Agent
Colleagues’ questions about policy
Every question about leave, working from home or expenses is different. The agent finds the right passage in your documents itself and answers with a link to the source. If it finds nothing, it says so.
In plain language
What is an AI agent, and do you need one?
An AI agent is software that decides for itself which step to take to finish a task. It looks up information, uses other software and decides what happens next. An AI workflow always follows the same steps. Here you’ll read the difference, where an agent really helps, what it may do on its own and how to start without the hype.
Written by Peter Dedene · last reviewed on 11 October 2026
At a glance
What you need to know
-
01
An AI agent chooses its own next step to finish a task. A workflow follows fixed steps. A chatbot only answers within a conversation.
-
02
There is no fixed definition. Even the European Commission calls the term still evolving. We use one test question: does the software decide the next step, or does the model?
-
03
An agent is not better by default. Anthropic, the maker of the Claude language model, also advises starting with the simplest solution and only building something more complex when it turns out to be needed.
-
04
What can be undone and stays inside your organisation, an agent may do on its own. Anything that goes outside or decides about a person, a colleague approves first.
-
05
Under the AI Act, an agent is not a separate category. The same rules apply as for any other AI system.
Chatbot, workflow or agent?
Three forms, one question: who takes the next step?
In practice the words get mixed up, and vendors call almost everything an agent. So we compare on one point: who decides what happens after the first step.
| Chatbot | AI workflow | AI agent | |
|---|---|---|---|
| What it does | Answers a question in a conversation | Fixed steps. The AI reads, sorts and prepares. | Finds what it needs and picks the next step |
| Who takes the next step | You, with a new question | The steps set in advance | The model, within the agreements |
| Example | Having a text rewritten in a chat window | Reading invoices and preparing them in your accounting | Answering a question about leave policy by finding the right documents itself |
| What to watch | What staff paste into it, and where that ends up | Do the steps still hold when your software changes? | How much access it gets, and who checks its work |
Where an agent helps
An agent pays off where every case is different
An agent only makes sense when you can’t fix the steps in advance. Four examples from practice, each with the honest answer: agent or workflow?
Agent
Preparing a file from several systems
For an application or a meeting, the agent pulls data from the CRM, the mailbox and the shared folder, and prepares a first overview. Your team adds to it and decides.
Workflow
Processing invoices
Every invoice goes through the same steps: read, check, prepare. A workflow is enough there, and it is more predictable.
Workflow, sometimes with one agent step
Sorting a shared mailbox
Labelling and routing follow fixed rules. If a reply has to pull information from several sources, that one step can be an agent.
What may an agent do on its own?
Two kinds of doors
Some doors swing both ways, others only one. We use that image to agree in advance what an agent may do on its own. It matches what OpenAI and OWASP, an international organisation for software security, recommend: let a person approve anything sensitive or irreversible.
On its own, because it can be undone
- Labelling an email
- Preparing a draft
- Looking up information in agreed sources
- Filling in a field that someone still checks
A colleague first, because it goes outside or concerns a person
- Sending an email or reply
- Paying or booking something
- A decision about a client, pupil or employee
- Sharing data outside your organisation
We set this out for each agent in the AI fact sheet.
What can go wrong
Real risks, and what helps
Even the strongest models can be misled. In a large test by the US NIST in 2026, attackers managed to mislead every model tested at least once. That’s why we build in checks rather than trusting the model blindly.
Hidden instructions
An email or document can contain instructions that mislead an agent. In June 2025 such an email made Microsoft 365 Copilot leak data without anyone clicking anything. The flaw has since been fixed.
What helps: an agent sends nothing outside without approval.
A wrong action
In July 2025 a coding agent at Replit deleted a database that was in use, against explicit instructions.
What helps: read-only access where reading is enough, and testing in a separate environment first.
Too much access
An agent that can see everything can also get everything wrong.
What helps: its own account with only the permissions the task needs.
Rising costs
An agent that keeps searching keeps running up costs. Gartner expects over 40% of agent projects to be stopped by the end of 2027, because of high costs, unclear value or too little control.
What helps: a limit on the number of steps, and measuring what a task costs.
How to start
Five steps, no hype
-
01
Choose one task
A task that comes up often and costs your team time every week.
-
02
First check what is enough
Does your current software already do it, or is a workflow enough? Often it is, and then we say so.
-
03
Write down the agreements
What may the agent do on its own, what data does it see and who checks? That goes in the AI fact sheet.
-
04
Test for four to six weeks
With real examples. We measure how much time it saves and how often someone has to correct it.
-
05
Decide with numbers
Roll it out, adjust it or stop.
Rather do it together?
Start with a one-hour Pitstop
You bring one task. Together we look at whether an agent, a workflow or your current software fits best.
Is ChatGPT an AI agent?
A plain conversation in ChatGPT or Copilot is not an agent: you ask a question and get an answer. But OpenAI and Microsoft now build agent features into them, which take steps on their own in a browser or in your Microsoft 365 environment. Then the same questions apply: what may it do on its own, and who checks?
What is agentic AI?
An umbrella term for AI systems that take steps on their own, from simple workflows to full agents. The word says little about how independent a system really is. So always ask: who decides the next step?
Does an AI agent fall under the AI Act?
Yes, like any AI system. The European Commission states explicitly that agents are not a separate category. If an agent talks to people or creates text or images, you must make that clear from 2 August 2026. If it helps decide about job applicants, education or access to support, stricter rules apply. This is not legal advice.
What does an AI agent cost?
That depends on the task, the systems and the number of steps. An agent usually costs more than a workflow, because every step calls on the model again. After a first conversation you get a proposal with a clear scope and a fixed price. In the pilot we measure what a task costs.
Can I build an AI agent myself?
Yes. With tools such as Microsoft Copilot Studio you can build a simple agent yourself, and for a test that’s fine. If the agent works with personal data or writes into your own systems, think about access, logging and who maintains it.
Does an AI agent replace staff?
Not with us. An agent takes over searching and preparatory work, so your team has time for what needs people. We don’t build systems meant to replace staff.
Not sure whether you need an agent?
Tell us which task costs your team time. In a one-hour Pitstop we look together at what fits best: an agent, a workflow or the software you already have. Sometimes the answer is: none of the three.
One hour · at our place or online
Sources
The facts on this page come from the publications below. As of 11 October 2026.
- Anthropic, Building effective agents (December 2024) anthropic.com (opens in a new tab)
- OpenAI, A practical guide to building agents cdn.openai.com (opens in a new tab)
- European Commission, AI Act Service Desk: How are AI agents addressed within the AI Act? ai-act-service-desk.ec.europa.eu (opens in a new tab)
- OWASP, LLM06: Excessive Agency (2025) genai.owasp.org (opens in a new tab)
- NIST CAISI, Insights from a large-scale AI agent red-teaming competition (March 2026) nist.gov (opens in a new tab)
- Gartner, Over 40% of agentic AI projects will be canceled by end of 2027 (June 2025) gartner.com (opens in a new tab)