Judgment-in-the-Loop: The Human Role AI Cannot Automate

In the first article of this series, I showed what happens when AI agents operate on bad context: wrong answers about my own blog, legal liability for Air Canada, medical misinformation at Google-scale. In the second, I mapped the architectural gap where quality checks should exist but don’t. This article is about the human who stands in that gap.

The Blog That Writes Itself (Sort Of)

This blog runs on five custom AI skills I built with Claude Code. /research pulls sources, verifies claims, and drafts outlines. /publish handles formatting, cross-referencing, and deployment. /audit checks for stale data and broken links. /plagiarism runs originality checks. /diagram renders D2 diagrams to SVG.

On paper, the pipeline looks automated. In practice, every article passes through a bottleneck that no AI skill can replace: me, reading the draft and asking, “Does this reflect what I actually think?”

That question is not about grammar or formatting. It is about whether the AI’s output matches my understanding of the subject, my judgment about what matters, and my sense of what the reader needs to hear. The AI can write a paragraph about Data Governance. It cannot tell you whether that paragraph belongs in this article, whether it oversimplifies a point I spent two years learning the hard way, or whether it makes a claim I am not willing to stand behind.

The industry already has a phrase for this: “human-in-the-loop.” But that phrase describes a checkpoint, not a capability. It tells you that a person is involved. It says nothing about what that person brings. What they bring is judgment.

Beyond Human-in-the-Loop

Andrej Karpathy amplified the term “context engineering” in June 2025 to describe the technical discipline of filling an AI’s context window with the right information: relevant documents, examples, instructions, tool outputs. It is an important concept, and some practitioners have extended it to acknowledge that human judgment plays a role in selecting and prioritizing what enters the context window. Anthropic’s engineering team frames the human as an architect of a “limited attention budget,” and Redis notes that “human judgment remains critical in selecting, structuring, and prioritizing context.” But even these expanded definitions stop at the input layer. Context Engineering is about what goes into the prompt. It says nothing about whether the resulting output is correct, appropriate, or aligned with the actual goal.

“Human-in-the-loop” addresses a different piece of the problem, but it is underspecified. It tells you that a human reviews output before it ships. It does not tell you what qualifies that human to judge, what they should be looking for, or whether their involvement is a genuine quality gate or a compliance checkbox.

Prompt engineering is narrower still: it focuses on how you phrase instructions. That matters, but it is a small slice of the problem.

I use the term judgment-in-the-loop to describe what I see as the missing piece: the ongoing human responsibility of ensuring that the context an AI system operates in is correct, complete, and current, and that the system’s outputs are evaluated against domain knowledge the AI does not possess. This is not established terminology; it is a framework I am proposing to make the human role in AI-augmented work more precise than “human-in-the-loop” allows.

The person with judgment in the loop knows enough about the domain to recognize when AI output looks right but is wrong. They know what “good” looks like. They know which trade-offs are acceptable and which are not. They carry institutional memory, stakeholder relationships, and judgment born from experience.

This is not a job title. It is a capability, one that becomes more valuable as AI handles more of the execution.

	Human-in-the-Loop	Judgment-in-the-Loop
What it asks	”Does a person review this?"	"Does a person who knows the domain review this?”
Who qualifies	Anyone with access	Someone with domain expertise and institutional context
When it happens	A checkpoint before output ships	Continuously: before, during, and after AI operates
What they check	”Does this look right?"	"Is this actually right, given what I know about this domain?”
Failure mode	Rubber-stamping plausible output	Catching “almost right” before it becomes production-wrong
Value source	Presence	Judgment
Example	A junior analyst clicks “Approve” on an AI-generated report	A senior architect catches that the report uses last quarter’s pricing model

The Evidence: Experience as the Differentiator

The intuition that domain expertise matters more in an AI-augmented world has strong empirical backing.

The Age Split

A Stanford Digital Economy Lab working paper circulated in 2025 analyzed employment trends across AI-exposed occupations using ADP payroll data. The finding is stark: early-career workers aged 22 to 25 experienced a 16% relative decline in employment in these roles, while workers over 30 saw employment grow in the same categories.

The researchers titled their paper “Canaries in the Coal Mine.” The pattern it reveals is not that AI replaces humans broadly. It replaces humans who lack the judgment to direct AI effectively. Experience is not just correlated with survival in AI-exposed roles; it is the mechanism.

