Every company that gives AI agents real work faces the same design question: when should a person say yes first, and when should the agent simply act? Ask for approval on everything and the agent saves no time, while the people approving learn to click without reading. Ask for approval on nothing and the first bad action reaches a customer before anyone notices.
The two common answers have names. Human in the loop and human on the loop describe where a person sits relative to the agent's actions. This guide defines both, compares them, explains why approving everything tends to fail, covers what regulators expect, and gives a practical way to decide which actions belong where, what an approval should show, and how to move work from one mode to the other as trust grows.
Key takeaways
- In the loop means approval before each action. On the loop means monitoring with the power to intervene or stop.
- Choose by consequence: reversibility, who is affected and how much is at stake.
- An approval is only as good as what the approver sees: the action, the evidence and the consequence.
- Move actions from in to on the loop when the record shows approvals are almost always accepted unchanged.
What do human in the loop and human on the loop mean?
Human in the loop (HITL) means the AI proposes and a person decides before anything happens. The agent drafts the email, prepares the change or suggests the payment, and waits. Nothing leaves until someone approves.
Human on the loop (HOTL) means the AI acts within set limits while a person supervises. The person sees what the agent is doing or has done, gets alerts when something looks wrong, and can pause, correct or reverse it.
A third mode completes the picture. Human out of the loop means the agent acts with no live supervision, and people only review results later, if at all. It suits very low-stakes, easily reversed work, and almost nothing customer-facing.
Anthropic's guide to building agents describes the in-the-loop pattern simply: agents "can then pause for human feedback at checkpoints or when encountering blockers" (Anthropic). The design work is choosing those checkpoints well.
Human in the loop vs human on the loop, side by side
| Human in the loop | Human on the loop | Human out of the loop | |
|---|---|---|---|
| When the person acts | Before each action | During or after, when needed | Rarely, in later review |
| Speed | Limited by the approver | Agent speed, with pauses on alerts | Agent speed |
| Best for | External, costly or irreversible actions | Frequent, reversible, internal actions | Trivial, fully reversible tasks |
| Main risk | Approval fatigue and rubber-stamping | Missing a problem until it has happened | Errors nobody sees |
| What it needs | Clear approval screens with evidence | Good logs, alerts and a stop switch | Strong limits and sampling |
Most real systems mix all three. The question is never which mode the company uses. It is which mode each kind of action gets.
Why does approving everything fail?
Approval on every step feels safest, and it is where most teams start. It tends to decay for a reason psychologists and regulators both recognise: people over-trust automated systems that are usually right. The EU AI Act names this directly. Its human oversight article requires that overseers stay aware of the tendency of "automatically relying or over-relying on the output" of a system, known as automation bias (EU AI Act, Article 14).
Anthropic's February 2026 study of how people supervise its coding agent shows the shift in practice. New users with fewer than 50 sessions used full auto-approve "roughly 20% of the time"; by 750 sessions, that rose to "over 40% of sessions." Experienced users also interrupted the agent more, in around 9% of turns against 5% for new users. Anthropic reads this as people moving from approving each step to watching and stepping in when something goes wrong: from in the loop to on the loop (Anthropic).
The same study argues against blanket rules. It warns that requiring "humans to approve every action" will "create friction without necessarily producing safety benefits," and recommends focusing on whether people can effectively monitor and intervene. It also found that on the most complex tasks, the agent stopped to ask for clarification more than twice as often as humans interrupted it. Good agents can be designed to ask at the right moments.
What do the rules expect from human oversight?
For high-risk systems, Article 14 of the EU AI Act sets out what effective oversight must allow. Overseers must be able to understand the system's limits, correctly interpret its output, "decide, in any particular situation, not to use the high-risk AI system" or to disregard, override or reverse its output, and interrupt it "through a 'stop' button or a similar procedure." For certain biometric identification uses, it goes further and requires that a result be "separately verified and confirmed by at least two natural persons." Under the current timeline these obligations apply to Annex III high-risk systems from 2 December 2027.
Singapore's Model AI Governance Framework for Agentic AI, released by IMDA in January 2026, takes a practical line for agents in general. It advises organisations to "define significant checkpoints or action boundaries that require human approval," to train people to evaluate those requests "and audit these approvals," and to complement this with automated monitoring and alert thresholds (IMDA).
Read together, the message is consistent: approvals at the checkpoints that matter, monitoring everywhere else, and the power to stop at all times.
How do you choose between in the loop and on the loop?
Decide per action, using two questions: can it be undone, and how far does it reach?
Anthropic's study offers a sense of proportion: in its sample of tool calls made through its API, "only 0.8% of actions appear to be irreversible." Most agent actions can be undone. The few that cannot, such as a sent email, a payment or a deleted record, are where approval earns its cost.
Add a third question for business actions: does it commit the company to someone outside it? A quote, a delivery date or a discount is a promise. Promises belong in the loop. Our AI agent accountability guide explains why the approver then answers for that promise.
What should a good approval show?
An approval is a decision, and a decision needs evidence. A screen with a draft and a button invites rubber-stamping. A useful approval shows five things:
- The action. Exactly what will happen: the email text, the record change, the amount.
