AI agents: what they are, how they work and where they help at work
AI agents receive a goal and carry out the steps towards it themselves: they look up information, use tools and prepare results. This guide explains how AI agents work, where they help in a company and which rules apply to data, permissions and approvals.
With a checklist for company use.
Key points
- An
AI agent plans the steps towards a goal itself and uses tools such as search, databases or email. A language model decides what happens next. - Good agents need
reliable knowledge from your sources. Without a source the model guesses; with a source every statement can be checked. - The more an agent may do on its own, the more
permissions, logging and approvals matter. Write actions belong behind approval by a person.
What are AI agents?
AI agents are software systems that pursue a goal on their own. A language model breaks the task into steps, picks suitable tools and assesses every intermediate result. This creates a loop of planning, acting and checking. It runs until the goal is reached or a limit applies, for example approval by a person.
The idea of alternating reasoning and acting is described, among others, in the research paper Yao et al., ReAct (arXiv 2210.03629). Fraunhofer IPA (German) gives an overview with a focus on manufacturing.
The difference from a chat is action. A chat answers a question. An agent can look things up, calculate, write a draft and hand it to the right system.
How do AI agents work? Five building blocks and a loop.
Most AI agents can be described with five building blocks, whether they sort complaints or prepare quotes. Anyone planning or assessing an agent checks these points and the loop that connects them.
Language model
The model understands the task and decides the next step. It can run with a provider or on your own server.
Goal and instructions
What should be achieved, in what tone, within which limits? Clear instructions make the behaviour predictable.
Tools
Interfaces to search, databases, email or the ticket system. Every tool is a capability and a risk at the same time.
Knowledge
Manuals, contracts, product data. The agent queries them when needed instead of answering from the model’s memory.
Limits and approvals
Permissions per tool, a log of every step and approvals for sensitive actions. This is what decides whether an agent is safe to run.
The loop
Plan, act, check, repeat. A cap on steps and cost stops an agent from running on forever.
Assistant, workflow or AI agent: what kinds of AI agents are there?
Not every task needs an agent. Often an assistant that answers with sources is enough, or a fixed workflow in which AI takes over single steps. The table shows four levels. With each level, the software makes more decisions itself.
| Level | What it does | Who decides the next step? | Example |
|---|---|---|---|
| AI assistant | answers questions from defined sources | the person | HR questions on leave and travel expenses |
| Workflow with AI | follows fixed steps, AI handles some of them | the defined workflow | classify incoming emails and route them |
| AI agent | plans the steps itself and uses tools | the model, within its limits | check a complaint and prepare a credit note |
| Multi-agent system | several specialised agents share one task | a coordinating agent | research, drafting and review by separate agents |
Rule of thumb: start with the lowest level that solves the task. An assistant with sources is quicker to introduce and easier to check than an agent.
Where AI agents help in a company.
AI agents pay off for tasks that come up often, touch several systems and follow clear rules. Six typical examples from different departments:
Customer service
Read the request, look up the order, draft a reply with sources. Credit notes go for approval.
HR
Create onboarding tasks, answer questions on policies, check documents for completeness.
IT support
Classify the ticket, suggest a fix from the IT wiki, prepare standard cases such as password guidance.
Sales
Create a draft quote from product data and pricing rules, gather open points for the call.
Purchasing and contracts
Find deadlines and termination clauses in contracts and suggest reminders.
Knowledge work
Research across many documents, summary with references, draft for the specialist team.
Without reliable knowledge, the agent guesses.
An agent is only as good as the information it works with. Language models do not know your manuals, contracts and product data. If knowledge is missing, the model fills the gap with plausible-sounding guesses. For an agent that acts on them, that costs more than in a chat.
That is why many companies separate agent and knowledge. The agent queries a knowledge layer and gets passages with sources back. The open Model Context Protocol makes this connection uniform. RAGSuite is an MCP server and an MCP client, both in the Community Edition. The actions themselves are carried out by the agent or by a workflow tool such as n8n that queries RAGSuite as its knowledge source.
Risks of AI agents and what helps against them.
The OWASP Top 10 for LLM Applications 2025 lists “excessive agency” as a risk of its own (LLM06): a system has more functions, permissions or autonomy than the task needs. Prompt injection, data protection and co-determination come on top. The table pairs each risk with a countermeasure.
For AI agents that interact directly with people, Art. 50(1) of the EU AI Act, Regulation (EU) 2024/1689 has applied since 2 August 2026: providers must design such systems so that people are informed they are interacting with an AI, unless this is obvious. Since 2 February 2025, Art. 4 has also applied: measures for staff AI literacy. This text is not legal advice.
| Risk | What can happen | Countermeasure |
|---|---|---|
| Too many permissions | The agent changes or deletes data it should not touch. | Read-only by default, enable write tools one by one |
| Prompt injection | Hidden instructions in an email or web page redirect the agent. | Treat external content as data, approval before actions |
| Wrong facts | The agent acts on an invented statement. | Query knowledge with sources, show sources in the result |
| Automated decisions | A decision with legal or similarly significant effects is based solely on automated processing (Art. 22 GDPR). | A person decides, the agent prepares |
| Co-determination | Logs allow conclusions about employees (Section 87(1) No. 6 of the German Works Constitution Act (BetrVG)). | Involve the works council early, settle purpose and evaluation |
| Runaway cost | A loop keeps running and uses compute. | Cap on steps and cost per run |
Preventing prompt injection →AI and the works council →EU AI Act Article 50 →
Before an AI agent goes live.
Tick what is done. Nothing is stored.
Questions about AI agents.
What is an AI agent in simple terms?
An AI agent is a program that receives a goal and plans the steps towards it itself. A language model decides which tool fits next, such as a search, a database or a ticket system. The agent checks the result and carries on until the goal is reached or a person has to approve the next step.
What is the difference between an AI assistant and an AI agent?
An AI assistant answers questions, for example from your team’s documents. An AI agent goes further and takes actions: it creates a ticket, writes a record or starts a workflow. That increases the responsibility. An agent needs clear permissions, a log and, for sensitive steps, approval by a person.
What kinds of AI agents are there?
Companies usually distinguish four levels. An assistant only answers. A workflow with AI follows fixed steps, with AI handling individual tasks inside it. An agent chooses its own steps. In a multi-agent system, several specialised agents share one task. The higher the level, the more limits and approvals matter.
May AI agents process personal data?
Yes, if there is a legal basis and the GDPR rules are followed, such as purpose limitation, data minimisation and data processing agreements. Decisions with legal or similarly significant effects must, under Art. 22 GDPR, in principle not be based solely on automated processing. Give an agent only the data it needs for the task. This text is not legal advice.
Does the works council have to agree to AI agents?
Often, yes. Under Section 87(1) No. 6 of the German Works Constitution Act, the works council has a say on technical systems designed to monitor behaviour or performance; under Federal Labour Court case law it is enough that they are objectively suitable for it. Many AI agents log who triggered what. Involve the works council early.
AI agents with knowledge you can check.
Jürgen works out with you which task suits an agent, what knowledge it needs and where an approval belongs.