May 23, 2026 · 6 min read

The human in the loop needs a job description.

Trustworthy AI depends on designing what the human is actually there to do. “Human in the loop” sounds reassuring — but until you can say which human, in which loop, with what authority, and to do what, the human is decoration.

In the past year I have watched “there is a human in the loop” end more conversations than any other sentence in this field. Someone raises a risk, the phrase lands on the table, next slide.

It shows up in governance decks and risk assessments wherever someone needs to say: do not worry, a person is still involved. What it hides is that nobody has decided anything yet.

Which human?
In which loop?
At what moment?
With what information?
With what authority?
To do what?
To do — which of these, exactly?
ApproveCorrectInterpretEscalateOverrideTeachTake responsibilityExplain it to someone affected

Fig. 1 — Six questions the phrase quietly skips. Each one is a design decision in disguise.

If the questions go unanswered, the human is not a safeguard. It is decoration.

It is a bit like putting a driving instructor in the passenger seat — but removing the pedals, hiding the speedometer and giving them no way to take over. Technically, there is a human in the car. Practically, the human cannot do much.

This matters more as AI becomes agentic. When AI only drafts text, the loop is simple: a person reads, edits and decides what to use. But when AI starts retrieving enterprise knowledge, recommending decisions, triggering workflows, updating records and coordinating between systems, the loop becomes part of how the organisation actually runs.

Simple loopAI drafts text — a person reads, edits, decides
ThenIt retrieves enterprise knowledge
ThenIt recommends decisions
ThenIt triggers workflows & updates records
And nowThe loop is part of the operating model

Fig. 2 — As the AI is allowed to act, the loop stops being a review step and becomes infrastructure.

The loop becomes part of the operating model. It has to be designed like any other service moment, with the same care we would give a handover between two people. And in the context of the enterprise processes, procedures, and work instructions.
01What happens before the AI acts?
02What context does it receive?
03What is it allowed to do?
04When does it stop?
05When does it ask?
06When does it escalate?
07What does the human see?
08What evidence is shown?
09How does the human challenge the output?
10What happens after the human intervenes?
11Does the system learn from it?
12Who owns the outcome?

Fig. 3 — The same questions a service designer asks of any moment of truth.

The research on the future of work lands in the same place: once AI can act, the constraint is no longer the model, it is how humans delegate, supervise, and evaluate. One large workplace study now tracks the ratio of humans to agents on a team as a management number. The loop is a management job. And the skills it asks for are the ones we spent twenty years calling soft: briefing well, judging output, giving feedback that changes behaviour. They turn out to be the hard ones. Nobody runs a team of twelve people with a checkbox.

Without that design work, organisations create an illusion of control. The human is present on the diagram, but powerless in the moment.

Placed in the loop as…But…
A reviewer placed at the end of the processgiven too little time to review properly.
An operator asked to approve a recommendationnot shown the source evidence behind it.
A manager accountable for the decisionunable to inspect how the system reached it.
A colleague expected to correct the AIgiven no way to feed the correction back.
A user told there is human oversightnever given a meaningful path to contest it.

Fig. 4 — Five humans in five loops, each prevented from mattering.

The AI oversight research is clear about this: a human who is present but powerless does not make the system safer. They make it look safer, which is worse, because the organisation relaxes.

Call it what it is: human-shaped governance language.

So the role of the human has to be explicit. “A human” is a placeholder, and in practice several very different jobs hide behind the same three words.

01Commissioner

Defines the task, intent, constraints and standard of quality before the AI begins.

Decision right — sets the brief
02Domain expert

Knows whether the output reflects reality, not just whether it sounds plausible.

Decision right — judges what is true
03Operator

Handles exceptions, edge cases and moments where the AI cannot continue safely.

Decision right — intervenes
04Reviewer

Checks quality, risk, fairness, compliance or tone before anything reaches a user.

Decision right — gates release
05Teacher

Turns corrections, examples and feedback into better future performance.

Decision right — improves the system
06Accountable owner

Decides what the AI may do, where the boundary sits, and when it must be stopped.

Decision right — owns the outcome

Fig. 5 — Six roles, not one. Each needs its own interface, training, decision rights and measure of success.

Confuse them, and you get a reviewer asked to commission, or an owner asked to operate — people held responsible for work they were never equipped to do. The six roles also cluster around three moments. The commissioner holds the brief. The domain expert and the operator hold the boundary, where the system reaches what it cannot judge and hands back. The reviewer and the accountable owner hold the close, where someone's name goes on the outcome. The teacher is what makes next month better than this one.

Workers do not need to be talked into this. When researchers asked fifteen hundred people across a hundred occupations how much agency they wanted to keep as agents arrive, the most common answer was a working partnership, with the human keeping a real say. On nearly half the tasks, people wanted more agency than the technologists considered necessary. People are asking for the loop to take them seriously.

A smoke alarm is useful because its role is clear. It detects a signal and alerts a human. The human checks the room, decides whether it is burnt toast or a real fire, and acts.

Clear role A smoke alarm
  • Detects a signal
  • Alerts a human
  • You check the room
  • You decide: toast, or fire
  • You act
You always know when it acts, and why.
Undefined agency An alarm that also acts
  • Opens the windows
  • Calls the fire brigade
  • Cuts the electricity
  • Locks the kitchen door
— and no one knows when, or why.

Fig. 6 — That second alarm is closer to the world we are entering with AI agents. The more a system can act, the more carefully the human role around it has to be designed.

From August 2026, European law makes this concrete. The AI Act requires organisations deploying high-risk systems to place oversight with people who have the competence, the training, and the authority to intervene, with the support to actually do it. Read as governance, that is a compliance line. Read as practice, it is a job description. The most useful test emerging in supervision is the override rate. A person who never overrides the system is not overseeing it. A person who always overrides it has been handed a system that was not ready. Between the two sits a number that says the role is real. Teams run it like any operating figure: low-stakes actions run alone, medium-stakes get a quick verification, high-stakes wait for domain authority.

AI raises the floor of a task; the human raises the ceiling: the judgement, context, care and accountability that make the work worth trusting. A person who clicks approve without understanding the output raises nothing. Neither does one asked to monitor forty agent actions an hour.

Designed poorly humans become part of the machinery. That is not the future of work I am interested in.

This is why human-in-the-loop design belongs close to service design: moments, roles, handovers, accountability. The system has to work operationally and humanly as well as technically.

If you own an AI-assisted process, try writing the job description this week. One page: which of the six roles this person holds, what they see before they decide, what they can refuse, what happens when they refuse it, and how often you expect them to. If the page is hard to write, the loop is not designed yet.

I am starting to collect these one-pagers across sectors. If you write one, I would like to see it.

Sources: Mollick — Management as AI superpower, Microsoft Work Trend Index 2026, Stanford Digital Economy Lab — Future of Work with AI Agents, Fink — Human Oversight under Article 14 of the EU AI Act, EU AI Act, Article 14, Institute for Systems Integrity — From Human-in-the-Loop to Human-with-Agency

Andreas Conradi · June 2026