Human-in-the-Loop

Human-in-the-loop is a way of using AI where a person must review and approve the AI's work before it takes effect. The system pauses at set checkpoints and waits. The AI prepares, and a person decides.

It sits on a spectrum. In a human-in-the-loop setup, nothing happens until a person approves. In a human-on-the-loop setup, the AI acts on its own while a person monitors and can step in afterward. Fully automated systems have no person involved in each decision. Which level fits depends on the risk of the decision, not on the tool.

Why it matters for lenders

Some lending decisions carry legal duties that a person must be able to stand behind. In the US, Regulation B, which carries out the Equal Credit Opportunity Act, requires that a statement of reasons for denying credit be specific and give the principal reasons. If a model made the call and no one can say why, the lender can't write a truthful adverse action notice. In May 2025, the Consumer Financial Protection Bureau (CFPB) withdrew its earlier circulars about complex algorithms, but that requirement is in the regulation itself and has not changed.

A sensible split for most commercial, small business and community lenders looks like this. AI can sort documents, pull numbers into a spread, list missing items, flag mismatches and draft a summary. A person confirms the documents, adjusts for one-time items, assigns the risk rating, makes the credit decision and confirms the true reasons for any decline.

The review has to be real. People tend to over-trust software and wave through its output, a problem known as automation complacency. Good practice gives reviewers the evidence behind each AI output, the authority to change it and a simple way to record why. Every review and override belongs in the audit trail.

How it works in practice

A CDFI uses an AI tool to help underwrite small business loans. For each file, the tool extracts figures from tax returns and bank statements, calculates key ratios and drafts a one-page summary with three flagged concerns.

The underwriter checks each extracted figure against its source page. She corrects one misread number and records why. She looks into the three flags, decides that one matters and writes her credit memo. The tool cannot approve, decline or send a notice. Those buttons only work for a named staff member with the right approval authority.

When the CDFI's funder later asks how a decision was made, the file shows what the AI prepared, what the underwriter changed and who approved the loan. The AI underwriting tool saved time, and accountability stayed with people.