Agents are employees without a contractAgents open cases, change data and send messages. Four levels of autonomy, a probationary period with fixed metrics and a one-page profile bring order to their use.AI agentsFurther reading
Topic · AI agents
Agents act. So they need to be managed.
Agents carry out work steps on their own. That changes who checks which work and who answers for the result.
An agent pursues a goal over several steps, using tools that write to systems, send messages or trigger orders. The check shifts from “Is the answer right?” to “May this step happen without approval?”
Before the first agent goes live, you therefore need a governing framework. It sets out which task an agent takes on, which permissions it gets, who supervises it and how its work can be traced afterwards. We make the case in our article “Agents are employees without a contract”.
Context
What’s changing and how we see it
What has changed
Agent features now come built into widely used office and developer tools. Open standards make it easier to connect agents to company systems, such as the Model Context Protocol (MCP), which Anthropic published in November 2024 and which has sat with the Linux Foundation's Agentic AI Foundation since December 2025. That also makes it easier to give an agent more access than its task needs.
Where the risks lie
The OWASP Top 10 for LLM Applications (2025 edition) lists excessive agency as a risk in its own right and names three causes: too much functionality, too many permissions, too much autonomy. On top of that comes manipulation through content the agent reads, such as doctored emails or web pages (prompt injection). The most effective defence against both is limits that sit outside the model, meaning permissions in the target system and human sign-off.
Our view
Which agent is technically the best is a secondary question for us. What matters more is which steps it may take without approval and who answers for it. We recommend introducing an agent the way you'd bring in a new colleague, with a job description, limited permissions, a named manager and a probation period after which its autonomy grows step by step.
What matters
Ground rules
Set levels of autonomy
Every agent gets a level, from making proposals to acting on its own within a clearly defined area. It only moves up once it has proven itself.
Limit permissions
An agent gets its own identity in the permission system and only the access its task requires.
Log every action
That way every decision can be explained afterwards.
Key questions
What you should clarify
Which steps may an agent take without approval?
Steps that can be undone and where the damage from an error is limited. Everything else it puts to a person until it has proven itself at the lower level.
Who is accountable for an agent internally?
A named person in the business unit who answers for the task and the result, and a team in IT responsible for permissions and operation.
How do we know an agent can be given more autonomy?
By measures agreed before the start, such as the share of proposals people had to correct. You set the thresholds; we help you choose them.
What applies to agents in HR?
Under the EU AI Act, AI systems that select or assess job applications count as high-risk (Annex III). Following the July 2026 amendment their obligations apply from 2 December 2027; in Germany the works council's co-determination rights already apply today.
How we help
Relevant capabilities
- Strategy & Consulting · DecideDeciding where AI pays off, and where it does not.
- AI strategy and roadmap
- Use-case portfolio and business cases
- Operating model
- AI governance
- Engineering · BuildWe build what we recommend.
- Agents and automation
- Platforms and integration
- Robotics and edge AI
Insights
Related articles
Agents are employees without a contractAgents open cases, change data and send messages. Four levels of autonomy, a probationary period with fixed metrics and a one-page profile bring order to their use.AI agentsFurther reading
Prototypes that force decisionsA good prototype answers a question you can't answer without it. Four rules and a one-page agreement make sure a decision is taken at the end.Generative AIFurther reading
The EU AI Act without the fussSince July 2026 there's a revised timetable, and Germany has a new supervisory authority. What companies using AI need to do now, and what can wait until 2027.AI governance & regulationFurther reading