such as your email, CRM or ERP. A chatbot only answers. An agent acts: it reads a request, looks up data, updates a record and first shows a person anything that matters.
The short definition is in our glossary under AI agent. On this page I answer the question "what is an AI agent?" in full. You will read how it works and how it differs from a chatbot, a workflow and Copilot. You will also see what AI agents look like inside a business, and when you are better off without one.
My starting point: begin with the work, not with the agent. Many tasks that reach us as an agent question are solved faster and more cheaply with an ordinary workflow. So that is what we tell people.
What is an AI agent, exactly?
An AI agent combines four things: a language model, instructions, access to systems and limits. The language model understands the question and chooses the next step. The integrations let it actually do something, and the limits decide what it may do on its own. Without all four, you have an assistant, not an agent.
So when people ask me what is an AI agent in practice, I check whether three things come together:
- A goal instead of 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 those three is missing, you have an assistant, a chatbot or a workflow with an AI step. That is not worse. In fact, it is often exactly what you need.
How does an AI agent work?
An agent works in a loop. It reads, picks a step, carries it out, looks at the result and then picks the next step. It repeats this until it reaches the goal, gets stuck or needs a person to decide. Every step goes into a log, so afterwards you can see what it did and why.
For example, take a customer email asking to move a delivery by a week. First, the agent reads the email and recognises the customer and the order number. Then it finds the order in the ERP and checks whether the new date works. After that, it prepares the change and a draft reply, and it puts both in front of a colleague for approval.
The tools an agent uses are connections to 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. That, too, is part of the answer to "what is an AI agent?": a model with the right keys. How to set that up is covered in connecting an AI agent to your existing systems and, per software package, on our API integrations page.
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. The chatbot talks; the agent finishes the job. For most directors, that difference is the quickest answer to "what is an AI agent?" It is also why an agent asks more of you: access, limits for each action, and logging.
A chatbot on your website tells a customer how to register a return. An agent registers the return, creates the return label and prepares the refund for approval. For customer questions by email and chat, we have a separate service: AI customer service.
AI agent, workflow or RPA: what is the difference?
A workflow follows fixed steps that you draw up in advance. RPA (robotic process automation) copies what a person does on a screen, such as clicking and retyping. An AI agent chooses the steps itself, based on what it reads. If the steps are fixed, a workflow is cheaper and easier to control than an agent.
- 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 old 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, 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 requires reading and deciding. How to choose per step is in AI agent or ordinary 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 Agent Builder you can make simple agents that follow instructions and use your own knowledge. With Copilot Studio you can build agents that also take actions with tools. Copilot itself therefore does nothing on its own initiative. So in a Microsoft environment, "what is an AI agent?" really means: what does Copilot Studio add?
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. What you can do with it, and where it stops, is in Copilot Studio and Copilot agents. For Copilot itself, see Microsoft 365 Copilot for business.
If you prefer ChatGPT or Claude, read ChatGPT for business, Claude for business or the comparison Claude, ChatGPT, Copilot or Gemini.
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 mainly differ in how many systems they touch and how much they may do on their own. The more systems involved, the more limits and logging you need.
- Personal agents: an employee asks an agent in ChatGPT, Claude or Copilot to research or prepare something. Few integrations, and a person reviews everything.
- Single-task agents: an agent that sorts incoming support questions, or prepares invoices. Clearly scoped, 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.
When several agents work together, each doing part of the job, people also call it agentic AI. That sounds attractive. But every extra agent is one more place where something can go wrong. Start with one.
What are examples of AI agents in a business?
The best examples are boring: work that comes back every day, involves language or documents and has a clear end point. Sorting support questions, checking purchase invoices and preparing quotes suit an AI agent well, because a person can check the result quickly.
- Sorting support questions: our support ticket triage agent reads new questions, sets the topic and urgency and routes them to the right team, with the context already attached.
- Checking purchase invoices: the AI invoice checker matches invoices against the order, flags differences and only escalates the exceptions.
- Preparing quotes: the AI quote assistant drafts a first version based on the request and your own pricing agreements, within the margins you have set.
You will find more examples in AI agent use cases for SMEs.
What can an AI agent take over in each department?
Almost every department has work that an AI agent can prepare or finish: sorting questions, reading documents, checking data and preparing a draft. The difference per department lies in the systems and in what a mistake costs. In finance, almost everything goes past a person first; in a shared inbox, less so.
So per department, the question "what is an AI agent?" gets a very concrete answer. We have worked out 62 of those tasks as an AI employee: an agent with one task, fixed limits and a person for the exceptions. By department:
- Finance: checking invoices, chasing payments, processing expense claims.
- Sales: following up leads, preparing quotes, entering orders.
- Support: sorting questions, handling returns, answering the phone.
- Operations: work orders, the shared inbox, documents and reports.
- HR: onboarding, HR questions, sick leave and payroll changes.
- IT, logistics, planning, procurement, quality, legal and marketing.
How to choose between an agent and an ordinary workflow for each process is explained in our pillar on automating business processes with AI.
What does an AI agent look like when it actually runs?
In a demo, the answer to "what is an AI agent?" looks like magic. An AI agent in production is less spectacular than that. It does one thing, every day, and passes the tricky cases to a person. What makes it useful is not the model. Instead, it is the limits, the logging and the colleague who reviews the exceptions.
