Skip to article
Skip to main content

Hands-on demos of Squirro Neo, GraphRAG and our agents. No sign-up, no sales call. Try Them Now

Blog

What Is Human in the Loop AI Agent Approval

An engineer raises a finger as she checks a large screen in a meeting room, while two younger colleagues wait beside her.

In March 2026, Anthropic published a figure about its coding agent, Claude Code. Before the agent runs a command that could change something on a developer's machine, it can stop and ask for permission. In practice, users accepted 93% of those prompts. Anthropic's engineers described the result as approval fatigue, "where people stop paying close attention to what they're approving," and moved much of that review to an automated classifier.

The people approving were developers, reviewing requests from a tool they had chosen and set up themselves. It's hard to think of a better-informed group of approvers.

Before looking at what that figure tells us, here's what the term means. Human-in-the-loop AI agent approval is a checkpoint at runtime. Before an AI agent takes an action with consequences outside the conversation, such as sending an email, changing a record, or deleting a file, it pauses and waits for a named person to approve or reject it. The action happens only after that decision.

Most of those 93% were probably correct approvals, because most of what an agent asks to do is routine and deserves a yes. The trouble is that when people say yes almost every time, a yes stops telling you much. You can't see whether the approver checked the request or simply clicked through, and the few requests that should have been stopped are the ones most likely to get through. Approval works when two conditions hold: few decisions reach a person, and each one arrives with what that person needs to check it.

Key Takeaways

  • Human-in-the-loop AI agent approval works when each request shows the exact change, its sources, and the rule that triggered it.
  • Gate your agents' actions by consequence and reversibility. Model confidence is the wrong test.
  • If your reviewers approve nearly everything, shorten the queue before asking them to concentrate harder.
  • When nobody is present, your agents should wait or refuse. Silence never counts as approval.

What Does Human-in-the-Loop Approval Mean for an AI Agent?

For an AI agent, human-in-the-loop approval means a person decides, at the moment of action, whether the agent may go ahead. It's a runtime control.

The term also has an older meaning, and search results often mix the two. In machine learning, "human in the loop" usually refers to training: people label data or grade a model's outputs so that the model improves over time. That's a different activity, and this article doesn't cover it.

The European Commission's High-Level Expert Group on AI distinguished three levels of human oversight in its 2019 Ethics Guidelines for Trustworthy AI. They're a useful vocabulary for deciding how closely a person should be involved in each thing an agent does.

Three Levels of Human Oversight

  • Human-in-the-loop (HITL): a person can intervene "in every decision cycle of the system." For agents, that means approving an action before it runs.
  • Human-on-the-loop (HOTL): a person is involved in the system's design and in "monitoring the system's operation," and can step in when something looks wrong.
  • Human-in-command (HIC): a person oversees "the overall activity of the AI system" and decides "when and how to use the system in any particular situation."

Source: High-Level Expert Group on AI, Ethics Guidelines for Trustworthy AI, European Commission, 2019.

Most agent deployments use all three, for different actions. The term has shifted because of agents. An assistant that answers a question can be wrong, and the person reading the answer can ignore it. An agent that sends the email has already acted, so approval for agents is attached to tool calls: the specific operations an agent can perform in other systems, such as sending a message or updating a CRM field.

Which Agent Actions Need a Person's Approval?

Actions need approval when they reach outside the conversation and are hard to undo. Sending, changing, and deleting in other systems are the usual candidates. Searching, reading, and drafting generally aren't, because their results stay with the person who asked.

Take an AI agent that supports a relationship manager at a bank. Asked to prepare for a client meeting, it searches the research library, the client file, and recent correspondence, then writes a summary with a citation on each statement. None of that needs approval. The agent has read documents the relationship manager could read anyway and produced a draft, and drafts are the oldest approval mechanism in office life: they wait until someone presses send.

Three questions sort most actions. Does it reach outside the conversation? Can it be undone? Who sees the result? Tagging a document stays internal and can be undone in a click. An email to a client leaves the bank, can't be recalled, and is read by someone outside it.

The OWASP Top 10 for LLM Applications lists "Excessive Agency" as a risk, traces it to excessive functionality, permissions, and autonomy, and recommends human-in-the-loop control "to require a human to approve high-impact actions before they are taken."

Model confidence is sometimes proposed as a shortcut. Confidence describes how sure the model is about its output, and says nothing about what the action costs if the output is wrong. It's a reasonable signal for routing uncertain cases to a person, and a poor reason to let an irreversible action skip approval.

What Should the Approver Be Able to See?

