Human in the loop AI - the decision every exec gets wrong

We built this blog as an agent loop, and the only places we kept a human were the three this piece argues for.

Where you place a human in an AI workflow is an operating model choice, and most leaders make it on instinct instead of on the two things that should decide it.

Short answer: Human in the loop is an operating model decision. The right question is where a human adds value that the loop cannot. Place humans where judgment, accountability, or irreversibility demand it. Everywhere else, a human in the loop is just a bottleneck.

Why is "do we trust the AI?" the wrong question?

The trust frame leads executives to add humans everywhere, which is how a faster AI workflow ends up no faster at all.

The cleanest evidence comes from engineering teams. Faros AI's analysis of AI adoption found that high-adoption teams merged 98% more pull requests, while PR review time rose 91% and the org-level DORA metrics did not improve. The AI wrote more code. The humans still had to review it, and review became the constraint. The throughput of the loop did not change, because a loop moves at the speed of its slowest link, and the slowest link was the person.

That is Amdahl's Law applied to org design. You can make one stage of a process arbitrarily fast, and the stages you did not touch decide the result. Drop a human review gate into the path of a high-volume agent and you mostly move the bottleneck onto a salaried calendar.

So the trust question quietly becomes an operating model question. The real decision is where a human changes the outcome enough to justify the time the gate costs.

Does a human in the loop actually catch the errors?

Often it does not. The assumption that a human reviewer reliably corrects an AI's mistakes is the part most governance decks never test.

Two findings should unsettle anyone designing on that assumption. The IAPP's review of human oversight in production AI concluded that human-in-the-loop oversight "does not reliably correct algorithmic errors and may even exacerbate errors." And in a study reported in AI & Society, 450 clinicians using an intentionally biased diagnostic tool saw their accuracy fall from 73% to 61.7%. They deferred to the machine rather than overriding it. The technical term is automation bias, and a tired reviewer clicking approve is its natural habitat.

There is a deeper problem under that one. Even a fully alert human is not a stable instrument. Kahneman, Sibony, and Sunstein's work on noise documents how much the same person's judgment varies on the same case from one day to the next. So a review gate can add cost, add delay, and add a second source of error all at once. The human is worth inserting only when the value they bring outweighs all three.

Where does a human actually add value the loop cannot?

Three things. A human earns a place in the loop when the decision needs judgment, when someone has to be accountable for it, or when it cannot be undone.

Judgment is the work that needs context the model cannot reach: a relationship, a read on the market, an ethical line, a bet on where the business is going. That is also the direction the org itself is heading. McKinsey's work on the agentic organization describes the shift plainly: humans move from executing activities to owning and steering end-to-end outcomes. The person stops doing the step and starts owning the result.

Accountability is the decision a regulator, a customer, or a board will hold a named person responsible for. The agent can draft it. Someone still has to sign it, and that signature does not transfer to a model.

Irreversibility is the one most teams overweight in theory and underweight in practice. Anthropic's study of 998,481 production tool calls found 73% already had a human in the loop, while only 0.8% of the actions were irreversible. Most loops are gating the wrong things. They put a human in front of work that could simply be checked after the fact and rerun if wrong, and the gate buys delay instead of safety.

How do you decide which AI decisions need a human?

Score each decision on two axes and let the score pick the model. The practitioner framing from MindStudio sets the autonomy boundary by the consequence of being wrong and the reversibility of the action. Low-stakes and reversible is a candidate for full autonomy. High-stakes or irreversible warrants a gate.

	Reversible (cheap to undo)	Irreversible (cannot be undone)
Low consequence	Full autonomy. Let the agent run and audit a sample after. Example: drafting internal copy, tagging tickets, enriching a list.	Human on the loop. The agent acts, a person monitors the stream and can halt it. Example: sending routine customer emails, posting scheduled content.
High consequence	Human on the loop, with logging and a fast rollback. Example: adjusting bids, reprioritizing a queue, updating a forecast.	Human in the loop. The agent proposes, a named person approves before anything commits. Example: pricing changes, hiring or firing, contracts, anything legal or financial that sticks.

The distinction between the two oversight models matters because they cost different amounts. Human in the loop means the agent waits for approval before it acts, so the agent never moves faster than the reviewer. Human on the loop means the agent acts and a person supervises, able to intervene, which keeps the human's judgment in the system without making them the bottleneck. The Faros data is what happens when you use the first model where the second would do. Most workflows live in the off-diagonal cells, and most of those want a human on the loop, not in it.

Want the version of this you can hand to your team? We wrote up the decision as a one-page skill. Get the skill

What happens when you remove the human entirely?

You can over-correct just as expensively in the other direction. The instinct after reading the bottleneck data is to pull humans out wholesale, and the clearest cautionary tale is Klarna.

Klarna replaced 700 customer service agents with AI, handled 2.3 million conversations in the first month, and cut response time from 11 minutes to under 2. Then it reversed course. CEO Sebastian Siemiatkowski's own account: "We focused too much on efficiency and cost. The result was lower quality." Customer service is exactly the off-diagonal case the matrix flags. High consequence for the relationship, low reversibility once a customer feels processed by a machine. The math worked. The judgment axis did not.

