Automated Credit Assessment Memo for Retail Loans

In retail banking and lending, credit committees depend on a Credit Assessment Memo (CAM) to make every loan approval decision. Today, credit analysts spend 3–8 hours manually preparing each CAM — gathering data from BSA, ITR, CIBIL, KYC, and property reports, cross-verifying figures, computing FOIR and LTV ratios, and drafting a narrative. Every case is a fresh marathon across 6+ systems.

This process is error-prone, analyst-dependent in quality, and creates a significant bottleneck in loan approval turnaround time. Inconsistent underwriting standards across analysts introduce credit risk, and CAM preparation blocks credit committee scheduling at peak volumes.

The Agent CAM is an Agentic AI system built on Salesforce Agentforce that automatically synthesizes BSA, CIBIL, KYC, ITR, and property data into a single, publication-ready Credit Assessment Memo — in under 15 minutes. Conservative underwriting is enforced by rule. Every figure is sourced and policy-referenced. A complete audit trail is maintained in Salesforce.

Instead of analysts manually compiling CAMs, the agent takes full responsibility for data synthesis, cross-verification, risk scoring, and recommendation generation — automatically.

Why is an Agent CAM Required?

Today's credit teams face a critical gap between loan application and credit committee decision:

• Credit analysts spend 3–8 hours per case gathering data from 6+ disparate systems
• Income figure selection is analyst-dependent — no conservative rule enforced
• FOIR computation sometimes uses BSA EMIs instead of conservative CIBIL obligations
• Recommendation language is vague with no policy traceability or specific figure citation
• CAM quality varies across analysts — no standardised underwriting output
• CAM preparation blocks credit committee scheduling at high-volume periods
• No systematic cross-verification gate — income mismatches and CIBIL staleness go undetected

This leads to delayed credit decisions, inconsistent underwriting quality, and avoidable credit risk exposure.

Agent CAM solves this by automatically orchestrating all data sources, enforcing conservative underwriting by rule, running a 9-step analysis pipeline, and generating a publication-ready CAM — with every figure sourced and every recommendation policy-cited.

Retail lending and credit teams commonly face challenges such as:

• Manual CAM Preparation is Time-Consuming: Credit analysts gather data manually from BSA, ITR, CIBIL, KYC, property valuation reports, and field investigation records. Cross-verifying figures across these sources and drafting a structured CAM narrative takes 3–8 hours per case, creating a bottleneck before every credit committee.
• Inconsistent Income Assessment: Different analysts apply different methods for identifying income: some use stated income, others use BSA income, and some average ITR across different year ranges. The conservative principle — always use the lowest of three verified sources — is not consistently enforced.
• Aggressive FOIR Computation: FOIR should be computed using CIBIL-reported obligations, which capture all existing EMIs including those not visible in the bank statement. When analysts use BSA-identified EMIs instead, FOIR is understated — resulting in credit approvals at higher actual risk than assessed.
• No Policy Traceability in Recommendations: Credit recommendations are written in analyst-specific language with no reference to the specific policy parameter, threshold figure, or deviation authority being invoked. This makes credit committee review difficult and creates compliance risk.
• Missing Cross-Verification Gate: There is no systematic check before CAM generation that verifies: income consistency across BSA, ITR, and stated sources; identity match between KYC and application; CIBIL report freshness; obligation gaps between CIBIL and bank statement; or property valuation alignment. These gaps surface only after a CAM is already drafted.
• Approval Bottleneck at Scale: At high application volumes, CAM preparation becomes the rate-limiting step in the credit process. Credit committees are delayed, loan approval turnaround inflates, and customer experience deteriorates. The problem compounds at peak lending periods without a scalable solution.

Agent CAM acts as an intelligent credit analyst that automatically orchestrates all loan assessment data sources, runs a structured 9-step analysis pipeline, and generates a publication-ready Credit Assessment Memo — within Salesforce Agentforce.