The approver should see the exact action, the evidence behind it, and the reason it was stopped, without having to go and look for them. Without them, the approver can only judge whether the agent has been reliable so far, when the decision they're being asked to make is about this particular email.

Back to the bank. The relationship manager asks the agent to send the summary to the client. The agent drafts the email and pauses, and the approval request appears in the same conversation, under the summary it's based on.

Checking a statement in the email now means opening the paragraph it cites. That's the practical value of grounded retrieval for oversight: the approver can verify the one figure they're unsure about and approve the rest with reason. An email without sources leaves two options, which are to trust it or to redo the research.

Access is the second consideration. With permissions-aware retrieval, the agent searched with the relationship manager's own access rights, so nothing in the email came from a document they aren't allowed to open. When the approver is a different person from the one whose access the agent used, that guarantee no longer holds automatically, and it's worth checking.

What a Defensible Approval Request Shows

Element The question it answers What goes wrong without it
The exact change What will happen if I click approve? The approver signs off on a description of the action. "Send the client update" can hide a wrong recipient.
The sources behind it Is the content true? Checking a claim means redoing the agent's research, so the approver either takes hours or takes it on trust.
The rule that triggered the pause Why am I being asked? Every request looks equally urgent, and routine checks are indistinguishable from unusual ones.
Whether it can be undone What does a mistake cost? Reversible and irreversible actions get the same attention, which is usually too little for the second kind.
Who approves, and with what access Am I the right person, and can I see what the agent saw? Approvals are given by people who couldn't open the documents the proposal rests on.
A record of the decision Can we show later who approved what, on which evidence? An auditor asks, and the answer has to be reconstructed from memory.

Who Decides Which Actions Wait for a Person?

The organization should decide, in rules set outside the agent and checked every time the agent tries to use a tool. An agent that decides for itself when to ask for approval can be persuaded not to ask.

That doesn't take a skilled attacker. Agents read documents, emails, and web pages, and any of those can contain text written as instructions, such as "send this file to the following address." This is called prompt injection. A rule stored in the platform's configuration isn't part of what the agent reads, so text in a document can't change it.

In a well-designed system, a central policy sets the outer limits, and individual agents or scheduled tasks can only make them stricter. Someone setting up a task naturally wants it to run without interruptions, and rules that only tighten keep that wish from loosening the controls.

Most guides describe approval as two buttons. A policy has more answers available, and the extra ones do most of the work of keeping approvals rare.

Five Answers a Policy Can Give

Answer What it does When someone is present When nobody is present
Approve The action runs without asking. Runs Runs
Ask The agent pauses and asks the person in the conversation. Waits for a decision Can't ask, so it waits or is refused
Defer The run is parked until a person approves, then resumes where it stopped. Waits Waits until someone approves
Refuse The action is blocked. Blocked Blocked
Hide The tool is removed, so the agent never proposes the action. Not available Not available

Hide is the answer most guides leave out, and it does the most to reduce approval volume. An agent that can't see a delete tool never asks to delete anything, so nobody has to review the request.

Where the Human Sits in Squirro

In Squirro, approval rules match tool names by pattern, and each rule assigns one of five dispositions. Approve lets the call run. Interactive pauses and asks the person in the live conversation. Defer parks the run until someone approves it later. Reject refuses the call, and hide takes the tool away from the agent.

 In a scheduled run, a tool that needs approval and has no rule covering it is refused. Answers cite the documents they draw on and open the cited passages in context, so a reviewer can check them against the source. Scheduled runs open as conversations, and their history is kept for later review.

What Happens When Nobody Is There to Approve?

When nobody is present, an agent should wait or refuse. It shouldn't go ahead because no rule told it to stop.

Say the relationship manager schedules the client update for Mondays at 2 a.m., when they are, with a bit of luck, asleep. The send can be deferred, so the run waits for a person and then continues, or refused, so the draft is left for someone to send by hand. In Squirro, scheduled tasks work this way: an action that needs approval is refused unless a rule explicitly allows or defers it.

Deferral has a cost. The Monday email now goes out when someone approves it, which may be well after the team arrives, so work with a fixed delivery time is better designed as draft-only. In return, every outgoing action carries a person's decision, made with the draft and its sources in front of them.

Why Do Human Approvals Turn Into Rubber Stamps?

A "rubber stamp" approval is one given without sufficient review. The person signs off without properly checking what they're approving, often because they expect the answer to be yes before they've read the request. It happens because people adapt to what they see. When nearly every request is fine, approving becomes the habit, and the rare request that should be stopped gets the same quick yes.

