From Manual Re-Keying to Validated Automated Extraction

Customer context

ClearLedger builds back-office reconciliation and reporting software for accounting teams, and their own operations team processes a high volume of incoming financial statements as part of their reconciliation workflow.

The challenge

ClearLedger's operations team manually transcribed data from incoming PDF statements — a mix of formats from different sources, some consistent, some not — into their reconciliation system. This was slow, occasionally error-prone, and created a processing backlog during month-end and high-volume periods when it mattered most.

Requirements

• Handle varied document formats

  Statements arrived in inconsistent layouts across different sources, not one standard template.

• Validate before acceptance

  Extracted data needed to be checked against business rules, not trusted blindly.

• Human review for uncertain cases

  The system needed a defined path for documents it couldn't confidently process.

• Direct system integration

  Validated data needed to flow into the existing reconciliation system without a manual export/import step.

Discovery

We spent the first phase of the engagement categorizing incoming document types by structure — a smaller number of highly consistent formats from major sources, and a long tail of less consistent formats from smaller ones. This split directly shaped the technical approach: consistent formats could use faster, cheaper structured parsing, while the long tail needed model-assisted extraction.

Solution

We built a document processing pipeline combining layout-aware structured parsing for consistent formats with model-based extraction for less predictable ones, followed by a validation layer that checks every extracted value against expected formats and cross-field consistency rules (line items summing to a stated total, dates falling in plausible ranges) before anything reaches the reconciliation system.

Technical architecture

Incoming documents enter a processing queue and are classified by format type to route them to the appropriate extraction path — structured parsing or model-assisted extraction. Extracted fields pass through a validation layer checking format correctness and cross-field consistency. Documents passing validation post directly to ClearLedger's reconciliation system via API; anything flagged appears in a review queue showing the source document and extracted fields side by side for fast human verification.

Implementation approach

We started with the highest-volume, most consistent document format, validated the full pipeline end to end against real historical documents with known-correct answers, and only then expanded to the less consistent formats — rather than trying to build one general solution for every format simultaneously. This let ClearLedger's team start realizing value early while the harder cases were still in development.

Key features

• Layout-aware extraction

  Combines structured parsing with model-based extraction for varied document formats.

• Cross-field validation

  Extracted data is checked against business rules before acceptance.

• Human review queue

  Low-confidence extractions are routed for review with source and extracted data side by side.

• Direct reconciliation integration

  Validated data flows automatically into the existing reconciliation system.

Integrations

• Reconciliation system

  Validated extracted data is posted directly via API.

• Document intake

  Automated ingestion from existing email and upload channels.

Challenges & decisions

We deliberately chose not to attempt full end-to-end automation for the least consistent document formats. For a small tail of genuinely irregular documents, we route directly to manual review rather than forcing an extraction attempt with predictably low confidence — a narrower automation scope that's reliable beats a broader one that requires constant correction.

Representative outcomes

Technology stack

In their words

Related services