Skip to content
Spark
Writing

Governance and compliance

Human-in-the-loop AI: what it means and when a regulator expects it

Human-in-the-loop is not a person clicking approve. It is a person who can realistically catch an error in the time they are given. Here is how to tell real oversight from theatre, and how to match it to the risk.

By Seb Lawson · 8 April 2026 · 4 min read

Human-in-the-loop is not a person clicking approve. It is a person who can realistically catch an error in the time they are given, with the information they need, and the authority to stop what happens next. If any of those three is missing, you do not have oversight. You have theatre. And a regulator can tell the difference in about ten minutes.

The phrase gets used as a reassurance. ‘Don’t worry, there’s a human in the loop.’ But the loop only counts if the human can act. A reviewer who sees forty AI outputs an hour, with no context and no realistic power to reject, is a rubber stamp with a pulse. The JFSC’s guidance is explicit that oversight must be meaningful, and meaningful is a high bar dressed in a plain word.

Real oversight versus theatre

Three things separate the two, and all three have to hold.

  • Time. The reviewer has enough of it to actually assess the output, not just glance at it. If the volume forces a two-second look, the review is decorative.
  • Information. They can see what the AI saw and how it reached the output. A person cannot sensibly approve a decision whose basis is hidden from them.
  • Authority. They can say no, and no sticks. If overriding the system is career-costly or practically impossible, the human is not in control of anything.

Theatre is easy to spot once you look for it. The approval rate is 99 percent. Nobody remembers the last time a human rejected an output. The reviewer cannot explain how the system reached its answer. Each of those is a sign the loop exists on the org chart and nowhere else.

Calibrate oversight to risk

Not every AI use needs a human standing over it. Uniform oversight is as wrong as no oversight, because it wastes scrutiny on trivia and starves the cases that matter. We use four modes, chosen by how much damage a wrong output could do.

  1. Approve. The AI proposes, a human signs off before anything happens. Reserved for high-stakes cases: onboarding decisions, client-facing letters, anything touching a payment or a regulatory return.
  2. Assist. The AI drafts, a human edits and owns the result. Good for medium-stakes work where a person is already in the flow, like drafting a file note or summarising a document.
  3. Exception. The AI runs on its own and only escalates the uncertain or unusual cases to a person. Efficient for high-volume, low-stakes work where most outputs are routine.
  4. Kill switch. The AI runs autonomously, but someone monitors it and can stop it instantly. For low-stakes automation where speed matters and the blast radius of a failure is small and reversible.

The art is putting each workflow in the right mode. Most firms find the mistake runs one way: they apply full approve-mode oversight to everything, the reviewers drown, and the scrutiny quietly degrades into theatre on the cases that actually needed it. Match the mode to the harm and the oversight stays real where it counts.

When a regulator expects it

A person is expected in the loop whenever an AI output can reach a client, move money, or feed a regulatory return without a human between the model and the consequence. That is the general shape of it across principles-based regimes like Jersey’s and the more prescriptive tiers in the EU. The higher the potential harm, the closer and more genuine the oversight a supervisor will expect to see.

The test a regulator applies is not whether a human was technically present. It is whether that human could and did exercise real judgement. So the burden is on you to show it: an approver by name, a record of what they saw, evidence that rejections actually happen. Without that, ‘there was a human in the loop’ is a claim you cannot back.

Oversight and audit are one control

Human oversight only holds up if it leaves a record. The approver’s name, what they reviewed, whether they changed anything, and when. That is where oversight meets the AI audit trail: the log is what turns ‘we had a human review it’ from a claim into evidence. Build the two together, because separately they are each half a control.

In our work, getting oversight right on a single high-stakes workflow is a couple of weeks, and it is the part of the work that keeps a firm defensible as it scales. Prove the pattern once, then reuse the same four modes across every system you build.

If you are not sure whether your current oversight is real or theatre, book a review. We will look at one live workflow, tell you which mode it should be in, and show you what a supervisor would actually want to see. No deck, no hype.

Let’s map how your business works

A free 30-minute call with Seb or Mark. Bring the spreadsheet you hate most and we’ll tell you honestly whether a system would fix it.