Bank Statement Analysis Agent

Automated OCR Extraction & Credit Analysis within Salesforce Agentforce

In retail banking and lending, credit decisions depend on accurate income assessment from bank statements. Credit officers currently spend 2–4 hours per application manually extracting transactions, computing income metrics, and identifying fraud signals — creating bottlenecks, inconsistency, and avoidable credit risk.

Statements arrive in multiple formats — PDFs (digital and scanned), Excel files, CSV exports, and image scans — each requiring different tools and expertise to process reliably. Manual extraction misses transactions, misclassifies income sources, and fails to detect fraud indicators.

The Bank Statement Analysis Agent is an Agentic AI system built on Salesforce Agentforce that automatically accepts bank statements in any format, performs OCR extraction, runs a 24-point fraud checklist, computes all credit metrics, and delivers a complete underwriting decision package to the credit officer — in seconds, not hours.

Instead of officers manually extracting and analysing data, the agent takes full responsibility for extraction accuracy, income verification, and credit signal generation — automatically.

Why is a Bank Statement Analysis Agent Required?

Today's credit teams face a critical gap between statement submission and credit decision:

• Bank statements arrive in 5+ formats — each requiring separate handling and tools
• Credit officers spend 2–4 hours per case on manual data extraction
• Manual review misses transactions, misclassifies income types, and overlooks fraud signals
• Income normalisation is analyst-dependent — inconsistent across the team
• Credit decisions are delayed, leading to poor applicant experience and SLA breaches
• Fraud indicators go undetected without a systematic, automated checklist

This leads to delayed credit decisions, increased credit risk, and reduced operational efficiency.

The Bank Statement Analysis Agent solves this by automatically processing statements in any format, extracting all transactions, computing income and obligation metrics, and delivering a complete credit signal — without any manual effort from the credit officer.

Retail lending and credit teams commonly face challenges such as:

• Multi-Format Statement Chaos: Bank statements arrive as PDFs (digital and scanned), Excel files, CSV exports, and image scans. Each format requires different tools, and scanned documents require OCR processing that is error-prone when done manually.
• Manual Extraction is Slow and Error-Prone: Credit officers spend 2–4 hours per application manually reading and transcribing transactions. Manual extraction misses entries, miscalculates running balances, and introduces data errors that affect credit decisions.
• Inconsistent Income Assessment: Different analysts apply different methods for identifying salary credits, normalising income across months, and handling outliers. This creates inconsistency in FOIR calculations and credit decisions across the team.
• Missed Fraud Signals: Round-trip transactions, balance spikes, irregular salary timing, duplicate narrations, and cash deposit anomalies are difficult to detect manually across hundreds of transactions. Without systematic checks, fraud risk goes undetected.
• No Standardised Credit Signal: Credit officers produce individual assessments in different formats and styles. There is no standardised GREEN / AMBER / RED credit signal that credit committees can rely on for consistent, policy-aligned underwriting.
• Scalability Constraints: As loan application volumes grow, manual extraction does not scale. Credit teams are forced to either hire additional analysts or accept longer turnaround times — both of which are operationally expensive.

The Bank Statement Analysis Agent acts as an intelligent credit analyst that automatically accepts bank statements in any format, extracts all transactions, and computes a complete set of credit metrics — within Salesforce Agentforce.

Instead of credit officers manually processing statements, the agent automatically:

• Accepts statements in 5 formats: PDF (digital), PDF (scanned), Excel .xlsx, CSV, PNG / JPEG
• Detects the format automatically and routes to the appropriate extraction pipeline
• Extracts all transactions with date, narration, amount (Dr/Cr), running balance, and payment mode
• Parses account metadata: bank name, account number, holder name, IFSC, statement period, opening and closing balance
• Verifies balance continuity: closing_balance[i] = opening_balance[i+1] across all pages
• Classifies every transaction: Salary, EMI, Rent, Insurance, UPI, Cash withdrawal, Credit Card
• Detects salary credits and normalises monthly income using median of last 3 consistent salary entries
• Identifies EMI obligations and computes total fixed monthly outflows
• Computes FOIR (Fixed Obligation to Income Ratio), ABB (Average Bank Balance), and DSCR (Debt Service Coverage Ratio)
• Runs a 24-point fraud checklist: round-trip transactions, balance spikes, irregular timing, duplicate narrations, cash anomalies
• Generates a composite GREEN / AMBER / RED credit signal with recommended action
• Produces a structured executive summary for the credit officer, ready for credit committee review

This enables credit officers to focus on credit decisions rather than data extraction.

The Bank Statement Analysis Agent is an Agentic solution built on Agentforce. The Agentforce orchestration layer autonomously runs end-to-end — detecting format, extracting transactions, running fraud checks, computing credit metrics, and deciding what credit signal to surface to the officer, without waiting to be asked.

