Anyone deploying AI agents, meaning systems that work through a task independently in several steps, quickly realises: the decisive question is no longer whether the tool is good enough. The decisive question is how clear the brief was and who ultimately takes responsibility for the result.
This is precisely what an interview with Iris Lorscheid, professor of data science, points to. Her argument: employees need genuine leadership strength when dealing with AI agents. At first this sounds surprising, because leadership is usually associated with personnel responsibility. In everyday work with agents, however, it describes a very practical competence – namely delegating tasks, assessing intermediate results and, if in doubt, stopping them.
Delegating is not the same as handing over
Managers know the pattern: anyone who passes on a task unclearly gets back an unclear result. With agents, this effect is immediately visible. A vaguely formulated brief leads to work that looks formally plausible but misses the goal – and that someone then has to correct laboriously.
The difference from human colleagues lies in the feedback channel. A team member asks questions when a brief is contradictory. An agent will still deliver a result if in doubt. That is why the effort shifts upstream: into describing the goal, the framework conditions and the boundaries.
What should be clarified before introduction
Before agents are integrated into development or business processes, a sober round on responsibilities is worthwhile. Experience shows that a few questions, answered bindingly, help more than an extensive rulebook:
- Scope of the brief: Which tasks may an agent take on – and which expressly not?
- Responsibility: Who approves the result and is accountable for it to clients, the specialist department or management?
- Data access: Which systems and information may the agent access, and with which rights?
- Escalation: What happens when an intermediate result is implausible – abort, query, manual takeover?
- Traceability: Is it documented which steps an agent has taken, so that a result remains verifiable later?
These points are not a formality. They determine whether an agent takes work off your hands or creates additional control work.
Control needs checkpoints, not mistrust
Leadership does not mean monitoring every step. It means setting sensible checkpoints. In software projects such checkpoints already exist: the review before the merge, the test suite, acceptance by the specialist department. Agents can be placed within this structure instead of bypassing it.
What matters here is distinguishing by risk. A suggestion for a text variant may pass through quickly. An intervention in production data, an invoice approval or a change to security-relevant code requires a human decision – even when the result appears convincing.
New requirements for employees
The shift from executing to instructing gives rise to a competence profile that many job descriptions do not yet reflect. What is needed is the ability to formulate a goal precisely. What is needed is professional judgement to assess a result without recalculating it line by line. And what is needed is the willingness to reject a result instead of waving it through under time pressure.
That is demanding because it presupposes professional depth. Anyone who cannot assess a task themselves cannot delegate it meaningfully either – to people no more than to software. For teams this means: agents do not replace onboarding, they increase the value of specialist knowledge when assessing results.
Organisational side effects
When individual employees begin assigning packages of tasks to agents, new workflows effectively emerge – often faster than process descriptions can keep up. This is not fundamentally problematic, but it should remain visible. Otherwise results depend on individuals and their informal approach, and no one can later reconstruct how a decision came about.
A simple organisational framework is therefore helpful: a named responsibility for the use of agents, a brief written definition of the permitted use cases and a regular look at where the use actually saves time. Without this feedback loop, the benefit remains an assumption.
Classification for web and software projects
For projects this means above all a shift in effort: less time executing, more time clarifying the brief, reviewing and accepting. Anyone who hooks agents into existing quality processes – with clear approvals, graduated rights and traceable steps – gains speed without additional risks. Anyone who leads them past these processes merely shifts the work into troubleshooting.