What is an AI agent? You hear everywhere that you need them and that AI agents are the future. But what exactly is an AI agent, and how do you decide if your company needs one? Over the past few months, we have had many conversations where agents were simply used to mean a chatbot with a better prompt. That is a shame, because that is not what they are. The technical difference is small, but the organizational one is huge. In our three-part blog series, we will take a closer look at AI agents. After reading this first part, you will know what components make up an agent, what the four collaboration models are, and which use case you can start with.
An AI agent is software that receives a goal and determines the path to achieve it independently. It breaks down the task, calls up tools, reviews the result, and corrects itself if something is not right. Only when the goal is reached or it gets stuck does it report back. The difference to everything we've seen before is summed up in one word: loop. A chatbot answers once. An agent works, checks its progress, and continues working. That is why it can complete a task, not just answer a question.
The stages can be described as follows:
Chatbot: The agent only knows what is in the model and what you enter. Useful for formulating, but blind to your company's needs.
Assistant: It sees your emails, files, and appointments. It formulates, summarizes, and suggests solutions. You have to execute them yourself.
Agent: It is authorized to act. It reads a ticket, searches the document repository, creates a response, and updates the system.
Agent Team: Several agents work together. One researches, one writes, and one reviews. They share interim results and escalate to you if anything is unclear.
When we build agents, we always use the same six building blocks. This is true regardless of whether you are using Copilot Studio, Microsoft Foundry, or your own platform.
The difference is best illustrated by a real-world example. Let's take a support ticket: «Outlook will not start after the update, and I cannot access my emails». Here is how an agent would handle it:
The value lies in steps 4 and 5. An assistant would have stopped in step 3 and shown you three links. An agent assesses and decides whether they can continue working. And step 5 contains the most important design decision: What can they do themselves, and when does the task need to be delegated to a human?
Note: When escalating a task, make sure that an agent always provides information about what they have already reviewed. Otherwise, the human will have to start over, and the time saved will be lost. This is a matter of how the task is worded, not a question of the model itself.
Tools give an agent access. But access is not the same as ability. An agent with access to Word does not know what your offer looks like. Skills fill precisely this gap. And they are the most underrated innovation of the last year. A skill is, at first glance, unremarkable: a folder containing a file called SKILL.md. This file describes, in plain language, how to perform a specific task. Optionally, it also includes scripts, templates, and reference documents. Anthropic introduced the concept in October 2025 and released it as an open standard in December 2025. According to the official directory, around forty products now support the same format – including OpenAI Codex, GitHub Copilot, Cursor, Gemini CLI, and VS Code. So, a skill you write once will work in multiple tools.
The real key is how they are loaded.
Discover: At startup, the agent only reads the name and description of each skill. This costs approximately 80 tokens per skill.
Activate: If the skill matches the task, it reads the complete SKILL.md file. This is where it gets expensive.
Execute: It follows the instructions and uses the provided scripts and templates.
The effect is enormous: An agent can know dozens of skills while consuming less context than a single, fully loaded skill. This solves a problem that was previously almost impossible to solve: providing an agent with a wealth of company-specific knowledge without overloading its context. For companies, the practical value is even greater than the technical one. A skill is the first artifact form in which you can document process knowledge in a way that both humans and agents can use it. It is version-controlled, reviewable, and has an owner. If you start an agent project today, ask yourself early on: Which five skills do we need, and who in the business unit will write them?
Note: Write skills like a good set of instructions for a new colleague. When does it apply, what are the steps in which order, and what is prohibited? The description in the file header is the most important part. It is what the agent uses to decide whether to even read the skill.
«Tools» is a collective term, and the differences between them are crucial. An agent can act in four different ways, and in a well-designed system, it combines them.
Let's be clear about this to avoid a lot of disappointment:
An AI agent is not a replacement for a sound process. If your process is broken, you are just automating the chaos.
An AI agent is the wrong choice for tasks with fixed rules. Approval based on four fixed criteria belongs in automation, not in a language model.