Instead of analysts manually compiling CAMs, the agent automatically:

• Ingests all loan data via a structured Apex payload (9 data objects: BSA, FSA, CIBIL, KYC, property, product policy, co-applicant)
• Runs Step 0: Cross-verification gate — income mismatch, identity match, CIBIL obligation gap, CIBIL staleness, property valuation mismatch
• Compiles the applicant profile: personal details, KYC status, employment stability score (1–5), FI report, banking relationship
• Assesses the loan request: EMI computation, LTV vs policy, sanction conditions, deviations (PRE/POST/ONGOING)
• Runs financial assessment: income cross-verification, FOIR using CIBIL EMIs (conservative), BSA signals, net worth, income confidence
• Evaluates CIBIL & credit bureau: score band, DPD 24M verdict, credit utilisation, hard enquiries, NPA history
• Assesses collateral: primary security, distress LTV, coverage ratio, insurance requirements, mortgage type
• Scores risk across 8 dimensions with product-specific weights: Income Stability, FOIR Adequacy, Asset Quality, Banking Conduct, CIBIL Score, Repayment Capacity, Collateral Coverage, Employment Stability
• Generates a credit recommendation: APPROVE / APPROVE WITH CONDITIONS / REFER / HOLD / DECLINE — with specific figures and policy references cited
• Produces an 8–12 sentence executive summary and runs a 12-item validation checklist
• Stores the complete structured JSON output (10 objects) in Salesforce Recommendation__c with full audit trail
• Records case_completion_time_minutes and number of agent reasoning runs for performance tracking

This enables credit officers to review a complete, policy-aligned CAM in under 15 minutes — rather than spending 3–8 hours preparing one.

Agent CAM is an Agentic solution built on Salesforce Agentforce. The Agentforce orchestration layer autonomously runs end-to-end — ingesting payload, executing 9 sequential reasoning steps, enforcing underwriting rules, generating the CAM, and committing the output to Salesforce — without waiting to be asked at each stage.

Each capability is delivered through Salesforce technologies as follows:

Step 0: Cross-Verification Gate

Before any CAM section is written, Agentforce runs five checks: income mismatch between BSA, ITR, and stated income (thresholds: >20% app vs BSA, >30% BSA vs ITR); KYC identity match; CIBIL obligation gap vs bank statement; CIBIL staleness (>90 days); property valuation mismatch (>15%). Block-level flags halt CAM generation; warn-level flags annotate the output.

9-Step Agentforce Pipeline

The Agentforce agent executes 9 sequential reasoning steps: Applicant Profile (Step 1), Loan Request & Conditions (Step 2), Financial Assessment with FOIR table using CIBIL EMIs (Step 3), CIBIL & Credit Bureau (Step 4), Security & Collateral (Step 5), Risk Matrix with product-specific weights (Step 6), Recommendation with policy citations (Step 7), Executive Summary (Step 8), and Validation Checklist (Step 9).

Weighted Risk Matrix

8 dimensions scored 1–5, each weighted by product policy (e.g. Home Loan: FOIR 20%, CIBIL 18%, Repayment 15%; Personal Loan: FOIR 25%, CIBIL 22%, Collateral 0%). Composite score = Σ(score × weight%) / Σweight%. Credit officer can adjust scores, edit weights in the Weight Editor modal, and challenge any dimension — each change triggers an agent reasoning run.

Agentic Credit Review Tab (LWC)

A Lightning Web Component Credit Review Tab surfaces all 9 CAM sections via a sidebar navigator. A right-panel Agent Reasoning Thread shows every step of the agent's thinking in real time. The credit officer can submit feedback via the Feedback Composer, triggering agent re-runs. The Commit button is enabled only after at least one run completes.

Architecture Principle: Deterministic Apex guarantees accuracy and auditability for financial computations (EMI, FOIR, LTV); the Agentforce prompt pipeline handles reasoning, narrative, and language; the Agentforce orchestration layer makes the system truly agentic — running end-to-end and committing output to Salesforce without manual intervention at each step.

9-Step Automated CAM Pipeline

A structured sequence of 9 reasoning steps executes automatically from a single trigger on the Salesforce Deal record. Each step produces a validated section of the CAM output. The pipeline enforces conservative underwriting rules at every step — income figure selection, FOIR computation source, LTV basis — with no analyst override possible without a formal deviation.

Benefits

• Complete CAM generated in under 15 minutes from a single trigger
• Conservative underwriting enforced by rule at every step — not by analyst discretion
• Each step's output is validated before the next step executes

Step 0: Cross-Verification Gate

Before any CAM section is written, Agent CAM runs five mandatory checks across all data sources. Income mismatch flags are raised if stated income differs from BSA by more than 20%, or BSA differs from ITR by more than 30%. Identity mismatch, CIBIL staleness, obligation gaps, and property valuation discrepancies are also detected. Block-level flags halt CAM generation entirely; warn-level flags annotate the relevant section.