The Jagged Frontier

Ethan Mollick’s study, published in Organization Science in March 2026, gave 758 BCG consultants identical tasks with and without AI access. For tasks within AI’s capability boundary, below-average performers improved by 43% with AI. Above-average performers improved by 17%. AI is an equalizer for routine work.

But for tasks outside AI’s capability boundary, a different pattern emerged: consultants using AI were 19 percentage points more likely to produce incorrect answers than those working without it. AI gave them confident, plausible, wrong output, and they could not tell.

Mollick calls this the “jagged frontier”: AI capability is not a clean line. It is irregular, with pockets of excellence next to pockets of failure. Knowing the shape of that frontier in your domain is what judgment-in-the-loop means in practice.

The Wage Signal

The market agrees. PwC’s Global AI Jobs Barometer (one billion job postings analyzed) found wages in AI-exposed industries rising 2x faster than non-exposed ones, with a 56% premium for AI-relevant skills. The market is pricing in the ability to work with AI effectively, not the ability to do what AI does.

The Productivity Myth and the Capability Expansion Reality

There is a narrative that AI makes everyone 10x more productive. The evidence says otherwise.

A Fortune/NBER survey of nearly 6,000 CEOs, CFOs, and other senior executives across four countries, published in February 2026, found that 89% reported no labor-productivity impact over the last three years. This is the Solow Paradox repeating: the technology is everywhere except in the productivity statistics.

The METR randomized controlled trial, published in July 2025, tracked 16 experienced open-source developers on real tasks. With AI tools, they were 19% slower. But they believed they were 20% faster. The perception gap is as important as the speed gap: people overestimate AI’s contribution to raw throughput.

I do not find these results surprising. I built an alarm clock app this year. I had envisioned it for years: specific behaviors, specific interactions, a clear picture of what “done” looked like. I could never build it because I am not a mobile developer. When AI coding agents reached a certain capability threshold (after Opus 4.5 and 4.6), I could finally get it built. The AI wrote the code. My domain knowledge, knowing what I wanted, how it should work, what trade-offs were acceptable, was the irreplaceable input.

AI did not make me 10x faster at mobile development. It made mobile development possible for me. In many workflows, the first visible value may be capability expansion rather than raw throughput.

Jeremy Utley at Stanford’s d.school puts it precisely: “Zero times any number is still zero.” AI often amplifies existing judgment, domain knowledge, and direction more than it replaces them. If you bring a clear sense of what good looks like, AI amplifies that. If you bring nothing, you get nothing, or worse, you get confident nonsense.

One developer’s anecdote puts a face on the pattern. Jay Moreno, with 20+ years in the field, wrote about building a complete SaaS product alone with AI in four months. His reflection: “This wasn’t about AI replacing developers. It was about AI amplifying someone who already knew what to build.” This is a single self-reported account, not controlled evidence, but it matches what the BCG and METR studies show at scale: domain knowledge is the input that makes AI output useful.

I see this pattern in my own daily work. I build tools with Claude Code for my day-to-day workflows: data processing scripts, analysis pipelines, automation for repetitive tasks. Building with AI forces you to think in examples, not abstractions. You cannot delegate effectively to an AI agent without being specific about what you want, why you want it, and how you will know it worked. That specificity comes from experience.

The Judgment Gap in Numbers

The Stack Overflow 2025 Developer Survey quantifies the trust gap: 66% of developers say their biggest frustration with AI coding tools is “solutions that are almost right, but not quite.” 46% actively distrust AI output accuracy.

This is the judgment gap stated plainly. “Almost right” is often worse than wrong, because wrong is obvious. “Almost right” passes code review, ships to production, and fails at 2 AM when the edge case hits. Detecting “almost right” requires knowing what “exactly right” looks like, and that knowledge lives in human heads, not in training data.

What Judgment-in-the-Loop Looks Like

Judgment-in-the-loop is not a single activity. It is a set of overlapping responsibilities:

Evaluate: Assess whether AI output meets the actual requirement, not just the stated prompt. The gap between what you asked for and what you needed is where errors hide.

Validate: Cross-check AI-generated claims, code, and recommendations against domain knowledge. Does this match what you know to be true? Does this align with how your systems, your organization, your industry actually work?

