Signals & data//ai-governance
AI governance & kill switch
Turn all AI off instantly, keep running on deterministic rules, and audit every AI output.
AI governance & kill switch
LoanPilot's AI can be switched off instantly, platform-wide, and every AI output is recorded for audit. This is built to the RBI Draft Guidance on Regulatory Principles for Model Risk Management, 2026 — an emergency kill switch, human-in-command, traceability, and customer disclosure.
The master kill switch
Admin → Configuration → AI governance → "AI models enabled (master kill switch)".
- On (default): AI models assist the analysis — the credit committee, specialist reviewers, salary / scenario / narration checks, risk-sense, name / loan match, and the Business & location read.
- Off: every AI model is deactivated at once. The platform keeps running on the deterministic rules / NLP engine, and every decision is referred to a human (never auto-approved).
It is a runtime setting, so it also works as an on-prem admin's emergency stop: if an AI model ever produces biased or wrong output, turn it off and the next analysis runs rules-only.
What still works when AI is off
Everything deterministic is untouched: statement parsing, the credit scores, FOIR / repayment capacity, EMI / EDI / credit-card offers, per-product eligibility, bureau merge, fraud / document checks, bounce and loan detection, loan servicing, and the Account Aggregator flow.
For the AI parts, paused mode is designed to still read well:
- Replay — if the same statement was analysed earlier with AI on, its recorded AI findings (verdict, reviewers, scenarios, risk-sense) are replayed so the report reads the same as before.
- Deterministic review — for a new statement, an LLM-free engine turns the computed facts into a plain review: a narrative plus strengths and concerns. The verdict stays Refer to manual review.
The AI-findings audit trail
Every analysis records every AI model output to a dedicated ai_findings store
(PostgreSQL in production), independent of the result:
- each AI output (committee decision, reviewers, scenarios, risk-sense, name / loan match, Business & location), with its model, timestamp, owner / team, and whether AI was live when it was produced;
- an
ai_usagerecord accounting for every AI step that ran (calls and failures per step) — so no AI touchpoint is unaccounted for; - a kill-switched run still records its outcome (marked AI-not-live), so the trail is always complete.
Read it at GET /api/analyses/{id}/ai-findings, or on the report under the
folded "AI findings (audit trail)" panel on the Decision tab.
Customer disclosure
Every analysis shows a standard disclosure line stating whether an AI model assisted the assessment and that a human officer makes the final decision.