Benefits

• Catches data quality issues before they contaminate the CAM
• Ensures credit officers receive a flagged, annotated CAM rather than a clean but inaccurate one
• Reduces credit risk from undetected income inflation or stale bureau reports

Conservative Income & FOIR Enforcement

Agent CAM always uses the lowest of three verified income figures: stated income, BSA normalised income, and ITR 3-year average. FOIR is always computed using CIBIL-reported obligations — not BSA-identified EMIs. These conservative rules are locked into the pipeline and cannot be overridden by individual analysts. Any deviation from policy requires a formal deviation record in Deviation__c with the appropriate approval authority.

Benefits

• Eliminates analyst-to-analyst income figure variance
• Prevents FOIR understatement from BSA-only obligation assessment
• Every income figure and FOIR computation is traceable to its source

Weighted Risk Matrix — Product-Specific Scoring

8 credit risk dimensions are scored 1–5 and weighted according to product policy. Home Loan policy weights FOIR at 20% and CIBIL at 18%. Personal Loan weights FOIR at 25%, CIBIL at 22%, and Collateral at 0% (unsecured). LAP weights Collateral at 15%. SEP Business Loan weights Income Stability at 18%. The composite score is computed as Σ(score × weight%) / Σweight% and mapped to a risk category: LOW (≥4.0), MEDIUM-LOW (≥3.0), MEDIUM (≥2.0), HIGH (<2.0).

The credit officer can adjust individual dimension scores using stepper controls (−/+), edit dimension weights in the Weight Editor modal (weights must sum to 100%), or challenge any dimension by pre-filling the agent feedback input for targeted agent review.

Benefits

• Risk scoring aligned to product-specific underwriting policy — not generic
• Officer judgment incorporated via score adjustments and agent re-runs, not manual overrides
• Every score adjustment triggers a logged agent reasoning run — full audit trail

Agentic Feedback Loop — Human-in-the-Loop

The Credit Review Tab's Agent Reasoning Panel shows the agent's full thinking chain in real time: every data check, every figure used, every flag raised, every risk dimension scored — with timestamps and run numbers. The Feedback Composer allows the credit officer to submit natural-language instructions ('Reduce loan to ₹40.5L and recheck FOIR', 'Add FD collateral of ₹5L') and the agent re-runs immediately, recomputing all affected metrics and updating the recommendation. Iterations continue until the officer is satisfied — then Commit locks the CAM to Salesforce.

Benefits

• Credit officer retains full control — the agent advises, the officer decides
• Every feedback round is logged with timestamp, input, and agent response
• Commit is only possible after at least one successful reasoning run

Policy-Cited Recommendations with Specific Conditions

Agent CAM generates one of five credit decisions: APPROVE, APPROVE WITH CONDITIONS, REFER TO CREDIT COMMITTEE, HOLD, or DECLINE. Every condition is specific and measurable — no vague language. Every rejection reason cites the specific policy parameter and threshold breached. Conditions specify the exact figure required (e.g. 'Reduce loan to ₹40,50,000 to bring distress LTV within 80% policy ceiling') rather than generic instructions.

Benefits

• Credit committee receives unambiguous, actionable conditions
• Policy compliance is demonstrable — every decision traces to a specific rule
• Reduced back-and-forth between credit officers and credit committee

Complete Salesforce Audit Trail

Every CAM run stores: the full agent output JSON in Prompt_Response__c, all 10 structured output objects in Salesforce fields, case_completion_time_minutes, the number of reasoning runs, which sections were auto-populated vs required human input, all data quality flags raised, and the committed-by officer's identity. The audit trail supports credit committee review, regulatory compliance, and portfolio analytics.

Benefits

• Every figure in the CAM is traceable to the source data object and field
• Complete reasoning history preserved — not just the final output
• Regulatory audit-ready by design — no separate documentation required

From loan application to CAM committed to Salesforce — fully within Salesforce Agentforce:

1. Loan Application Received

Credit officer creates a Deal (Opportunity) record in Salesforce. BSA Agent 2 processes the bank statement and populates BankStatementAnalysis{}. FSA Agent 3 processes financial statements and populates FinancialStatementAnalysis{}. KYC, CIBIL, and property data are linked to the Deal.

