Which AI agent actions need human approval?

What does "human in the loop" mean for AI agents?

An AI agent can plan and carry out steps across systems: read emails, look up records, draft documents, update data and trigger actions. "Human in the loop" means a person reviews and approves certain actions before they take effect. The design question is which actions. Approving everything makes the agent pointless. Approving nothing hands the business's commitments to software.

A useful rule: the agent may do anything that is reversible and internal on its own; it must ask before anything that is costly, external, sensitive or permanent.

Which agent actions need human approval?

Action type	Examples	Suggested control
Read and analyse	Search records, summarise documents, compare invoices to orders	Autonomous, logged
Draft and prepare	Draft replies, prepare a payment file, propose a journal entry	Autonomous; output waits for review
Internal, reversible updates	Tag a ticket, update a status, route a request	Autonomous within limits; sampled review
External communication	Send an email to a customer, reply to a supplier	Human approval until quality is proven, then sampled for low-risk templates
Money and commitments	Release a payment, place an order, offer a discount, sign a contract	Always human approval, following the approval matrix
Master data and access	Change supplier bank details, create users, change limits	Always human approval by someone other than the requester
Decisions about people	Screen candidates, assess credit, flag an employee	Human decision; agent only assists
Irreversible actions	Delete records, cancel contracts	Human approval, plus a recovery plan

Adjust the limits to the business's own appetite for risk. A value threshold often works well: below it, the agent proceeds and a sample is reviewed; above it, a person approves. Our guide on controls before automating approvals shows how to build such a matrix.

What does the EU AI Act require on human oversight?

The AI Act's general application date was 2 August 2026, when the majority of its rules started to apply, including the Article 50 transparency rules (European Commission, AI Act implementation timeline). The human oversight duty in Article 14 applies to high-risk AI systems, and its start has moved. Amendments published in the Official Journal on 24 July 2026 set 2 December 2027 for stand-alone high-risk AI systems and 2 August 2028 for high-risk AI embedded in products (European Parliament Legislative Train; European Parliament press release).

Many business agents, such as those handling invoices or customer queries, are unlikely to fall into the high-risk categories, but check your own use case. Article 14 is still a good design checklist. It asks that people overseeing a high-risk system can understand its capacities and limits, stay aware of "automation bias", correctly interpret its output, decide not to use it or override it, and interrupt it through "a 'stop' button or a similar procedure" (AI Act, Article 14).

How do you make human approval meaningful?

• Show the evidence. The approver sees what the agent saw, what it proposes and why, not just an "Approve" button.
• Keep the queue small. If people approve hundreds of items a day, they stop reading. Use thresholds and sampling so attention goes where risk is.
• Watch for automation bias. Track how often approvers change or reject proposals. A rate near zero may mean nobody is really checking.
• Separate duties. The person who asked the agent to act should not be the only approver of high-value actions. See segregation of duties in digital workflows.
• Log both sides. Record the agent's proposal and the human decision together.
• Keep a stop switch. Someone can pause the agent at once and fall back to the manual process.

How do you widen an agent's autonomy safely?

Start with approval on every external or financial action. After a set period, review the log: where proposals were accepted unchanged at a high rate and errors were rare, raise the threshold or move to sampling for that action type. Where approvers often corrected the agent, keep the gate and improve the process or instructions. Widen one action type at a time, and write each change down.

Update, 27 October 2026: NIST published design tenets for agent identity management. They say agents should have an identity and access lifecycle distinct from human users, that access should be bounded to the minimum scope needed for a task, and that separation of duties should prevent an agent from assuming conflicting privileges (NIST, 27 October 2026). These complement the approval gates above: limits on what an agent can reach, plus approval for what it may commit.

Linkgrow's AI agent workflows service designs approval points like these into agent use cases. Check your starting point with the AI agent readiness checklist or the free gap finder.

Frequently asked questions

Does every AI agent action need human approval?

No. Reading, analysing and drafting can usually run on their own with logging. Approval belongs at the point of commitment: money, external messages, master data, decisions about people and irreversible actions.

When does the EU AI Act human oversight rule apply?

Article 14 applies to high-risk AI systems. After the amendments published on 24 July 2026, it applies from 2 December 2027 for stand-alone high-risk systems and 2 August 2028 for those embedded in regulated products.

What is automation bias?

The tendency to accept a system's output without enough scrutiny. It is a key reason to keep approval queues small and to track how often approvers change the agent's proposals.

Can an AI agent approve its own actions?

It should not approve actions that move money, commit the business or change access. A separate, accountable person should approve those.