Each capability is delivered through Salesforce technologies as follows:

Phase 1: OCR Extraction

Agentforce Phase 1 prompt processes the uploaded statement. For digital PDFs and structured files, content is parsed directly by the Apex pre-processor. For scanned PDFs and image files, OCR via Vision API extracts raw text. All transactions are extracted, classified, and stored in custom Salesforce objects (BS_Transaction__c, BS_Monthly_Analysis__c).

Phase 2: Credit Analysis

Agentforce Phase 2 prompt runs credit computations on extracted transaction data. FOIR, ABB, and DSCR are calculated by Apex services. The 24-point fraud checklist runs systematically. Income normalisation identifies salary credits and computes median income. The GREEN / AMBER / RED signal is generated by the Agentforce orchestration layer based on all computed metrics.

Format Detection & Pre-Processing

Apex file handler inspects the uploaded document, detects format (PDF digital, PDF scanned, Excel, CSV, PNG/JPEG), and routes it to the appropriate extraction pipeline. Format normalisation ensures a consistent transaction schema regardless of input type.

Credit Officer Workbench

A Lightning Web Component dashboard surfaces all extracted metrics, fraud flags, income assessment, and the executive summary to the credit officer. The officer reviews, approves, declines, or refers the case — all within Salesforce, with a complete audit trail.

Architecture Principle: Deterministic Apex guarantees accuracy and auditability for financial figures; the Agentforce prompt pipeline handles OCR extraction and language summarisation; the Agentforce orchestration layer makes the system truly agentic — running end-to-end and delivering the credit signal without waiting to be asked.

Multi-Format Statement Acceptance

Accepts bank statements in any of 5 formats: PDF (digital), PDF (scanned image), Excel .xlsx, CSV, and PNG / JPEG image files. Automatically detects the input format and routes to the appropriate extraction pipeline — no manual pre-processing required.

Benefits

• No format-specific handling required from the credit officer
• Consistent extraction quality across all input types
• Reduced statement rejection and resubmission cycles

Phase 1: Automated OCR & Transaction Extraction

Agentforce Phase 1 extracts every transaction from the statement with full metadata: date, narration, amount (debit/credit), running balance, and payment mode (UPI/NEFT/NACH/RTGS/IMPS/Cash). Account metadata is parsed and stored. Balance continuity is verified automatically across all pages.

Benefits

• 100% transaction extraction — no entries missed
• Balance continuity verification flags tampered or incomplete statements
• OCR quality scoring guides confidence assessment per document

Phase 2: Credit Metric Computation

Agentforce Phase 2 automatically computes all credit underwriting metrics:

• FOIR: Total monthly EMIs + obligations divided by normalised income. Benchmarked per product — ≤50% for home loans, ≤55% for personal loans
• ABB: Mean of daily closing balances over the statement period, compared to 3× proposed EMI benchmark
• DSCR: Monthly net income divided by total debt obligations. Values > 1.25 indicate healthy repayment; < 1.0 is a critical flag
• Income Normalisation: Median of last 3 consistent salary credits, excluding outliers, classified as SALARIED / SELF_EMPLOYED / MIXED with confidence score

Benefits

• Consistent, policy-aligned FOIR computation across all cases
• Eliminates analyst-to-analyst variance in income assessment
• Benchmarked metrics enable instant comparison against credit policy thresholds

24-Point Fraud Checklist

The agent runs a systematic 24-point fraud detection checklist across all extracted transactions, covering:

• Round-trip transactions: credits immediately followed by equal debits
• Balance spikes: sudden large credits inconsistent with income pattern
• Irregular salary timing: salary credits on inconsistent dates month-to-month
• Duplicate narrations: identical transaction descriptions on different dates
• Cash deposit anomalies: large or frequent cash credits inconsistent with profile

Benefits

• Systematic fraud detection across every transaction — not sampling
• Fraud flags surfaced to credit officer before credit decision
• Reduced credit risk from statement manipulation

GREEN / AMBER / RED Credit Signal

The agent generates a composite credit signal from all computed metrics and fraud indicators:

• GREEN: Proceed — all metrics within policy, no fraud flags, income stable
• AMBER: Proceed With Conditions — borderline metrics or minor flags; specific conditions recommended
• RED: Decline or Refer to Credit — metrics breach policy or significant fraud flags detected

The recommended action (PROCEED / PROCEED WITH CONDITIONS / REFER TO CREDIT / DECLINE) and an executive summary for the credit officer are generated automatically.

Benefits

• Single, consistent credit signal aligned with underwriting policy
• Credit committee receives standardised assessment — not analyst-specific narrative
• Faster credit decisions with reduced ambiguity

Salesforce-Native Credit Officer Workbench

A Lightning Web Component workbench within Salesforce surfaces all extracted data, computed metrics, fraud flags, income assessment, and the executive summary to the credit officer. The entire workflow — upload, analysis, review, and credit action — operates within Salesforce with a complete audit trail.

