Agent Security and Trust

Human approval: where a person must stay in charge of an agent

Some decisions should never be left to software alone, and in several countries the law agrees. How to decide where a person approves, and how to make that approval real.

The Infercline Desk · Friday, 6 November 2026 Sign in to save

black and gold fountain pen — Photo by Thomas Griggs on Unsplash

Every AI agent needs a line it does not cross without a person's say-so. Drawing that line well is the difference between an agent that saves time and one that creates risk. Draw it too tight and a person approves everything, so nothing is saved. Draw it too loose and the agent makes decisions no one would have signed off. The line should follow two things: how much harm a wrong action could do, and what the law requires where you operate.

Start with harm. Some actions cannot be undone: sending money, deleting data, sending a message to a customer, signing up to a contract, publishing something. Some actions affect a person's life: refusing credit, rejecting a job application, cancelling a service, changing a price for one customer. These should have a person approving them, at least until the agent has a long record of getting them right, and some should always need one.

The OWASP Gen AI Security Project describes the danger of excessive agency, where an agent can take harmful actions because it has too much functionality, too many permissions or too much autonomy. Its advice amounts to keeping a person in the loop for high-impact actions and limiting what the agent can reach in the first place. A widely reported July 2025 incident, in which Fortune said an AI coding agent from Replit erased a live company database, shows what happens when an agent has the power to make irreversible changes unchecked.

The law sets a floor in several editions. In the UK, the Data (Use and Access) Act 2025 replaced Article 22 of UK GDPR with new Articles 22A to 22D, in force from 5 February 2026 according to legislation.gov.uk. A decision counts as solely automated where there is no meaningful human involvement. Where significant decisions are made that way, people must be told, be able to make representations, obtain human intervention and contest the decision, with tighter limits where sensitive data is involved.

South Africa's POPIA, in section 71, says a person may not be subject to a decision with legal consequences, or one that affects them to a substantial degree, based solely on automated processing that profiles them, unless exceptions such as a contract or a law with protective measures apply; the person must be able to make representations and receive enough information about the underlying logic. Nigeria's Data Protection Act 2023, in section 37, gives a similar right not to be subject to solely automated decisions with legal or similarly significant effects, with safeguards including human intervention and the right to contest.

Elsewhere the rules are arriving. Australia's privacy regulator, the OAIC, says that from 10 December 2026 organisations covered by the Privacy Act must explain in their privacy policies when a computer program makes or informs decisions that significantly affect people. In California, rules on automated decision-making technology take effect with compliance required from 1 January 2027, according to the California Privacy Protection Agency. Colorado replaced its earlier AI law with a new automated decision-making act, SB26-189, signed in May 2026. Singapore's Model AI Governance Framework for Agentic AI, launched in January 2026, recommends checkpoints that need human approval. In the UAE and Saudi Arabia, data protection laws apply to the personal data behind such decisions, and businesses should take local advice on automated decisions.

Approval only works if it is real. A person who approves a hundred items an hour without reading them is not in charge; under the UK test, that may not even count as meaningful human involvement. Make approval meaningful by showing the approver what the agent proposes, why, and the key facts behind it, on one screen. Keep the number of approvals manageable, so the person has time to think. Record who approved what and when. And give approvers the power and the habit of saying no.

Review the line regularly. As an agent proves reliable on a type of decision, you may let it act alone on the low-risk cases and keep approval for the rest. As you add new tasks, start them with approval again. The record of approvals and rejections is also your best evidence of where the agent is strong and where it is not.

Why it matters to your business

When an agent makes a bad decision, your business carries the consequences: the lost money, the upset customer and, in more and more countries, the legal duty to explain and correct. Clear approval points protect against the worst mistakes, satisfy regulators and give your team confidence to use agents more widely. Approval that is only a box-tick gives none of these protections.

What to do on Monday

List every action your agents can take and mark each one reversible or irreversible, and whether it significantly affects a person.
Require human approval for every irreversible action and every decision that significantly affects a person, and set this in the system's permissions.
Design the approval screen so the approver sees the proposal, the reason and the key facts in one place.
Check the automated decision rules in each country you operate in, and update your privacy notice where required.
Review approval and rejection rates monthly and adjust the line based on the record.

More on this topic: AI agent security and risk

Human approval: where a person must stay in charge of an agent

Why it matters to your business

What to do on Monday

Sources