Human-in-the-loop
Human-in-the-loop means a person stays involved in an AI system's operation, reviewing, approving, or taking over cases the system routes to them. The term covers arrangements as different as signing off a single action and handling an entire conversation, which is why it needs unpacking before it means very much.
Human-in-the-loop is often presented as a safety feature, and in regulated financial services it is frequently an obligation. That distinction matters, because a feature can be tuned down when it costs too much and an obligation cannot. It is also not a single design: where the person sits changes what the arrangement is actually worth.
How does human-in-the-loop work?
The useful question is where the person sits relative to the action, and there are three answers with very different properties.
In the loop: the system prepares, and a person approves before anything happens. Slowest, and the only arrangement that prevents a wrong action
On the loop: the system acts, and a person monitors and can intervene. Faster, and catches patterns rather than individual errors
Out of the loop, with review: the system acts, and people sample afterwards. Fastest, and finds problems after customers have experienced them
Vendors use human-in-the-loop for all three. A deployment described as having a human in the loop may mean approval on every refund, or it may mean somebody reads 2% of transcripts next week. Asking which is the first question worth putting to a vendor.
When is human-in-the-loop a regulatory requirement?
For most routine volume it is a design choice. For defined categories it is not.
The EU AI Act requires human oversight for high-risk AI systems under Article 14, and creditworthiness assessment falls inside that classification. In UK conduct terms, Consumer Duty requires firms to avoid foreseeable harm, and the categories where that bites are consistent: vulnerability, complaints, and financial difficulty.
The practical consequence is that one number reads two ways. A high handoff rate looks like failure against a containment metric and like compliance against a control narrative. Our guide on AI copilots versus autonomous agents works through where firms draw that line.
What does human-in-the-loop cost?
Oversight is not free. Every case routed to a person is a case the automation did not close, and the economics of the whole programme run through that number. Three costs worth planning for:
Throughput: approval gates add latency, and on voice that latency sits in front of the customer
Attention: reviewers approving hundreds of near-identical actions stop reading them, which is the well-documented failure of any rubber-stamp control
Staffing shape: the remaining queue is harder on average, so the team needs more experience rather than less
The third catches teams out. Automating routine volume leaves people with a concentrated residue of difficult work, and planning headcount as though the remaining cases resemble the old average produces a stretched team.
Frequently asked questions
What is the difference between human-in-the-loop and human-on-the-loop?
Position relative to the action. In the loop means a person approves before anything happens, so a wrong action is prevented. On the loop means the system acts while a person monitors, so a wrong action is caught afterwards. That distinction decides what the control is actually worth.
Does human-in-the-loop slow everything down?
Only where it is applied indiscriminately. Most production deployments in finance run autonomously on routine volume and apply oversight to defined categories, which keeps the cost proportionate. Applying approval to every action removes most of the benefit of automating in the first place.
Is human-in-the-loop enough to satisfy a regulator?
Not on its own. A reviewer who approves everything without reading it is not oversight, whatever the process document says. What tends to be examined is whether the person had the information, the time, and the authority to disagree, and whether there is evidence of them doing so.
Does human-in-the-loop prevent hallucination?
It catches some of it, and it is a poor primary defence. A reviewer checking a draft may spot an invented figure, and may equally approve it, because a plausible wrong answer looks much like a right one. Automated grounding checks catch more, more consistently, and at lower cost.