Signals & data//ai-review
AI review
What the AI does — and the firm line it never crosses.
AI review
LoanPilot uses AI in two bounded ways. In both, deterministic rules own every borrower-facing number; the AI advises.
The committee
A panel of specialist reviewers reads the evidence and the rule baseline, then the committee decides the score, decision, conditions and a plain explanation. It applies this account's configured policy (the approve / conditional thresholds from your rules — not fixed numbers).
Guardrails:
- The AI's one number lever — a suggested maximum EMI — is clamped so it can only lower the EMI, never raise it.
- A rise over the rule baseline must be evidenced; an unexplained rise beyond the limit is capped and the case is sent to manual review.
- If the AI can't finish, or is turned off, the decision falls back to the rules and is sent to manual review — never auto-approved.
Risk-sense
A second-opinion reviewer that catches what the fixed rules don't: round-tripping, income-vs-lifestyle mismatch, payer concentration, mis-declared business, unusual money movement. It returns a concern level and evidence-based findings, shown on the Risk tab.
Borrower intelligence
A third, advisory agent reads a whole customer dossier — the KYC
profile, every business, the location / market / climate intelligence, the declared
loans and the bank-statement monitoring signals — and reasons across every
perspective (repayment capacity, business viability, local market, climate exposure,
stability, leverage, data quality) and across scenarios (base / stress / upside). It
returns a summary, a per-business read (viability + climate impact + market
position), a current-market analysis, a portfolio view, strengths, risks, a
recommended_stance, conditions and a confidence. It runs on the same stack as
the committee and is metered the same way.
Safety
All statement text the AI sees is wrapped as untrusted data, so content inside a statement can't issue instructions to the model. The same holds for borrower intelligence: location-intelligence text, business notes and declared values are treated as untrusted data, not instructions.