- The reason. Why the agent proposes it, tied to a goal or request.
- The evidence. The messages, documents or figures it relied on, with links.
- The consequence. What changes for the customer, the money or the schedule if approved.
- The choices. Approve, edit, reject with a reason, or hand to someone else.
Reasons for rejection matter most for improvement. "Wrong price" or "customer is on hold" teaches more than a silent no. The setting approval rules playbook has templates.
How to move an action from in the loop to on the loop
- Start everything external in the loop.
For the first weeks, every outbound or financial action waits for approval.
- Track acceptance per action type.
Record how often each type is approved unchanged, edited or rejected, and why.
- Set a graduation bar.
For example, a high share approved unchanged over a set number of cases, with no serious errors. Write the bar down before you start.
- Graduate with a safety net.
Move the action to on the loop with logging, alerts on unusual cases and a sample reviewed each week.
- Demote on evidence.
If errors rise, move it back. Graduation is a decision, not a one-way door.
When many agents graduate at different speeds, keeping track becomes an operating task of its own, covered in AgentOps.
Common approval design mistakes
- Approval on everything, forever. It wastes time and trains people to stop reading.
- No evidence on the screen. An approver who cannot see the facts cannot oversee anything.
- On the loop with no alerts. Supervision that depends on someone reading every log will not happen.
- No stop switch. If pausing an agent needs an engineer, nobody will pause it in time.
- The wrong approver. Approvals routed to whoever is available instead of whoever owns the decision.
- Counting approvals as safety. A high approval count says little. Edits, rejections and incidents tell you more.
How to measure approval design
| Measure | What it tells you |
|---|---|
| Accepted unchanged, edited, rejected | Where the agent is ready to graduate and where it is not |
| Median time to approve | Very fast approvals of complex actions suggest rubber-stamping |
| Rejection reasons | The context the agent is missing |
| Interventions on the loop | How often supervisors pause or correct, and why |
| Incidents by mode | Whether actions on the loop are as safe as those in it |
IMDA's advice to audit approvals is worth taking literally: once a quarter, sample approved actions and check whether the approver had what they needed to decide.
How GeniOS approaches human in the loop AI
GeniOS keeps people in the loop for anything that leaves the company. It reads where a company's work happens, such as email, calendar, chat and the business tools already in use, read-only to start. It forms the company's goals from its own emails, decks and meetings, and checks every team's pending work against them.
What reaches a person is one situation, what to do, why, one owner and a date, which is the evidence an approval needs. The owner answers Delegate, Done, Later or Not relevant. Delegate hands work to a person or an AI agent, and nothing is sent until a human approves.
An item closes only on proof, such as the email that went out, and GeniOS learns from each outcome. Over time that record shows which kinds of work hold up, which is how a company moves step by step toward the autonomous company while people keep the final say.
Worked example: drawing the loop for a buyer week
The names, dates and numbers in this example are made up to explain the idea. Vaultex makes safety footwear in Greater Noida and was GeniOS's first design partner. It is setting approval rules for the agents that help Harsh in sales and Anushka in production.
The week's work includes a quote to Bata for 12,000 pairs of S3 safety shoes at $12 a pair, about $144,000, before Friday's vendor review, with Bata asking for 60-day payment terms. Reebok samples are on Line 2. Adidas's vendor audit needs IS 15298 test reports. Steel toe-cap prices have just risen 8%.
| Action | Mode | Why |
|---|---|---|
| Bata quote and payment terms | In the loop, Harsh approves, finance confirms terms | External promise on a $144,000 order, set on a cost that just changed |
| Reebok ship date reply | In the loop, Anushka confirms | A delivery commitment that depends on Line 2 |
| Adidas test report pack | In the loop at first | External, and the auditor sees what is sent |
| Updating the order tracker | On the loop | Internal and easy to correct |
| Reminders to colleagues | On the loop | Internal, low stakes, reversible |
On the Bata approval, Harsh sees the draft at $12 next to the supplier email about the 8% rise and rejects it with the reason "old cost." The revised quote goes out after the terms are settled, and the item closes when the sent email exists. After two months, the Adidas-style document packs have been approved unchanged nearly every time, so Vaultex moves them on the loop with a weekly sample. Quotes stay in the loop.
Conclusion
Human in the loop and human on the loop are both tools, and the skill is matching each action to the right one. Keep approval for actions that are external, costly or hard to undo, and make every approval show the evidence. Put everything else on the loop, with logs, alerts and a stop switch. Then let the record decide what graduates, and audit approvals so they stay real. That is how a company gets speed from its agents and keeps the final say. To see approvals built around evidence and proof, see pricing or book a call.
Sources
- Anthropic, Building effective agents, 19 December 2024
- EU Artificial Intelligence Act, Article 14: Human Oversight, Regulation (EU) 2024/1689
- Anthropic, Measuring AI agent autonomy in practice, 18 February 2026
- IMDA Singapore, Model AI Governance Framework for Agentic AI, first published 22 January 2026