LendingIQ logo
← Agent catalogue

AI Agent Profile · LendingIQ · Bengaluru

Chief Risk Officer AI

Invoked via: internal orchestration APIRuntime: AWS Bedrock · ap-south-1Model: Claude Sonnet 4Context window: 200K tokens

DivisionRisk & governance

Resume

What this agent does

The CRO AI is a reasoning agent that reads credit files, applies LendingIQ's risk policy, queries live data sources, and produces structured risk opinions — including a risk grade, key concerns, policy exceptions, and a recommended next action. It does not approve or reject loans autonomously. It prepares the risk case so that human credit officers can decide faster and with more information.

Primary functions

Credit File Assessment

Triggered at origination

Invoked when: new loan application submitted

  • Reads the full application: financials, bureau report, bank statements, GST returns, ITR — whatever documents are passed in context.
  • Checks stated income vs GST turnover vs ITR declared income. Flags where numbers conflict and asks for clarification or raises a concern.
  • Applies LendingIQ credit policy (retrieved via RAG) — checks LTV, FOIR, leverage ratio, sector exposure limits. Lists every policy condition met or breached.
  • Generates a risk grade (Low / Moderate / High / Decline) with the 3–5 specific reasons that drove the grade, not a black-box score.
Output: Structured JSON risk opinion — grade, concerns, policy exceptions, recommended conditions, confidence level, and a plain-language paragraph for the credit file.

Portfolio Early Warning

Triggered on schedule / CBS event

Invoked when: monthly refresh OR repayment missed

  • Reads the latest repayment data, updated bureau scores, and GST filing status for accounts in the watch-list.
  • Identifies behavioural signals: declining GST filings, bureau score drop, overdraft utilisation spike, bounce in ECS — and narrates what the combination suggests.
  • Does not predict NPA probability with a calibrated number. Produces a qualitative "deterioration narrative" with supporting evidence from the data passed to it.
Output: Early warning report per account — signals observed, what they may indicate, suggested action (call borrower / restructure / provision / refer to collections).

Policy Guardrail Check

Triggered pre-sanction

Invoked when: proposal reaches sanction stage

  • Runs a final checklist against the current credit policy document (v4.2) — every hard limit and soft guideline, not a cached summary of it.
  • Returns a pass/fail/exception-needed verdict for each policy condition, with the exact clause cited and the applicant's actual figure against the limit.
  • Cannot approve exceptions itself. Prepares the exception memo language if one is needed, for a human authoriser to sign off.
Output: Policy compliance checklist with clause references, figures, verdicts, and — where exceptions exist — a draft exception justification.

Risk Report Generation

Triggered on schedule / on demand

Invoked when: board pack due OR ad-hoc query

  • Synthesises portfolio data passed to it — sector mix, vintage performance, NPA buckets, concentration — into a readable narrative report.
  • Answers specific risk questions in natural language: "What is our exposure to real estate developers above ₹5 Cr?" — by reasoning over the data provided.
  • Does not maintain a live database. It reads whatever structured data is injected into its context at the time of invocation.
Output: Narrative risk report in the format specified — board summary, detailed MIS, or regulatory draft — with section headings, key observations, and data tables.

Knowledge base

LendingIQ Credit Policy v4.2

The live policy document, retrieved via RAG at query time. Not baked into the model — always reads the current version.

RBI Master Directions

IRACP norms, NBFC scale-based regulations, PCA framework. Loaded as a grounding corpus during fine-tuning.

Basel III / IV Framework

Capital adequacy, LCR, NSFR, credit risk standardised approach. Applied when asked to assess capital impact.

General Credit Reasoning

Pre-training knowledge of credit analysis, financial ratios, lending products, and risk frameworks from public sources up to knowledge cutoff.

Sector Reference Data

Industry benchmarks for MSME, real estate, agriculture, and services sectors — used to contextualise borrower financials.

Live Data via Tool Calls

Bureau scores, GST returns, ITR, CBS repayment data — fetched at runtime. Not stored. Not recalled across sessions.

How decisions are formed

Input it needs
Application documents, financial statements, bureau report, policy document (via RAG). The better the input, the better the output. Garbage in, garbage out applies fully.
Reasoning method
Chain-of-thought reasoning over documents. Checks facts, applies rules, identifies contradictions, weighs concerns. Does not run a statistical model — reasons in language.
Confidence signal
Returns a confidence level (High / Medium / Low) based on data completeness. Low confidence always triggers human review — the agent flags its own uncertainty.
Explainability
Every verdict cites the specific document, figure, or policy clause that drove it. No black-box outputs. The credit officer can see exactly why.
Consistency
Highly consistent on rule-based checks. May vary slightly on judgment calls — e.g., how seriously to weigh a borrower's growth narrative — across invocations.

Hard guardrails

Will notApprove or sanction a loan. It produces a risk opinion; a human signs the sanction letter.
Will notOverride a policy exception without human authorisation. It drafts the exception memo; the authoriser approves it.
Will notRetain borrower data between sessions. Each invocation starts fresh — no memory of previous applications.
Will notFabricate data it cannot verify. If a document is missing or a figure is unverifiable, it says so explicitly rather than estimating.
Will notMake credit decisions for cases it flags as outside its competence — legal disputes, fraud investigations, related-party exposures — it refers these to humans.

Known limitations

Does not maintain calibrated PD probabilities. Risk grades (Low / Moderate / High) are qualitative assessments, not actuarially validated default probabilities.Use a dedicated scorecard model if a precise PD number is required for provisioning or capital calculation.
Knowledge cutoff applies to general credit reasoning. It may not know about a regulatory change issued after its training date unless the updated document is passed in context.Always inject the current policy document via RAG — do not rely on the model's baked-in knowledge for policy details.
Cannot read scanned documents or images. Financial statements must be machine-readable text or structured data. Scanned PDFs must be OCR-processed before passing.Invest in an upstream OCR layer for document ingestion.
May be over-literal on policy rules. Edge cases that a seasoned credit officer would handle with discretion may be flagged as hard exceptions by the agent.Calibrate policy language carefully. Ambiguous rules produce ambiguous verdicts.
Qualitative inputs (borrower management assessment, site visit notes) are taken at face value. The agent cannot independently verify them.Qualitative assessments fed to the agent should be validated by humans before use.
Agent Profile · Chief Risk Officer AI · LendingIQ · BengaluruLast updated April 2026 · For internal use

Important Reads

Learn more about how to deploy Chief Risk Officer AI to your lending workflow.