Benefits

• No external tools required — entire workflow within Salesforce
• Complete audit trail: every metric traceable to the source transaction
• Credit committee-ready output generated automatically

Scalable Processing Architecture

Multiple statements can be queued and processed in parallel. Custom Salesforce objects store all extracted data in a structured, queryable format — enabling reporting, dashboards, and downstream integration with credit decision systems and Agent CAM.

Benefits

• Handles high application volumes without additional analyst headcount
• Consistent processing time regardless of statement complexity
• Structured data storage enables credit portfolio analytics and reporting

From statement upload to credit decision — fully within Salesforce:

1. Lead Created

Credit officer creates a Lead record and uploads the bank statement in any of 5 accepted formats (PDF digital, PDF scanned, Excel, CSV, PNG/JPEG).

2. Lead → Deal Conversion

Lead is converted to a Deal record. Bank statement is linked to the Deal and queued for processing.

3. Statement Selection

Credit officer selects the statement on the Deal record. Can upload a fresh statement or reuse a previously processed one.

4. Agent: Phase 1 — OCR Extraction

Agentforce Phase 1 runs automatically: format detected, all transactions extracted, account metadata parsed, balance continuity verified, OCR quality flags set.

5. Agent: Phase 2 — Credit Analysis

Agentforce Phase 2 computes FOIR, ABB, DSCR, runs income normalisation, detects salary credits and EMI obligations, executes the 24-point fraud checklist.

6. Analysis Surfaced on UI

Structured results appear on the Lightning credit workbench. Credit officer reviews metrics, fraud flags, income assessment, and the GREEN / AMBER / RED credit signal.

7. Credit Action Recorded

Officer approves, declines, or refers the case to the credit committee. Action is recorded in Salesforce with a complete audit trail. All data is stored in custom objects for downstream use by Agent CAM.

All results are stored in Salesforce custom objects. Audit trail maintained end-to-end.

Key Risks & Mitigations:

OCR Accuracy on Scanned Statements

• Likelihood: HIGH
• Impact: HIGH
• Mitigation: Diverse test scan library; manual review fallback flag (LOW_QUALITY_SCAN) surfaced to credit officer; confidence scoring per extraction.

Token Limits for Large Statements

• Likelihood: MEDIUM
• Impact: HIGH
• Mitigation: Chunked processing strategy; paginated Agentforce prompt execution; statement split by month for large 24-month statements.

Excel / CSV Format Variations Across Banks

• Likelihood: HIGH
• Impact: MEDIUM
• Mitigation: Format normalisation layer with bank-specific parsers; column mapping configuration for major Indian banks.

Demo Data Privacy (PII)

• Likelihood: HIGH
• Impact: HIGH
• Mitigation: Fully synthetic and masked test data used exclusively in demo environment. No real customer data at any stage.

Why This Matters — Key Differentiators:

100% Salesforce-Native

No external systems required. The entire workflow — statement upload, OCR extraction, credit analysis, and workbench — operates within Salesforce Agentforce. No data leaves the Salesforce environment.

Auditable Every Step

Every credit metric is traceable to the exact source transaction used in its computation. The complete audit trail is maintained in Salesforce custom objects, enabling credit committee review and regulatory compliance.

Indian Lending Native

IST timezone handling, Lakhs/Crores number format, NACH/ECS/UPI/NEFT/RTGS narration parsing — all handled natively. Built specifically for the Indian retail lending context.

Agentic CAM Generation

The Bank Statement Analysis Agent feeds directly into Agent CAM, which synthesises BSA output alongside CIBIL, ITR, KYC, and property data into a publish-ready Credit Assessment Memo — within 15 minutes of triggering, replacing 3–8 hours of manual preparation.

• Faster Credit Decisions: Reduce bank statement processing time from 2–4 hours to seconds, enabling faster loan approval turnaround and improved applicant experience.
• Increased Credit Officer Productivity: Eliminate manual extraction and data entry, freeing credit officers to focus on credit judgment and applicant engagement rather than data processing.
• Consistent Underwriting Quality: Standardised FOIR, ABB, and income assessment methodology applied identically across all applications — no analyst-to-analyst variance.
• Improved Credit Risk Management: Systematic 24-point fraud checklist detects manipulation signals that manual review misses, reducing exposure to fraudulent applications.
• Scalable Operations: Process high application volumes without proportional increase in analyst headcount. The agent scales to volume; the team scales to judgment.
• Regulatory Compliance Ready: Complete audit trail in Salesforce with every transaction, metric, and credit signal traceable to source data — audit-ready by design.
• Seamless CAM Integration: BSA output feeds directly into Agent CAM, completing the full credit workflow from statement to published Credit Assessment Memo within 15 minutes.