Human in the loop: when a person must approve, and how the request reaches them
Approval steps protect customers only if they are rare, specific and strong. Decoupled authentication lets an agent pause and ask a person for a recorded decision.
"Human in the loop" is easy to promise and hard to do well. Two questions decide whether it means anything: when a person must approve, and how the request reaches them.
When. For high-risk AI systems, the EU AI Act requires design that lets people effectively oversee the system: understand its limits, stay aware of automation bias, override its output and stop it. In the UK, the FCA's Consumer Duty expects firms to deliver good outcomes for retail customers and avoid foreseeable harm, which applies whatever technology a firm uses. In practice, approval belongs at the steps where a mistake is costly or irreversible: a new payee, an amount above a threshold, a change to standing instructions.
How. An agent often works when the customer is not looking at a screen. Decoupled flows solve this. With OpenID's Client-Initiated Backchannel Authentication, the agent's application asks the customer's identity provider to authenticate them, and the customer approves on a separate device, such as their banking app. The answer comes back by polling, notification or push, and is recorded.
Making approvals meaningful:
- Ask rarely. Constant prompts train people to click yes.
- Show the specific risk: the amount, the payee, what is new.
- Bind the approval to exactly what was shown.
- Measure how often reviewers change or reject something.
An approval that is never refused is not oversight; it is a delay.
Sources: EU AI Act, Article 14: Human oversight (AI Act Service Desk), OpenID Connect Client-Initiated Backchannel Authentication Core 1.0, FCA: Consumer Duty