An AI agent is not a replacement for good data. It can only find what is discoverable and has legitimate access.
An AI agent is not a staff reduction program. The cases that work save time – they do not replace roles.
So far, it is been about technology. Now for the part that drives the decision. For its Work Trends Index 2026, Microsoft surveyed around 20,000 knowledge workers in ten countries and analyzed anonymized usage signals from Microsoft 365. Three figures are truly relevant:
The figure that impressed us most is a different one: According to the report, organizational factors explain around 67% of the impact of AI, while individual factors account for only about 32%. Put simply: It is not how skillfully individual people prompt that determines the outcome. It is whether the company transforms work. And at the same time, only 26% of executives say their organization has a clearly defined AI strategy. That is where the competitive edge lies.
Note: The term frontier firm comes from the Work Trends Index. It refers to a company that purchases intelligence on demand, works in human-agent teams, and establishes a new role: the Agent Boss. A person who builds, commissions, and manages agents.
The report of 2026 describes four patterns of how people and agents collaborate.
The leap from Editor to Director is the real step. It is not a technical issue, but a question of trust and governance: You are delegating an entire task, no longer just a draft. This requires protocols, permissions, and a defined escalation path; otherwise, delegation becomes a loss of control.
Three criteria: Does it happen often? Is the data located in a place I have authorized access to? And is an error noticeable and correctable before it causes problems? If the answer to any of these is no, it is not a good first use case.
«Saves time» does not survive any budget round. Therefore, calculate before building, and be deliberately conservative. Example: Ticket triage in a service desk with 60 tickets per day: Today, pre-sorting and enriching a ticket takes about four minutes. That is four hours per day. Realistically, the agent pre-processes 60% of the tickets completely, makes a mistake on 15%, and a human corrects it. That leaves about two hours per day – not four. The corrections take time and should be factored in. Consider this in terms of model and platform costs, as well as the one-time setup. If it does not pay for itself in under six months, it is the wrong first use case. This calculation is not pretty, but it is defensible. And it almost always leads to a narrower definition of the use case. Which is good, because narrower agents perform better.
Inventory: Which agents are already running in your tenant? The answer is almost always «more than you thought.»
Remember: Without a registry, there is no governance.
Identity: Does each agent have its own identity with its own set of permissions? If not, change that before going live.
Data classes: Which data is an agent allowed to see, and which is strictly off-limits? This is a decision for the business department, not IT. Measurement: How will you know in eight weeks that it is working? If you cannot say now, you will not be able to later.
Start with the most intriguing case instead of the most useful one. The most intriguing case is usually the one where everything is missing: data, permissions, process.
Build without a way back. Every action needs a way to undo it – reopening a ticket, saving an email as a draft instead of sent. Without a way back, every mistake becomes an incident.
Start with a messy knowledge base. If two documents contradict each other, the agent will contradict itself too. And you will look for the error in the model, even though it is in the folder.
Do not set a budget limit. Agents run in the background, and consumption models don't remember that you wanted to save.
An AI agent is not simply a better chatbot, but an actor with a mission, tools, and an identity. Whether your company needs one is not determined by the technology, but by whether there is recurring work, data that is readily available, and where you can identify errors early. If so, start with precisely such a case, give the agent its own identity, and measure the results.
A chatbot answers a question once. An agent receives a goal, plans the route itself, calls up tools, checks the result, and continues working until the task is completed or it gets stuck. The difference lies in this checking and correction step.
Four things:
an inventory of which agents are running,
a separate identity with its own rights for each agent,
a decision by the department regarding which data an agent is allowed to see,
and a benchmark after eight weeks to determine if it worked.
With one that occurs frequently, whose data is stored in an accessible location, and where an error is noticeable and correctable. In practice, this usually involves ticket triage, preparing quotes, or research and summarization.
The issue is not so much the license as the resource consumption: An agent running hourly in the background costs significantly more than one that operates on a click-by-click basis. Calculate conservatively for correction costs and set a budget limit for your total costs (but not for tokens per session) before the first agent goes live.