Agentic AI and organization: which roles agents change

Agentic AI first changes the roles that gather and pass on information, and it creates new roles for steering and quality assurance of agents. An agent receives a goal and plans the steps itself. A person approves its result, or it flows straight into the process. The delivery perspective therefore asks which roles and decision paths must change for agents to arrive in operations.
What distinguishes an agent from classic automation
Classic automation follows a fixed sequence. Someone defines the rule, the system executes it, and exceptions land with a person. An agent receives a goal and breaks it down into sub-steps. It selects tools and changes its approach when a step fails. For the organization this has an immediate consequence: people no longer prescribe individual steps. They define goal and boundaries and set the quality standard.
This changes how work is cut. Until now, a role consisted of a sequence of activities. In the future it consists of a result that people and agents work towards together. Organization design decides what this cut looks like. The technical foundations are described in Agentic AI: how AI agents take over workflows.
The starting position in Germany is broad enough for this question. According to the Federal Statistical Office, 26 percent of companies with ten or more employees were already using artificial intelligence in 2025, and among companies with 250 or more employees the figure was 57 percent. Where AI already runs in daily work, the step to agents is a question of organization. Where deployment touches employee data or co-determination, additional rules apply, outlined further below. This article provides an overview and does not constitute legal advice.
Which existing roles does agentic AI change first?
The first roles to change are those with a high share of recurring information work whose output can be checked. A study of the labor market impact of large language models estimates that for around 80 percent of the US workforce at least ten percent of tasks could be affected. For about 19 percent it is at least half. Agents amplify this effect because they take over entire sub-processes.
The executive team recognizes roles under pressure to change by three features:
- High repetition: The task occurs daily in a similar form.
- Clear definition of output: A second person can check whether the result is correct.
- Digital access: The information required is held in systems.
Case handling with a fixed review scheme changes, as does the preparation of decision papers. The role shifts from producing to reviewing. The person assesses the agent's output and handles the cases the agent does not resolve.
Management roles change too. A team lead will steer a mix of people and agents. Capacity planning and quality control need new standards for this. Performance assessment takes into account what a person achieves together with agents. The team lead thus becomes more accountable for results and distributes fewer activities. For span of control this means it is measured by the number of people and by the number of exceptions coming back from the agents.
Skills shift in the same direction. What is needed is the ability to formulate goals and quality criteria, to assess results and to handle exceptions. Whoever steers agents needs process knowledge and judgment. Programming skills are rarely required.
Which new roles does a company need for AI agents?
A company needs a named person for every agent who answers for its results, plus three further responsibilities covering process, quality and platform. Four responsibilities have to be filled for operation, in small units often by one person:
- Agent owner: The agent owner is accountable for the goal and performance of an agent and develops it further, much as a product owner does for a product.
- Process owner: The process owner decides which parts of the process an agent takes over and sets the autonomy level.
- Quality review: Quality review defines test cases and samples and measures output quality continuously.
- Platform support: Platform support is responsible for the agents' access rights and logging as well as their costs.
The assignment is binding. Without an agent owner, errors have no addressee and the agent loses the trust of the business units. Legal liability towards third parties lies with the company that deploys the agent. The internal assignment decides who notices and fixes errors. The article Delivering agentic AI describes the route from prototype into ongoing operation.
Which autonomy levels exist for AI agents?
Four autonomy levels, assigned per task, are sufficient for most fields of use. They range from a mere proposal to execution with exception handling and can be mapped to the common terms of oversight.
Human in the loop means that a person confirms every single decision before it takes effect. Human on the loop means that the agent acts on its own while a person monitors operation and can intervene at any time. The first form costs time per case, the second requires reliable metrics and a clear intervention rule.
Two criteria determine the level: the extent of damage an error causes and how measurable quality is. Tasks with high potential damage start at level 1 or 2. Moving up follows measurements. This requires a decision rule: the agent owner applies for the move with data, the process owner decides. Accountability thus stays with people, even when agents do most of the work.
As soon as an agent is technically capable of monitoring the performance or behavior of employees, the works council must be involved before launch under section 87(1) no. 6 of the German Works Constitution Act. This is frequently the case with agents that log work steps. How companies plan this consultation early is shown in Works council and AI implementation.
Implementation as a delivery task
Niehues Consulting supports organizations in this endeavor as a delivery partner using the Niehues Delivery Model. Marin Niehues works as a management consultant from the delivery perspective and builds organizations that deliver. His projects range from VC-funded scale-ups to DAX 40 groups, with assignments in 20+ countries on six continents.
- Understand: The roles and tasks that agents change are analyzed, and applications are prioritized by benefit. The guiding question is: who decides, who delivers, who uses?
- Build: Prototypes prove the benefit in the real process. The role model is developed in parallel, guided by the question of which roles and decision paths operation needs.
- Introduce: The organization sets responsibilities and enables staff until an internal or external team runs the agent. Here the question that counts is: who owns operation from day one?
How the whole route from the first application to regular operation is structured is described in the article From pilot to scale. The entry point is a free initial consultation of 30 minutes, in which roles and starting level of your first agent are clarified.
Frequently asked questions
Which jobs will agentic AI change first?
Activities with high repetition and a clear output change first, such as case handling with a fixed review scheme or research for decision papers. Roles shift from producing to reviewing. Those who process cases today will assess the agent's output and resolve the exceptions.
What is an agent owner?
An agent owner is the person accountable for the goal and performance of an AI agent. They develop the agent further, apply for a move to a higher autonomy level and are the point of contact for errors. The role corresponds to the product owner for a product.
Does a company need dedicated roles for AI agents?
Yes, accountability for every agent must be assigned to a person. At its core this means agent owner and process owner, with quality review added. Smaller companies bundle these responsibilities in a few people.
What is the difference between human in the loop and human on the loop?
With human in the loop, a person confirms every decision of the agent before it takes effect. With human on the loop, the agent acts on its own and a person monitors operation with the ability to intervene at any time. The second form requires robust metrics.
How much autonomy should an agent have at the start?
The start is at proposal or draft with approval. The lower the potential damage of an error and the better quality can be measured, the sooner the move to sampling or exception handling follows.
How does the manager's role change when steering agents?
The manager becomes more accountable for results and distributes fewer activities. They plan capacity for people and agents together and handle the exceptions that agents hand back. Span of control is therefore also measured by the number of these exceptions.
Who is liable for an AI agent's mistakes?
Towards third parties, the company that deploys the agent is liable. Internally, an agent owner should be named for every agent who notices errors and is accountable for fixing them. Individual liability questions are a matter for legal counsel.
Does the works council have to approve AI agents?
Yes, in Germany, as soon as an agent is technically capable of monitoring the performance or behavior of employees. This is frequently the case with agents that log work steps. Plan the consultation before launch.
How does the executive team recognize an agent's success?
By process metrics. Lead time and error rate are compared before and after deployment, along with the share of exceptions and the cost per case.
Let’s talk about your initiative.
In 30 minutes we clarify where your organization stands and which step will pay off most now.
Request a consultation30 minutes · free · no obligation



