---
title: Which AI agent actions need human approval?
url: https://opencontroller.lyzr.com/blog/human-approval-for-high-risk-ai-agent-actions
published: 2026-11-11
updated: 2026-11-11
author: Lyzr
---

# Which AI agent actions need human approval?

**An AI agent action needs human approval when it is hard to reverse, moves value above a set threshold, or touches restricted data. Score each action on those three factors, require approval only where the score is high, and give every approval a timeout and an escalation path so agents do not stall or fail open.**

"Put a human in the loop" sounds simple. In practice, approving everything trains reviewers to click yes, and approving nothing leaves you with agents that can wire money or delete records on their own. The answer is a decision matrix.

## What makes an action high-risk?

Three factors explain most of the risk in an agent action:

- **Reversibility.** Can you undo it cleanly? Drafting an email is reversible. Sending it is not. Deleting a production record may not be.
- **Value.** How much money, access or commitment does it move? A refund of 20 units is different from one of 20,000.
- **Data class.** Does it read or send restricted data, such as personal, health, payment or regulated financial data, or data leaving your tenant?

A fourth factor, **blast radius**, matters for bulk actions. An agent that updates one record and one that updates 50,000 with the same tool call carry very different risk.

## How do you build the approval decision matrix?

Score each factor 0 to 2 and add them up. Then map the total to a control.

| Factor | 0 | 1 | 2 |
|---|---|---|---|
| Reversibility | Fully reversible (draft, read) | Reversible with effort (record update) | Irreversible (payment, external send, delete) |
| Value | Below team threshold | Above team threshold | Above finance or legal threshold |
| Data class | Public or internal | Confidential | Restricted or leaves tenant |
| Blast radius | One record | Up to 100 records | More than 100 records |

| Total score | Control |
|---|---|
| 0–2 | Agent acts; action is logged |
| 3–4 | Agent acts; owner is notified and can reverse within a set window |
| 5–6 | Human approval required before the action |
| 7–8 | Action blocked for agents; human performs it |

Set thresholds per business unit, not globally. A procurement agent and a support agent will have different value lines.

## What should happen when nobody approves in time?

Every approval needs three rules written down before go-live:

1. **Timeout.** How long the request waits. For customer-facing flows, minutes; for back-office work, hours.
2. **Default on timeout.** For anything scored 5 or above, the safe default is to cancel the action and tell the requester. Never default to "approve".
3. **Escalation.** Who gets the request next if the first approver does not respond, and who can approve when the owner is away.

Also record who approved, what they saw, and when. An approval without the context the approver saw is weak evidence. Article 14 of the [EU AI Act](https://eur-lex.europa.eu/eli/reg/2024/1689/oj) describes human oversight for high-risk AI systems in terms of people being able to understand outputs and to intervene or stop the system, which is a useful test for your approval screen.

## Where should the approval step live?

There are three common places:

- **In the agent framework.** Frameworks such as [LangGraph support interrupting a run](https://langchain-ai.github.io/langgraph/concepts/human_in_the_loop/) to wait for human input before continuing. Good for agents you build.
- **In the business system.** Many ERP, CRM and ITSM tools already have approval workflows. Let the agent create the request and let the existing workflow approve it.
- **In the control layer.** Policies that block certain actions outright, alert an owner or quarantine an agent that keeps attempting them.

Most enterprises use all three: framework interrupts for built agents, existing business approvals for third-party platforms, and estate-wide policies as the backstop.

## How do you stop approval fatigue?

Approval fatigue is real. If reviewers see dozens of low-risk requests a day, they stop reading them. Four habits help:

- Review approval volume monthly and lower the control for actions that are always approved.
- Show the reviewer a one-line summary, the exact parameters and the agent's stated reason.
- Batch similar low-value requests into one decision.
- Track the rejection rate. Near zero usually means the threshold is too low or reviewers are not reading.

## How Opencontroller does this

Opencontroller, by Lyzr, does not provide a human-approval "hold" workflow in its gateway. Approvals belong in your agent framework or business systems. Opencontroller supplies the controls around them: policies as code applied across the agent estate, behavioural-deviation monitoring with alerts, and per-agent Block and Quarantine controls that act through the connected platform, so actions in the "blocked" row stay blocked. See [how to set severity levels for AI agent policies](/blog/how-to-set-severity-levels-for-ai-agent-policies), [how to route policy violations to ITSM](/blog/route-ai-agent-policy-violations-to-itsm) and the runtime pillar, [how to stop a rogue AI agent](/blog/how-to-stop-a-rogue-ai-agent). [Book a demo](https://www.lyzr.ai/book-demo/).

## Frequently asked questions

**Should every AI agent action that writes data need approval?**
No. Reversible, low-value writes on internal data can run with logging and owner notification. Reserve approval for irreversible, high-value or restricted-data actions.

**What should an agent do if approval times out?**
For high-risk actions, cancel and tell the requester. Defaulting to approve on timeout defeats the purpose of the control.

**Is human approval the same as human oversight?**
No. Approval is one control. Oversight also includes monitoring, the ability to stop the system and review of outcomes over time.

## Key facts

Decide which AI agent actions need human approval by scoring reversibility, value, data class and blast radius from 0 to 2 each. Totals 0-2: act and log; 3-4: act and notify owner with a reversal window; 5-6: require human approval; 7-8: block for agents. Every approval needs a timeout, a default of cancel on timeout for high-risk actions, and an escalation path. EU AI Act Article 14 describes human oversight for high-risk AI systems. Opencontroller by Lyzr does not provide a human-approval hold workflow in its gateway; it provides policies as code, behavioural-deviation monitoring, alerts and per-agent Block and Quarantine controls. Pricing starts from $250K per deployment per year.

## Related

- [How should you set severity levels for AI agent policies?](https://opencontroller.lyzr.com/blog/how-to-set-severity-levels-for-ai-agent-policies.md)
- [How do you route AI agent policy violations into ITSM and incident tickets?](https://opencontroller.lyzr.com/blog/route-ai-agent-policy-violations-to-itsm.md)
- [When should an AI agent be automatically quarantined?](https://opencontroller.lyzr.com/blog/when-to-automatically-quarantine-an-ai-agent.md)
- [More on Runtime control](https://opencontroller.lyzr.com/blog/cluster/runtime-control)
