In short: When AI systems produce more variants, more decisions land on fewer people. I call this decision density. Human in the loop therefore means more than clicking “approve” at the end of a workflow. Leaders have to define which decisions a system may make on its own and which must always land with a person.
Ten campaign variants in the time it used to take to produce one sounds like relief at first. What actually happens looks different. Every one of those variants needs someone to sign off on it, and that job ends up on one person’s desk.
People expected something else. When systems started compiling reports, formatting presentations and combing through competitive data, hopes ran high: fewer rounds of alignment, more time for what really matters. Those routines have in fact disappeared. What remains, and is even growing, are the moments when someone has to judge whether a result is actually right.
The more tasks a system takes on, the more decisions land on the people who carry the responsibility.
When an AI rolls out ten campaign variants in the time it used to take to produce one, someone has to decide which one fits the brand, which one is risky and which one to pull before it goes public. A leader makes that call. No algorithm makes it for them.
A concrete example
In practice
I saw this recently in a workshop with a marketing team that had just introduced a new copywriting tool. Within a week, the team was producing twice as many variants per campaign as before. The real bottleneck shifted to approval. The team lead suddenly faced twenty copy variants instead of two, each held to the same standard: on brand, legally sound, factually correct. The tool had saved her time. It had also handed her more responsibility, concentrated in a single spot.
Where leadership gets squeezed today
I call this decision density. There are more moments in which a person has to judge whether an automatically generated statement is accurate, whether a system is operating at the edge of its competence, or whether a campaign that runs flawlessly from a technical standpoint is still wrong. Budgets show the same pattern. A system can run ten allocation scenarios in seconds, but how much risk a company is willing to take still has to be decided by someone who lives with the consequences. That applies to campaign budgets just as much as to the question of which customer data a system is allowed to process in the first place.
How to shape this density
Organizations that take this seriously deliberately make room for it: clear escalation paths, defined limits within which systems may decide on their own, and designated moments that must always land with a human. At its core, this is decision architecture. It doesn’t emerge on its own. It is designed, in workshops, in guidelines, in the way approval processes are set up.
Skip that design work and the density grows anyway, just without any order. Approvals pile up with individuals, responsibility gets distributed at random, and burnout quickly becomes a side effect of technological acceleration.
That is the heart of my book “Die neue Betriebslogik” (The New Operating Logic). Marketing is a particularly good place to watch this shift, because technology has worked its way deeper into day-to-day work here than in almost any other business function. What shows up here first will catch up with other areas in the coming years, from sales to product development.
If I may leave you with one question: take a close look at where decisions are piling up in your own organization, at the points where someone ends up taking the heat when something goes wrong. That is exactly where leadership in the age of AI begins, and exactly where it pays to start.
What does human in the loop really mean?
The term human in the loop is often used as if it were about a person clicking “approve” at the end of a workflow. That falls short. The German researcher Jörg von Garrel describes the approach as a control loop of human and machine, in which the human monitors, decides and intervenes as soon as a situation exceeds the machine’s limits. The human stays involved across the system’s entire lifecycle, from development to ongoing improvement.
This is exactly where decision density comes from. Fewer people handle routine cases themselves. More people have to teach an agent what a good case looks like, review its output and decide the exceptions. That role needs time on the calendar and a place in performance reviews. If it gets done on the side, the agent’s quality drops within a few months.
Every leader should know one trait of agents: an agent that’s wrong rarely sounds unsure. In McKinsey’s Superagency report, about half of employees name AI inaccuracy and cybersecurity risks as concerns. That argues for well-placed review steps wherever a mistake really costs something.
What should leadership teams review now?
- Clarify which decisions in your AI-supported workflows have a defined owner, and which simply ended up with someone.
- Measure the number of approvals per week for each accountable person, before and after a system goes live.
- Talk to the people who give the most approvals today about which of them an agent could handle on its own within firm limits.
Frequently asked questions
What is decision density?
I use decision density to describe how AI systems take over routine work while producing more output that a person has to decide on. Those decisions concentrate on a few people, often the manager.
Does human in the loop mean every output gets reviewed?
No. A person stays part of the control loop, but how deeply they review depends on the risk. Low-risk routine cases can run without individual review if samples and metrics show the quality holds.
Further reading: How Much AI Agent Autonomy Is Right? Three Levels · The AI-First Operating Model
Sources: von Garrel, J. (2026): Analyse der Mensch-KI-Interaktion: Vertrauen, Akzeptanz und der Human-in-the-Loop-Ansatz. Springer Gabler · McKinsey (2025): Superagency in the workplace