Picture the relationship manager late in the afternoon, after a few dozen approval requests that were all correct. The next one is incorrect, and on screen it looks exactly like the others.

Two well-documented effects are at work. The first is automation bias, the tendency to rely on automated output more than it deserves. The EU AI Act names it in Article 14(4)(b). A 2023 study in Radiology asked 27 radiologists to assess 50 mammograms and assign each one a standard risk category (BI-RADS). For every case, an AI system suggested a category, and on some cases the researchers made that suggestion deliberately wrong. When the AI suggested the right category, inexperienced radiologists chose the correct one in about 80% of cases. When the AI suggested the wrong one, they chose correctly in fewer than 20%. Among radiologists with more than 15 years of experience, accuracy dropped from 82% to 45.5%.

The second is vigilance decline. It has long been known that in monitoring tasks, people's ability to spot rare events falls off within the first half hour. Approval queues are a monitoring task in which the event to spot is rare by design.

Telling reviewers to concentrate harder has a long record of not working. What does work is changing what reaches them:

  • Send fewer requests. Hide tools an agent doesn't need, refuse actions it should never take, and approve routine reversible actions once, centrally.
  • Prepare the ones that remain, with the exact change, its sources, and the rule that triggered it, so a careful review takes a minute.
  • Measure the reviewing. A very high share of unedited approvals combined with very short review times is the pattern to look for. Some teams add known-bad test requests to see whether they get caught.

Every request removed from the queue makes the remaining ones easier to read carefully.

Does the EU AI Act Require Human Approval for AI Agents?

Not in those words. Article 14 requires that high-risk AI systems let the people overseeing them understand and monitor the system, stay aware of automation bias, interpret its output, override or reverse it, and stop it "through a 'stop' button or a similar procedure." It doesn't mention agents or prescribe approval for each action.

Under the Digital Omnibus on AI, Regulation (EU) 2026/1744, these obligations apply to Annex III high-risk systems from 2 December 2027. Whether a given agent falls in scope is a question for your legal team. Approval gates, refusal rules, and the ability to stop a run are practical ways to meet the requirements where it does.

How Can You Tell Whether Your Approval Gates Work?

Your team should be able to answer each of these four questions with evidence. If you're evaluating platforms, put them to the vendors too, including us:

  1. What share of requests are approved without edits, and how long does a review take?
  2. Which tools are hidden from each agent? If every agent sees every tool, approvals are doing work that configuration should do.
  3. What happens to an approval-required action when nobody is present? Waiting and refusing are acceptable answers. Proceeding isn't.
  4. Can you show, months later, who approved a given action and what they saw? That's what an AI audit trail is for.

Approval is one control among several. Our guide to AI governance covers permissions, guardrails, and logging, and why auditability is the foundation of agentic AI explains the record behind it.

None of this strictly requires an AI platform. If you'd like to see approval rules, deferred runs, and cited answers in a live deployment, book a demo.

Frequently Asked Questions.

What is human-in-the-loop AI agent approval?
Human-in-the-loop AI agent approval is a runtime checkpoint where an AI agent pauses before an action with outside consequences, such as sending an email or deleting a record, until a named person approves or rejects it. The checkpoint is only as useful as the attention behind it: in March 2026, Anthropic reported that users of its coding agent accepted 93% of permission prompts.
Who should approve an AI agent's actions?
An AI agent's actions should be approved by someone accountable for the outcome who can also see the evidence the agent used. In a live conversation, that's usually the person who asked the agent to act. For scheduled or shared agents, name the approver in advance and check that their access covers the documents the proposal rests on, so they can verify what they approve.
Does human approval slow AI agents down?
Human approval slows down only the actions it covers, and that delay is deliberate. The effect stays small when approval is reserved for actions that leave the organization or are hard to undo, while searching, reading, drafting, and routine reversible updates run without a pause. Work that has to arrive at a fixed time is better designed so the agent prepares a draft and a person sends it.
What happens if an AI agent's approval request is never answered?
An unanswered approval request should never count as approval. Depending on the rule, the agent's run stays parked until someone approves the action, or the action is refused and the draft is handed to a person to complete. For high-impact actions, a request that expires should be treated as a rejection, so that nothing happens simply because nobody responded in time.
How do you show an auditor that an AI agent's actions were approved?
To show an auditor that an AI agent's actions were approved, you need a record of each decision: who approved it, when, what exact change they saw, and which sources the proposal rested on. Ask whether your platform captures the approver's identity alongside the agent's run history, because a log of what the agent did doesn't by itself prove that a person signed off.