Correct: Identify and fix errors that AI cannot detect. These are often errors of context: correct syntax but wrong business logic, accurate data in the wrong time frame, valid analysis applied to the wrong question.

Guide: Shape the AI’s operating context proactively. Choose what goes into the prompt. Decide which tools the agent has access to. Define the evaluation criteria. Set the boundaries.

Decide: Make judgment calls that AI cannot. Which trade-off is acceptable? Which stakeholder concern takes priority? When is “good enough” actually good enough, and when does it need to be perfect?

These are not purely mechanical tasks, though parts of each can be partially automated through validation rules, monitoring pipelines, and escalation design. The irreducible core is final contextual judgment in ambiguous or high-stakes situations: knowing when the automated checks are insufficient, when the edge case matters, when the trade-off requires a human call. That judgment comes from doing the work, making mistakes, and learning the difference between what looks right and what is right.

The Emerging Role Landscape

The market is beginning to formalize this judgment capability into roles. “Context Engineer” has appeared in job postings since late 2025, describing someone who designs and maintains the information architecture that AI systems operate within. “Director of AI Governance & Risk” is increasingly common, reflecting the need for senior judgment about where and how AI should be deployed. “Trust Engineer” describes someone who builds the validation and monitoring infrastructure that keeps AI systems honest.

The WEF Future of Jobs Report 2025 projects 39% of existing skill sets will be transformed by 2030. The skills that survive are judgment-in-the-loop skills. But Deloitte’s 2026 survey found that while 66% of leaders say human-AI interaction design matters, only 6% say they are leading at it. The gap between recognizing the need and building the capability is enormous.

This connects directly to the AI Governance framework I wrote about in February. The three lines of defense model, where the first line builds, the second line validates, and the third line audits, is a stewardship structure. Each line represents a different type of applied judgment at a different point in the AI lifecycle. AI Governance, as an organizational capability, is judgment-in-the-loop formalized at enterprise scale.

Do Next

Priority	Action	Why it matters
Immediate	Audit your current AI usage for judgment gaps. Where are AI outputs going to production without domain expert review?	”Almost right” output passing unchecked is your highest-risk failure mode.
This quarter	Identify who has judgment-in-the-loop by name, not by title. Who in your org actually knows what “good” looks like for each AI-augmented workflow?	Judgment-in-the-loop is a capability, not a role. You need to know who has it.
This quarter	Redesign AI-augmented workflows around evaluation, not speed. Measure quality of AI-assisted output, not just quantity.	The METR study shows speed gains are illusory. Quality gains require human judgment.
Next quarter	Invest in domain expertise development, not just AI tool training. Send people deeper into the subject matter, not just wider into the tooling.	AI multiplies domain knowledge. Zero times any number is still zero.
Next quarter	Build feedback loops from production outcomes back to AI system design. Track where AI output required human correction and why.	These correction patterns reveal the shape of AI’s jagged frontier in your domain.
Ongoing	Treat AI Governance as judgment-in-the-loop at organizational scale. Staff your second line of defense with people who have domain depth, not just compliance checklists.	Governance without domain expertise is theater.

The Human in the Gap

This three-part series started with a simple observation: AI agents fail when they operate on bad context. The architectural layer that should catch these failures largely does not exist yet. And the human capability needed to fill that gap, the judgment to evaluate whether context is correct, complete, and current, remains one of the most valuable skills in the AI-augmented workforce. Some stewardship tasks can be partially automated through validation and monitoring, but the final contextual judgment in ambiguous situations resists full automation.

Judgment-in-the-loop is not a temporary role that AI will eventually absorb. It is the role that AI creates. Every advance in AI capability expands the surface area where human judgment is needed: more powerful models produce more plausible output, which makes detecting errors harder, not easier. More autonomous agents make more decisions, which makes oversight more important, not less.

The value has shifted. It used to live in execution: who can write the code, process the data, generate the report. It now lives in judgment: who knows whether the code is correct, the data is relevant, the report answers the right question.

AI did not make me faster. It made me capable of things I could not do before. The difference is everything, and the ability to recognize that difference is what judgment-in-the-loop is about.

This article is related to The Practitioner’s Guide to AI Agents, a nine-part series on building, evaluating, and improving AI agents.