You can find worked examples in our cases: an AI agent dashboard for a retailer, an AI assistant that fills in the SUGB equal pay form from the collective labour agreement and the AI platform for Spectre. If you want to write up a first role yourself, use the blueprint for your first digital employee or the list of 50 tasks for an AI employee.
When do you not need an AI agent?
More often than you think. A director who asks "what is an AI agent?" usually wants to know whether they need one. If a task always follows the same steps and the input is clean, an ordinary integration or workflow is faster, cheaper and easier to manage. An agent only pays off when the steps differ per case and the volume is high enough.
- The rules are fixed: an amount above a threshold goes to a second approver. You do not need an agent for that.
- The volume is low: a five-minute task done ten times a month will not earn back the build and the upkeep.
- A mistake cannot be undone: then a person should decide, not an agent.
- The data is not digital: then there is other work to do first.
More situations where AI is not the answer are in when AI is not a good idea.
What are the risks of AI agents?
The strength of an agent is also its risk: it takes steps without asking. That is why you set limits, approvals and logging before it may do anything. Four risks keep coming back: errors that carry through, permissions that are too broad, misleading input, and not being able to trace what happened.
- Errors that carry through: a wrong assumption in step two carries through to step six. So build in fixed moments where a person checks.
- Permissions that are too broad: only give an agent access to the systems and actions the task requires.
- Misleading input: an email or document can contain instructions that try to talk the agent into something. See prompt injection.
- Not being able to trace what happened: log what the agent saw, what it decided and what it did. Also when nothing went wrong.
You can read more about made-up answers in preventing AI hallucinations.
What may an AI agent do under the GDPR and the EU AI Act?
In most business processes an AI agent may simply do its work, as long as you sort out three things. First, since 2 February 2025 you have to take measures to support the AI literacy of the people who work with AI. Second, if the agent handles personal data, the GDPR applies. Third, if it helps decide on recruitment, credit or access to services, it is high risk.
For those high-risk uses, the heavy requirements apply from 2 December 2027, after the EU AI Act was amended in July 2026 (Regulation (EU) 2026/1744, EUR-Lex). Think of risk management, logging, documentation and human oversight. An agent that only sorts support questions or prepares invoices usually falls outside that category.
My advice: build every agent as if it does fall under those rules. Logging, limits for each action and a person at every irreversible step cost little extra, and they also make an agent easier to manage. What else you need to arrange is in the EU AI Act and the GDPR: what to sort out.
Should you 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 has to write to your ERP or CRM, handles sensitive data or must run in production with logging and support. At that point it is no longer an experiment, but software you have to maintain.
What each of those tools can do, and where you hit the limit, is in build your own AI agent. If you want an agent that finishes work in your own systems, with approval and logging, you can have us handle AI agent development. An agent that only answers your phone is a separate service: the AI phone assistant.
How much does an AI agent cost?
That 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 make the call. On top of that you pay for using the language model and for support after go-live. Always set that against what the work costs today.
On what does an AI agent cost? you can run the cost check. How to calculate the return is in the AI business case.
Which terms come up around AI agents?
The same words keep coming back as soon as someone asks "what is an AI agent?" in a sales meeting. If you know them, quotes and product pages become a lot easier to read. A language model is the engine, while RAG and MCP are the ways an agent reaches your own information and systems.
- Agentic AI: several agents that finish a task together.
- Large language model: the model that lets the agent read and write.
- RAG: the agent first searches your own documents and bases its answer on them.
- MCP: an open standard that lets an agent reach your systems.
- Generative AI: AI that creates text, images or code.
Where do you start?
Once you know the answer to "what is an AI agent?", start with one task that comes up often, involves language or documents and has an outcome a person can check quickly. At first, let the agent only make suggestions. Then measure how often they are right, and only after that give it more room. That way trust grows from numbers, not from a demo.
Not sure yet which work would gain the most, or whether an agent is even 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 do not need an agent. If you then want to know how to introduce AI step by step, read the AI implementation roadmap for SMEs.
Frequently asked questions
What is an AI agent, in short?
Software that gets a goal, decides for itself which steps are needed and carries out those steps in other systems, such as email, CRM or ERP. An agent acts; a chatbot answers.
What is the difference between an AI agent and a workflow?
A workflow follows fixed steps that you draw up in advance. An agent chooses the steps itself, based on what it reads. If the steps are fixed, a workflow is cheaper and easier to control.
What are examples of AI agents?
An agent that sorts support questions and routes them to the right team, an agent that matches purchase invoices against the order and flags differences, and an agent that drafts a quote from your own pricing agreements. You will find more examples under AI employees.
Is ChatGPT an AI agent?
At its core, ChatGPT is an assistant that answers what you ask. With actions or integrations through MCP, it can also do things in other systems. Whether something is an agent depends on how it is set up: does it get a goal, does it choose its own steps and can it act in other systems?
Are AI agents safe?
For clearly scoped tasks, yes, as long as the agent only gets the permissions the task requires, a person approves irreversible actions and every action is logged. The more steps and systems involved, the more those limits matter.
Sources
- Regulation (EU) 2026/1744 (digital omnibus on AI), EUR-Lex
- Overview of Copilot Studio agent builder, Microsoft Learn
