So the failure mode is symmetric. Put humans everywhere and you get the Faros bottleneck. Pull them out everywhere and you get the Klarna reversal. The discipline is placing the human on the two-by-two.

What do the companies getting this right do differently?

They treat human placement as a designed process, and it shows up in their results. McKinsey's State of AI 2025 found that 65% of AI high performers have defined human-in-the-loop validation processes, against 23% of everyone else. The high performers have not removed humans. They have decided, on purpose, where the humans go.

This is no longer an edge consideration. Gartner's April 2026 CEO survey of 469 executives found 80% expect AI to force high-to-medium overhauls of their operational capabilities, and that the share expecting automation to stay limited to specific tasks falls from 54% today to 13% by the end of 2028. As more of the work runs inside agent loops, the human placement decision stops being a one-off setting and becomes the operating model itself. It pairs directly with where you draw the line between routine and judgment across each function and with which workflows you hand to an agent first.

Frequently asked questions

What is human in the loop AI and when do I need it? Human in the loop AI is a workflow where an agent proposes an action and a person approves it before it takes effect. You need it when a decision carries high consequence and is hard to reverse, like pricing, hiring, or anything legal or financial. For low-stakes, reversible work, it usually just adds delay.

What is the difference between human in the loop and human on the loop? Human in the loop means the agent waits for a person's approval before acting, so the agent moves at the reviewer's pace. Human on the loop means the agent acts and a person supervises, able to step in or halt it. On the loop keeps human judgment in the system without making the person the bottleneck.

How do you decide which AI decisions need human approval? Score the decision on two axes: the consequence of being wrong and how reversible it is. High consequence or irreversible decisions warrant an approval gate. Low-stakes, reversible decisions are candidates for full agent autonomy. MindStudio's framework sets the autonomy boundary on exactly these two factors.

Does human in the loop slow down AI workflows? It can, badly. Faros AI found that teams with heavy AI adoption merged 98% more pull requests while review time rose 91%, with no org-level throughput gain. A loop runs at the speed of its slowest link. Put a human review gate in front of a high-volume agent and the human becomes that link.

What happens when you remove humans from the loop entirely? You risk a quality collapse on decisions that needed judgment. Klarna replaced 700 service agents with AI, cut response times sharply, then reversed course because quality fell. Removing humans works for low-stakes, reversible tasks. It backfires on high-consequence relationships and irreversible calls.

Get the framework

We turned the consequence-by-reversibility matrix into a one-page skill: the two axes, the four oversight models, and a worked example for each cell, so you can place humans on real workflows instead of by instinct. Get the skill

If you would rather not download anything, we write about building AI-native operating models, and you can join the newsletter to get the next piece.

Sources

1. Faros AI, "The AI Productivity Paradox" (2025). https://www.faros.ai/ai-productivity-paradox
2. McKinsey, "The State of AI in 2025" (November 2025). https://www.mckinsey.com/capabilities/quantumblack/our-insights/the-state-of-ai
3. Gartner, "CEO Survey on AI and Operational Capability" (April 2026, 469 respondents). https://www.gartner.com/en/newsroom/press-releases/2026-04-23-gartner-survey-reveals-80-percent-of-ceos-say-artificial-intelligence-will-force-operational-capability-overhauls
4. Anthropic, "Measuring AI Agent Autonomy in Practice" (2025). https://www.anthropic.com/research/measuring-agent-autonomy
5. Fortune, "Klarna AI, humans, and return on investment" (May 2025). https://fortune.com/2025/05/09/klarna-ai-humans-return-on-investment/
6. AI & Society (Springer), study on automation bias in clinical diagnosis (2025). https://link.springer.com/article/10.1007/s00146-025-02422-7
7. Kahneman, Sibony, and Sunstein, "Noise: A Flaw in Human Judgment" (2021). https://www.researchgate.net/publication/362590233_Noise_A_Flaw_in_Human_Judgment_by_Daniel_Kahneman_Olivier_Sibony_and_Cass_R_Sunstein_Little_Brown_Spark_2021_464_pp_US_3200
8. MindStudio, "Classify AI Agent Actions by Risk" (2025-2026). https://www.mindstudio.ai/blog/classify-ai-agent-actions-by-risk
9. McKinsey, "The Agentic Organization" (2025). https://www.mckinsey.com/capabilities/people-and-organizational-performance/our-insights/the-agentic-organization-contours-of-the-next-irrigate-paradigm-for-the-ai-era
10. IAPP, "Human in the loop in AI risk management: not a cure-all approach" (2024-2025). https://iapp.org/news/a/-human-in-the-loop-in-ai-risk-management-not-a-cure-all-approach

Christopher Kliebenstein builds AI-native operating models and agent workflows for founders and operators at Kliebenstein AI Studio.