> 📥 The full article is available as a PDF (Danish). Download it at the top of the page.
The hardest problem in enterprise AI right now is a question of decision rights. It just doesn't look like one.
A few weeks ago I sat in a room with a leadership team doing everything right. Strategy. Pilots across multiple departments. Copilot rollout. Governance committee. Budget line. Two new data scientists. Signed contract with an enterprise vendor.
I asked them one question. When the model is in the loop, who decides what?
Silence.
It's not unusual. I see it everywhere I work — retail, finance, public sector, life sciences. What kills AI projects is almost always the same. A question nobody thought to ask out loud.
This article is about that question. And about the small frame I now bring into every room to force it onto the table.
"Human-in-the-loop" doesn't mean anything anymore
The phrase shows up in every AI policy I've read in the past two years. Board notes. Vendor pitches. EU AI Act compliance documents.
It tells you there's a human somewhere in the system. That's it.
It doesn't tell you what the human is doing. When. With what authority. With what information. Or what happens when the human disagrees with the machine.
I've seen organizations celebrate "human oversight" while the human was a junior employee clicking approve on 200 model outputs per hour, because the alternative was missing their SLA.
That's rubber-stamping with productivity targets. Calling it oversight doesn't change it.
The fix is to be concrete about what judgment we're actually asking humans to exercise.
The Judgment Loop
Four questions. One picture. Designed to be answered by a leadership team in under ninety minutes, and revisited every time the system changes.
For any AI-supported decision in your organization:
1. What is the model deciding?
A prediction is not a decision. A credit risk score is a prediction. The decision is whether to approve the loan, what rate to offer, whether to escalate the case. Conflate the two and accountability quietly disappears.
2. What is the human deciding?
Almost no one answers this explicitly.
Does the human decide whether the model's output is accepted? Whether the case warrants further investigation? Whether to override based on context the model can't see?
Each of those is a different job. Different information. Different training. Different time on task.
"The human reviews it" is not an answer.
3. What does the human need to make a good decision?
This is where most deployments break.
If the job is catching model errors, the human needs to see the model's confidence, which features it weighted, and ideally a comparison case.
If the job is adding context the model lacks, that context has to be visible. Not buried in a separate system.
If the job is taking accountability, the human needs the authority to actually override.
Most "human-in-the-loop" implementations give the human a button and a stopwatch. Call it what it is. Theater.
4. What happens when the human and the model disagree?
This is where you find out if the loop is real.
If disagreement triggers a learning event — logged, reviewed, fed back into the model and the process — you have a Judgment Loop.
If disagreement triggers a delay, a red KPI, or a quiet override that doesn't go anywhere, you have a compliance ritual. A classic weakness of compliance functions.
That's it. Four questions. Draw them on a whiteboard. The arrow runs both ways between human and model. That's the whole point.
What the loop reveals
The first time a leadership team works honestly through the Judgment Loop, three things usually happen.
First: silence. They don't actually know the answer to question 2. They thought they did. They've written it down somewhere. But when they have to articulate it for the specific decision in front of them, the words don't come.
Next: a fight. Usually between the function that owns the process and the function that owns the technology. Both thought they had decision rights. Neither had written it down.
Finally: relief. Because the Loop translates "are we being responsible enough with AI?" — an abstract worry — into a concrete piece of design work that can be done, reviewed, and revised.
That's what good judgment in operations looks like. More concrete human involvement, not less.
Why this isn't in your AI strategy deck
There's a reason the Judgment Loop isn't in your AI strategy.
AI strategy, as it's sold to most organizations, optimizes for two things. What we'll do with AI. And how we'll avoid getting sued for it.
The first lives in roadmap slides. The second lives in policy documents.
Neither contains the operational reality. Who decides what, with what information, under what authority, the moment the model delivers.
That reality lives in process design. And in most companies process design is done by people who weren't in the AI strategy meeting.
That's the gap. The Judgment Loop is the tool I use to close it.
Tools change. Judgment accumulates.
The frontier model you deploy today is obsolete in eighteen months. The agent framework you standardize on gets replaced. The vendor you picked gets acquired or pivots.
What you build around your AI — who decides what, with what information, with what authority — that accumulates.
It sharpens with every disagreement you log. Clarifies with every edge case your team works through. Over time it becomes the real competitive advantage.
Not the model. The judgment around the model.
That's what I mean when I say AI, Built Human. The Loop is the system. Take the humans out and you don't have autonomous AI. You have irresponsible AI.
One thing to try this week
Before your next AI deployment review, take one production-ready use case. Just one.
Walk your team through the four Judgment Loop questions, out loud, in the room.
If you can answer all four in clear sentences, you're in the minority that reaches production with something durable.
If you can't, you just found the work that actually needs doing.
That's not a failure. That's the start.
📥 Download the full article as a PDF (Danish) at the top of the page.
🎓 Want the Judgment Loop applied to a specific deployment? Book a strategic advisory engagement or an AI Readiness Sprint. First conversation is free.
📬 Subscribe to the newsletter "AI, Built Human" on Substack — bi-weekly insights on the judgment behind the machine.
Stefano Vincenti · AI Advisor & Trainer · aitrainer.dk · External Lecturer, IT University of Copenhagen · Cofounder & CTO BotTellMe · Partner, TryZone