Maker-checker for AI-assisted credit decisions

AI is now doing real work in loan files. It reads documents, extracts values, analyses bank statements and drafts the credit note. That work is useful. It also raises a question every credit and risk head has to answer clearly: when the system has done most of the preparation, who actually made the decision?

The answer has to be a named person. Not "the system", not "the model", and not "the vendor".

Why a named person must stay accountable

There are practical reasons. When an account goes bad, an auditor raises an observation or a borrower complains, someone has to explain why the loan was sanctioned on those terms. A model cannot be called into a credit committee or held to a delegation of powers. A person can.

There are regulatory reasons too, and RBI has been consistent about them.

RBI's outsourcing directions for NBFCs, issued in November 2025, say that an NBFC shall not outsource core management functions, and they name decision-making functions such as giving sanction for loans, including retail loans.

RBI enforces this. In January 2025 it announced a penalty on an NBFC that had outsourced one of its decision-making functions, the sanction of loans, to its digital lending app partner.

For digital lending, RBI's credit facilities directions for NBFCs, also from November 2025, say the NBFC remains fully responsible and liable for all acts and omissions of its lending service providers, and that outsourcing in no manner dilutes or absolves the NBFC of its obligations.

On AI specifically, RBI's FREE-AI committee report of August 2025 is a framework rather than a binding direction, but its position is plain: final authority should rest with humans, who should be able to override AI, and accountability cannot be delegated to the model or the underlying algorithm. It also recommends that regulated entities adopt a Board-approved AI policy.

Taken together, the message is consistent. Tools can prepare, analyse and recommend. A person inside the lender decides and answers for it.

What to record

Maker-checker only works if the record shows what happened. For each decision on a loan file, record at least:

• Who. The maker who prepared the file and the checker who approved it, as named individuals, signed in through the lender's own identity system, not a shared login.
• What. The decision itself, the terms, and any deviation from policy, with the reason.
• On what evidence. The values that were relied on and the documents and pages they came from, including which flags were raised and how each was resolved.
• What the AI contributed. Which values were extracted or derived by the system, which were corrected by a person, and which model version did the work.
• When. Time stamps for each step, from file creation to sanction.

The record should be append-only. Corrections are added as new entries, never written over old ones, and the log should be tamper-evident, so that any later change to history can be detected.

Designing sign-off that is not a rubber stamp

The risk with AI-prepared files is not that people stop checking. It is that checking slowly turns into clicking "approve" on a tidy summary. A few design choices make that harder:

1. Show evidence, not just conclusions. The checker should see each key value next to the page it came from, not a summary paragraph. Approving evidence is a different act from approving prose.
2. Make flags impossible to skip. Anything unverified, mismatched or missing must be resolved or explicitly accepted with a reason before sign-off is possible.
3. Separate the roles. The person who prepared or corrected a file cannot also approve it. The system should enforce this, not rely on habit.
4. Tie approval to authority. The checker's approval limits should follow the lender's delegation of powers, and a file above those limits should route upwards automatically.
5. Review the reviewers. Sample approved files periodically and check whether flags were genuinely examined. Look for approvers whose acceptance of flags is consistently higher than their peers, and talk to them.
6. Feed corrections back. When checkers correct the system, those corrections should be captured and used to improve it under review, so the same error is not presented again and again.

Where the tools fit

A sensible arrangement gives AI the preparation work and keeps the decision with people. Documents are read, values are linked to their pages, statements are reconciled, and anything that cannot be confirmed is flagged. A maker assembles the file; a checker with the right authority approves it; both are recorded.

That is how Console, the workflow and control layer of Inferova, is designed: each step in a process is routed to a person or to AI, maker-checker sign-off is built in, and every action goes into an append-only, tamper-evident audit trail.

If you would like to see how this works on your credit process, you can request beta access at inferova.com/demo.

Sources