Human-in-the-Loop (HITL)

Who uses it

Lenders, banks and credit unions, servicers, third-party agencies, and in-house collections teams that run AI agents or models, plus the vendors that build them. The term isn’t defined in the FDCPA, the TCPA, or Regulation F. The UK’s Financial Conduct Authority defines it in its Oct 2026 AI publication as “A control where a person reviews, approves, challenges or can override an AI output or decision before it is used.” Where a given workflow calls for human review is a question teams often take to their compliance team or counsel.

How it works

• A spectrum, not a switch. NIST’s AI Risk Management Framework (AI RMF 1.0, January 2023) says “Human-AI configurations can span from fully autonomous to fully manual. AI systems can autonomously make decisions, defer decision making to a human expert, or be used by a human decision maker as an additional opinion.” Its MAP 3.5 reads: “Processes for human oversight are defined, assessed, and documented in accordance with organizational policies from the GOVERN function.” (NIST AI 100-1)

• Checkpoints. Common review points in credit operations include settlement offers outside set ranges, hardship or payment plan approvals, disputes and complaints, payoff quotes, and anything flagged as a possible bankruptcy, attorney representation, or cease request. Where each checkpoint sits is a design and risk decision.

• Authority and capability. The FCA writes that “Effective oversight is not simply about reviewing AI outputs, but about ensuring people have the right authority, information, capability and escalation routes to challenge or intervene where needed.” (FCA)

• Automation bias. The FCA’s glossary describes it as “The tendency for people to place too much trust in an automated output or recommendation, even where it may be wrong or incomplete.” It lists “Excessive reliance on human-in-the-loop controls” among areas for improvement, saying “Human review may mask weaknesses in the model,” and describes a risk of review “being a ‘box-ticking’ exercise, rather than meaningful oversight.”

• Records. The FCA describes a smaller firm that recorded “whether staff needed to review AI outputs, whether this happened, and whether they approved the outputs, overrode them or escalated issues,” which it says “created an audit trail showing how oversight operated in practice, rather than relying on an assumed control.”

• Related terms in common use. “Human-on-the-loop” usually means a person monitors a system that acts on its own and can step in, rather than approving each action. A “human handoff” or transfer is different again: the consumer asks for a person. New York City’s proposed rules call that a “single-step transfer mechanism” (proposed 6 RCNY 5-310).

What it means for AI agents (operations questions)

• Which agent actions run on their own, which wait for approval, and which are never automated? Is that map written down, and who signs off on changes to it?

• Does the reviewer see enough to judge the output (the account history, the source of a number, the rule that fired), or only the proposed answer?

• Is every review logged with who reviewed, what they saw, what they decided, and how long it took?

• Does anyone measure override rates? An override rate near zero can mean the model is good, or that nobody is really looking. The FCA describes one firm that had reviewers assess cases “without being shown the AI’s proposed outcome” to compare the two.

• When the agent is unsure, does it stop and route to a person, or keep going? What sets that threshold?