Doc Judge: field-level OCR trust scores for SMB lender document intake

Statement Gate is going into a crowded market. Heron Data already markets per-field confidence scores, checks that totals reconcile, and sends flagged items to review, specifically for MCA funders. Ocrolus built its business on people checking only the uncertain parts of bank statements and pay stubs. MoneyThumb charges per statement and reconciles against the statement summary. Kaaj is a funded newcomer targeting the same lenders. General platforms like Extend and Sensible ship confidence scoring and AI review agents. A new entrant can still differentiate in three ways. First, on-premise deployment for lenders: Unstract is the only self-hosted bank statement option I found, and it charges by page rather than being built around lenders and reviewers. Second, lower per-document pricing from a small model, compared with people-heavy services. Third, published proof that its confidence scores actually predict errors, which I found no incumbent publishing; recent research argues the common confidence signals break down at practical thresholds. Judge Only is less contested. I found no product that audits a lender's existing parsing vendor and reports error rates by document type, only generic audit advice and a public complaint about Ocrolus's accuracy guarantee, which suggests the pain exists. Its risks are that you need ground truth to prove error rates and a judge model can only estimate them, that vendors may block it or add their own accuracy reporting, and that lenders may buy it only once, during vendor selection. The strongest wedge is probably to use Judge Only to show a lender where their current vendor fails, then sell Statement Gate on-premise as the replacement for those document types.

• Heron Data

  Parses bank statements for MCA funders and small-business lenders. Markets a confidence score on each field so reviewers know where to look first, checks that totals reconcile, checks page counts and period coverage, and sends variances to a review queue. This is the closest match to Statement Gate. It is a hosted SaaS, not self-hosted.

• Ocrolus

  The established human-in-the-loop service for bank statements, pay stubs and tax forms for fintech and small-business lenders. A Philadelphia Fed presentation names PayPal's small business lending as a customer. Also offers fraud detection and cash-flow analysis. A public review criticizes the accuracy of its Perfect Audit product.

• MoneyThumb

  Bank statement conversion and analysis popular with MCA funders. Charges per statement, not per page. Claims it automatically reconciles 95% of statements against the statement summary. Sells a separate fraud tool called Thumbprint.

• Kaaj

  Seed-funded ($3.8M) AI underwriting startup for small-business lending that parses bank statements. Competes directly with MoneyThumb for MCA and small-business lenders.

• Unstract

  Bank statement extraction API that can be deployed on-premise. Uses its own proprietary OCR (LLMWhisperer) and charges by page. The only self-hosted option for bank statements I found, but it is not built specifically for lenders.

• Extend (Review Agent)

  General document extraction platform. Its Review Agent is an AI reviewer that improves confidence scores and flags fields to check. Close to the judge concept, but horizontal rather than built for lenders.

• Sensible

  Document extraction platform that combines logical validation rules, OCR quality scores and LLM-reported uncertainty to flag fields for review.

• LlamaIndex bank statement extraction

  General LLM document parsing vendor with a dedicated bank statement extraction offering and its own extraction benchmark.

Checked 2026-10-28