How to Detect Financial Statement Discrepancies Using AI in Underwriting

Financial statement discrepancies are where credit risk and fraud risk overlap. Reviewers find them inconsistently because the work is mechanical, fatiguing, and not always the highest-value task in the day. AI is suited to this work because the comparisons are repeatable and the audit trail is the point.

Artificial Intelligence (AI)

What Counts as a Material Discrepancy

Not every difference between documents is material. A practical definition: a discrepancy that, if left unaddressed, would change the credit decision; a discrepancy that, if left unaddressed, would mask a material misrepresentation; or a discrepancy that, in aggregate with other discrepancies, suggests a pattern.

The institution's policy should specify what constitutes material at the field level (dollar thresholds, percentage variances, categorical inconsistencies). This matters beyond credit quality. Patterns of misrepresentation are the raw material for suspicious activity monitoring, and the FFIEC BSA/AML Examination Manual expects institutions to have documented processes for identifying and escalating red flags rather than relying on individual reviewer intuition.

The Discrepancies AI Surfaces Reliably

Period-over-period inconsistencies. Cash on the balance sheet that does not reconcile to the cash-flow statement. Receivables that move without supporting collections. Inventory changes inconsistent with cost of goods sold.

Cross-statement inconsistencies. Net income on the income statement that does not reconcile to retained earnings on the balance sheet. Cash flow from operations that does not reconcile to the income statement after adjustments.

Tax-return-to-statement inconsistencies. Revenue on the tax return that does not match the income statement.

Document-version inconsistencies. A draft financial statement that conflicts with a final version.

Reasonableness flags. Margins, ratios, and trends outside the typical range for the industry or for the borrower's history.

Each of these is grounded in submitted documents. Each is citation backed.

What AI Does Not Decide

The configured agent persona surfaces discrepancies. It does not decide whether the discrepancy indicates fraud; decide whether the credit decision changes; decide whether to file a SAR; or decide whether to close the relationship.

The credit officer, the underwriter, and the BSA officer (where appropriate) make these calls. Suspicious activity determinations and filings under 31 CFR 1020.320 are judgment decisions that stay with people. Similarly, where discrepancy findings contribute to a decline, the adverse action reasons owed under Regulation B must be specific and accurate, which is exactly what CFPB Circular 2022-03 requires of institutions using complex algorithms in credit decisions.

Step 1. Configure the materiality thresholds

The institution should document, in writing: the percentage variance that constitutes a flag for each statement line; the dollar threshold that constitutes a flag; the categorical inconsistencies that always flag regardless of magnitude; and the exception types that always escalate to senior review.

These thresholds are policy, not just configuration. They are also governance artifacts. Threshold-driven analytical tools fall within the model risk management principles of SR 11-7, and the agencies' revised interagency model risk management guidance issued in April 2026 (SR 26-2) carries the same principles-based expectations forward to AI-supported analysis: document the logic, validate it, and control changes to it.

Step 2. Run the structured analysis

The configured agent persona performs period-over-period variance analysis with documented thresholds, cross-statement reconciliation, tax-return-to-statement reconciliation, document-version comparison, and reasonableness checks against the borrower's history and industry context.

Step 3. Surface the structured findings

The output is a structured discrepancy report listing the discrepancies identified, the materiality classification, the supporting documentation and citations, and the proposed reviewer disposition path.

Step 4. Reviewer disposition

The reviewer addresses each finding. For each finding, the reviewer documents the disposition (cleared, requires clarification, requires escalation), the rationale, and the supporting documentation if additional material was requested.

Step 5. Escalate where appropriate

Some discrepancies escalate by policy: discrepancies above a specific dollar threshold, discrepancies of a specific categorical type, or patterns suggesting misrepresentation. Escalations that raise potential suspicious activity route to the BSA officer, whose obligations are framed by the Bank Secrecy Act and the institution's monitoring program.

Step 6. Capture the trend data

Over time, the discrepancy data is useful at the portfolio level. Patterns of discrepancies in a specific industry, geography, or relationship-manager portfolio can surface broader risk signals.

Common Pitfalls

Reviewer fatigue from noise. If the materiality thresholds are too tight, every case generates many low-value flags. Calibrate to the policy.
Cleared exceptions without reasoning. Each clearance should have documented rationale.
Industry context not incorporated. A trend that is normal for an industry should not be flagged.
Inattention to patterns. Individual discrepancies are surfaced; patterns across customers should also be analyzed.

Discrepancy detection is one input into a larger audit-ready credit workflow.

How StandardC AI Approaches This

StandardC AI's intelligence layer performs structured discrepancy analysis through configured agent personas calibrated to the institution's materiality thresholds and policy. Outputs in StandardC AI Report identify each discrepancy with citations to the supporting documents and the reasoning structured for reviewer use.

Frequently Asked Questions

Does this work with QuickBooks-style financials in addition to CPA-prepared statements?

Yes. The configuration accommodates different statement formats.

What about borrowers without tax returns yet (start-ups)?

The configuration adjusts the reconciliation expectations based on what is available.

Can this support fraud-detection workflows?

It supports fraud-relevant analysis by surfacing patterns and inconsistencies. Fraud determinations are made by the BSA officer and fraud team.

Who decides whether a discrepancy warrants a SAR filing?

The BSA officer. The AI surfaces the discrepancy with citations; the suspicious activity determination and any filing under 31 CFR 1020.320 remain human decisions.

Authoritative Sources

‍