Agentic Credit Analysis

Blake Williams, Rohan Goel, Jake Pappo, Avery Geller, Mueed Rehman22 Oct 2026

Problem

Ramp has to keep checking the credit risk of every business customer as card spend grows. Bureau data on businesses is thin, and bank cashflow misses an unpaid invoice or a coming loan payment. Those signals sit in financial statements. Each statement has its own layout, a borrower may misstate a line item, and a lender may want more conservative accounting than the borrower used. Large exposures therefore often go to manual underwriting, which scales only with headcount.

Example from the article: a business capitalizes routine repairs and adds the resulting depreciation back to adjusted earnings. Ramp's credit analysts may treat these as recurring operating costs, so the agent applies Ramp's definition instead of the business's calculation.

Solution in brief

An LLM agent extracts line items with their periods, then aggregates them under Ramp's canonical definitions. Deterministic code recalculates totals and flags fallbacks. A benchmark of 200 analyst-reviewed statements measures the result, and analyst corrections feed the next review.

Questions this case answers
  • How to extract structured figures from heterogeneous documents with an LLM agent when every document has a different layout and aggregate numbers must be exactly reproducible
  • How to evaluate an extraction pipeline against expert-reviewed values when the target is reliable inputs for a downstream decision, not exact reproduction
  • How to turn messy financial statements (PDFs, spreadsheets, scans) into a standard schema for credit decisions without manual analysts?
  • How to stop an LLM agent from miscalculating or double-counting figures when it extracts financial data?

Tools and methods

  • Format-specific extraction — recover formulas from spreadsheets and digital PDFs, cross-checked OCR for scans
  • Semantic definitions — canonical per-line-item rules with vertical exceptions
  • Deterministic reconciliation — metrics like EBITDA recalculated in code
  • Analyst-reviewed benchmark — 200 statements, recall and precision within 5%
  • Analyst feedback loop — corrections with reasons feed the next review
Continue reading — it's free

Google or a one-time email code. No password.