Guide · Human-in-the-loop AI

Human-in-the-loop for AI agents, when no one can review everything

AI now issues refunds, changes orders, sends emails and ships code, thousands of times a day. "A person checks it" can't mean a person reads every one. Here's how to keep people in charge of AI's work, and keep quality control, at that volume.

Human-in-the-loop (HITL) means a person reviews, approves or corrects an AI system's work before it counts. When AI acts thousands of times a day, nobody can review it all, so the checking happens in three layers, the way a good QA department works:

  • Sign-off first for anything that can't be undone, like a large refund, a payment or a deletion. It waits for a person.
  • Simple rules block what's clearly out of bounds, with no judgment call needed.
  • Spot checks on a fair random sample of everything else. Because the sample is fair, the results tell you how often the AI gets it wrong across all its work, for example "fewer than 1% of refunds were wrong, with 95% confidence."

What human-in-the-loop means

The phrase describes where a person sits relative to an automated decision. There are three common positions:

In the loop

A person approves before it happens

The action waits for a human decision. Highest control, slowest. Right for anything that can't be undone: large refunds, payments, deletions, production deploys.

On the loop

A person monitors and can step in

The system acts on its own; people watch the stream and intervene. Fast, but only as good as what people actually notice.

Out of the loop

No person involved

Fully autonomous. Acceptable only where mistakes are cheap, reversible and bounded by hard limits.

Over the loop

People own the rules and the evidence

People set the policy, review a verifiable sample, and sign the irreversible cases. This is the position that scales to AI agent volume.

Human in the loop vs human on the loop

The two are often confused. The difference is when the human decides: before the action runs (in the loop) or while it's running and afterward (on the loop).

Human in the loopHuman on the loop
When the person decidesBefore the action runsDuring or after; the system acts first
SpeedWaits for a personFull machine speed
Best forIrreversible, high-value or high-risk actionsHigh-volume, reversible actions
Main failureApproval queues pile up; people rubber-stampProblems slip by unnoticed in the volume
Evidence it workedA signed approval for each actionUsually none, unless the monitoring is measured

Most AI agent deployments need both, split by risk: in the loop for the small share of actions that can't be undone, and on the loop for the rest. The weak spot is "on the loop": watching a stream isn't the same as checking it. A random sample with a stated confidence bound turns on-the-loop monitoring into evidence.

Why classic human-in-the-loop breaks with AI agents

  • Volume. An agent can take more actions in an hour than a team can read in a week. "Review everything" quietly becomes "review nothing."
  • Rubber-stamping. After a few hundred approvals, people stop looking. Without a way to measure attention, nobody knows it has happened.
  • Queues as bottlenecks. If every action needs approval, the agent saves no time and people start approving in bulk.
  • AI reviewers. Swapping the person for another model makes the watcher share the failure modes of what it watches. Text inside an agent's output can even talk an AI reviewer into approving it.
  • Volume attacks on caps. A per-action limit ("refunds up to $25") can be beaten by doing many small actions.

A human-in-the-loop design that scales

The answer isn't more reviewers. It's putting people exactly where they add the most, and making every human decision count as evidence.

allow
Within policy: runs immediatelyDeterministic rules check each action in milliseconds. It runs and goes on a tamper-evident ledger.
hold
Risky or irreversible: waits for a signatureA named person approves or denies, with their own signature on the record. No answer by the deadline means deny, never approve.
block
Out of bounds, or a breaker has trippedThe action never runs. Speed limits ("more than 20 refunds or $2,000 in 10 minutes") trip a circuit breaker that pauses the whole category until a person resets it.

Then, after the fact: a sample people can't game

Everything that ran is eligible for review. A secret seed, committed before the period starts, selects a random sample; reviewers answer one plain-language question per action (Yes, No, or Not sure); the seed is revealed afterward so anyone can recompute exactly which actions should have been checked. Known-bad planted test cards are mixed in, so the report shows whether reviewers were really looking.