2. CAM Payload Assembly

CAMPayloadBuilder.cls assembles all 9 data objects into a single structured JSON payload. Completeness is validated — any missing mandatory field is flagged as DATA_MISSING before the agent is invoked.

3. Agent CAM Triggered

Credit officer clicks 'Generate CAM' on the Deal record, or the system triggers automatically on Deal status change. The Agentforce agent receives the assembled payload and begins execution.

4. Step 0: Cross-Verification Gate

The agent runs five cross-checks across all data sources. Block-level flags halt execution with a HOLD recommendation. Warn-level flags are annotated in the relevant CAM section and surfaced to the officer.

5. Steps 1–9: CAM Sections Generated

The agent sequentially executes all 9 pipeline steps: Applicant Profile, Loan Request & Conditions, Financial Assessment, CIBIL & Credit Bureau, Security & Collateral, Risk Matrix, Recommendation, Executive Summary, and Validation Checklist. Each step produces structured JSON output.

6. Credit Officer Reviews on Credit Review Tab

The CAM surfaces on the Salesforce Credit Review Tab. The officer navigates 9 sections via the sidebar, reviews the Agent Reasoning Panel, adjusts risk scores if needed, and submits feedback via the Feedback Composer. The agent re-runs on each feedback submission.

7. Officer Commits CAM to Salesforce

When satisfied, the officer clicks Commit. Recommendation__c is locked (Committed__c = true), Opportunity.Stage advances to 'Credit Committee', and the complete CAM output is preserved in Salesforce. All inputs are disabled post-commit.

All CAM sections, agent reasoning runs, flags, and conditions are stored in Salesforce custom objects. Audit trail maintained end-to-end.

Step 0: Cross-Verification Gate

Runs five mandatory checks:

• Income mismatch (BSA vs stated >20%; BSA vs ITR >30%)
• KYC identity match
• CIBIL obligation gap
• CIBIL staleness (>90 days)
• Property valuation mismatch (>15%)

Flags:

• INCOME_MISMATCH
• IDENTITY_MISMATCH
• UNDISCLOSED_OBLIGATIONS
• STALE_CIBIL
• PROPERTY_VALUATION_MISMATCH

Step 1: Applicant Profile

Personal details, KYC verification status, employment type and stability score (1–5 scale based on employer category and tenure), FI report (Auxilo opinion, address type, locality rating), banking relationship vintage and conduct.

Step 2: Loan Request & Conditions

EMI computation (P×r×(1+r)^n/(1+r)^n−1), LTV vs policy maximum (market and distress basis), loan purpose assessment, sanction conditions grouped by type (PRE-DISB / POST-DISB / ONGOING), and deviations (auto-triggered and manual, with approval authority level).

Step 3: Financial Assessment

Income cross-verification table (stated / BSA / ITR — uses lowest), FOIR table using CIBIL obligations (conservative rule), bank statement signals (ABB vs proposed EMI, dishonour count, cash pattern), net worth summary, FSA signal, income confidence level (HIGH / MEDIUM / LOW).

Step 4: CIBIL & Credit Bureau

CIBIL score and band (PRIME / NEAR-PRIME / SUBPRIME / HIGH-RISK), DPD 24-month verdict (CLEAN to NPA_HISTORY), 30/60/90+ DPD instance counts, credit utilisation ratio, recent hard enquiries (6 months), settled accounts and write-off history.

Step 5: Security & Collateral

Primary security description, independent valuation vs applicant stated (lower used), distress sale value (80% of valuation), conservative LTV (loan / distress value), total collateral coverage ratio, insurance requirements (life and property), and proposed mortgage type.

Step 6: Risk Matrix

8 dimensions scored 1–5 with product-specific weights. Composite = Σ(score×weight%) / Σweight%. Risk category: LOW (≥4.0), MEDIUM-LOW (≥3.0), MEDIUM (≥2.0), HIGH (<2.0). Key risks and mitigating factors listed. Officer can adjust scores and weights via Credit Review Tab — each change triggers an agent reasoning run.

Step 7: Recommendation

One of five decisions: APPROVE, APPROVE_WITH_CONDITIONS, REFER_TO_CREDIT_COMMITTEE, HOLD, or DECLINE. Sanctioned amount, recommended rate, tenure, and EMI specified. Each condition is measurable and figure-specific. Each rejection reason cites the exact policy parameter breached. No vague language permitted.

Steps 8 & 9: Executive Summary & Validation

8–12 sentence formal credit narrative covering all 7 required elements: applicant profile, loan purpose, income assessment, FOIR position, CIBIL standing, collateral, and recommendation rationale. 12-item validation checklist runs automatically. case_completion_time_minutes, sections_auto_populated, and sections_requiring_human_input recorded in CAMHeader.

Key Risks & Mitigations:

Agentforce Context Limits for Large Payloads

• Likelihood: MEDIUM
• Impact: HIGH
• Mitigation: Payload chunking strategy; priority field ordering (critical fields first); test with maximum-size payload (property + co-applicant + guarantor) in Phase 2. Partial CAM with HOLD state for incomplete payloads.

Agent 2 (BSA) or Agent 3 (FSA) Schema Changes

• Likelihood: LOW
• Impact: HIGH
• Mitigation: Payload schema version field agreed before Phase 2; schema change control process with Agent 2 and Agent 3 teams; schema freeze before integration testing begins.

AI Hallucination: Income or FOIR Errors

• Likelihood: MEDIUM
• Impact: HIGH
• Mitigation: Step 9 validation checklist as automated guard rail; 10 FOIR edge-case unit tests in Phase 3; deterministic Apex computation for all financial figures; agent produces narrative and recommendation, not financial calculations.

Credit SME Availability for QA Review

• Likelihood: MEDIUM
• Impact: MEDIUM
• Mitigation: Two structured review sessions booked in advance; written question set prepared to maximise SME session efficiency; written sign-off obtained before demo preparation begins.

Demo Data Privacy (PII)

• Likelihood: HIGH
• Impact: HIGH
• Mitigation: Fully synthetic and anonymised loan scenario (Sayali Badge is a fictional applicant). No production customer data at any stage of development, testing, or demo.

Why This Matters — Key Differentiators:

100% Salesforce-Native

No external systems required. The entire CAM workflow — payload assembly, 9-step agent pipeline, credit review, officer feedback, commit — operates within Salesforce Agentforce. No data leaves the Salesforce environment.

Conservative Underwriting Enforced by Rule

Agent CAM always uses the lowest income figure, always uses CIBIL obligations for FOIR, and always uses independent valuation for LTV. These rules are locked into the pipeline. Any exception requires a formal deviation with the appropriate approval authority.

Human-in-the-Loop Agentic Design

Agent CAM is not a black box. The Agent Reasoning Panel shows every step of the agent's thinking in real time. The credit officer can challenge any dimension, adjust any score, and resubmit feedback — the agent re-runs on every change. The officer commits only when satisfied.

Auditable Every Step

Every CAM figure is traceable to its source data object and field. Every reasoning run is logged with timestamp. The complete Prompt_Response__c stores the raw agent output. Credit committee review, regulatory audit, and portfolio analytics are all supported by the same audit trail.

Indian Retail Lending Native

IST timezone, Lakhs/Crores number format, CIBIL bureau integration, FOIR benchmarks by Indian product type, SARFAESI compliance for property insurance, RERA property verification, NACH/ECS/UPI narration handling — all built natively for the Indian retail lending context.

• Faster Credit Decisions: Reduce CAM preparation time from 3–8 hours to under 15 minutes, enabling faster credit committee scheduling and improved loan approval turnaround.
• Increased Credit Analyst Productivity: Eliminate manual data gathering, cross-verification, and narrative drafting — freeing analysts to focus on credit judgment and applicant engagement.
• Consistent Underwriting Quality: Conservative income figure selection, CIBIL-based FOIR computation, and policy-cited recommendations applied identically across all analysts, products, and deals.
• Improved Credit Risk Management: Step 0 cross-verification gate detects income mismatches, stale CIBIL reports, undisclosed obligations, and property valuation gaps before any credit decision is made.
• Scalable Credit Operations: Process high application volumes without proportional increase in analyst headcount. The agent scales to volume; the credit team scales to judgment.
• Credit Committee Efficiency: Standardised, policy-aligned CAM with specific conditions and policy references enables faster, more confident credit committee decisions.
• Regulatory Compliance Ready: Complete Salesforce audit trail with every figure sourced, every flag logged, and every decision traceable to a specific policy rule — audit-ready by design.
• Seamless BSA Integration: Agent CAM ingests BSA Agent 2 output directly, completing the full credit workflow from bank statement processing through to published CAM within 15 minutes of trigger.