The key number: about 299 clean random checks prove a category's problem rate is below 1% at 95% confidence (a one-sided Clopper-Pearson bound). That holds whether the category had ten thousand actions or ten million. Human review grows with the confidence you want, not with volume. In practice, plan on about 350 random checks per category each month: that proves the category is under 3% every month and under 1% every quarter, even when reviewers find the handful of problems a good agent produces; 299 prove under 1% in a month only if they find none.

How much human time it takes

Agent actions per monthRandom checks a month (5 categories, <1% a quarter)Review time at 2 min a card
10,0001,750≈ 58 h
1,000,0001,750≈ 58 h
100,000,0001,750≈ 58 h

About 60 hours a month is a part-time job for one reviewer, and it stays the same at a hundred million actions. Held actions add time on top, so the policy decides how many actions are truly worth a signature. The review-time calculator lets you try your own numbers, and the oversight staffing calculators plan a staged rollout and compare firms from 10 people to a million.

Human-in-the-loop examples by industry

IndustryHold for a signatureSample afterwardSpeed limit
Customer supportRefunds over the order total, credits over policyReplies, small refunds, address changesRefunds per 10 minutes
Finance operationsPayments over a threshold, new vendor bank detailsInvoice coding, reconciliationsPayment total per hour
Coding agentsProduction deploys, migrations, deletesCommits, config changesCommands per minute
Insurance claimsDenials and large payoutsRoutine approvalsPayout total per day

Human-in-the-loop checklist for AI agents

  • Sort actions by reversibility. Anything that can't be undone gets a hold, not a sample.
  • Write the rules down as code, not prompts. A model shouldn't decide whether a model's action runs.
  • Add speed limits, not just per-action caps, so volume can't slip past them.
  • Make "no answer" mean deny, with a deadline on every hold.
  • Sign every human decision, so the record shows who approved, not just that someone clicked.
  • Sample at random with a committed seed, so neither the agent nor the business can steer what gets checked.
  • Plant test cases to measure each reviewer's catch rate.
  • Report a bound, not a feeling: "below 1% at 95% confidence" beats "we spot-check."

Frequently asked questions

What is human-in-the-loop AI?

Human-in-the-loop AI is a design in which a person reviews, approves or corrects an AI system's output or action before it takes effect. For AI agents, it usually means holding high-risk actions for a person's approval and sampling the rest for human review.

What is the difference between human in the loop and human on the loop?

In the loop, a person approves before the action runs. On the loop, the system acts on its own while a person monitors and can intervene. Most AI agent deployments use both: in the loop for irreversible actions, on the loop (backed by a verifiable random sample) for the rest.

How many AI agent actions should a human review?

Enough to support the confidence you need, not a fixed percentage. About 299 clean random checks prove a problem rate below 1% at 95% confidence, regardless of how many actions there were. Because a good agent still makes a few mistakes, plan on about 350 checks per category each month: that proves under 3% every month and under 1% every quarter. Irreversible actions should be held for approval individually.

Does the EU AI Act require human oversight?

Yes, for high-risk AI systems. Article 14 requires them to be designed so people can effectively oversee them, including understanding outputs, deciding not to use them, and intervening or stopping them. Article 12 requires automatic record-keeping (logs).

Does human-in-the-loop slow AI agents down?

Only for the actions you choose to hold. Everything within policy runs at machine speed; sampling happens after the fact, so it adds no delay. The design goal is to hold the few actions that can't be undone, not everything.

Can an AI be the human in the loop?

No. An AI reviewer shares the failure modes of the system it reviews and can be manipulated by the content it reads. AI can help by flagging actions for extra human review, but it should never be the one that clears them.

Try it

Be the human in the loop for an AI support agent

Approve a held refund from a phone, trip a circuit breaker, review a verifiable sample with planted tests, then try to cheat the audit.