YTJK Systems

Home / Insights / System design

Document processing with AI: intake, extract, verify

System design · Get AI into production · Reduce cloud and AI cost

Problem

Skilled staff re-keying data from documents, with backlogs and errors.

Approach

Classify, extract and validate automatically; route low-confidence fields to people.

Outcome

Faster processing, lower cost per document and a full audit trail.

Document processing is one of the clearest returns on AI: high volumes, structured outputs and measurable accuracy. The design challenge is not extraction itself but knowing when the result can be trusted, and routing everything else to a person efficiently.

The business problem

Invoices, application forms, claims or contracts arrive by email, upload and post. Staff read each one and type the details into core systems. Backlogs build at peak times, errors cause rework, and there is no consistent record of who checked what.

Constraints

  • Many document types and layouts, including scans and handwriting.
  • Personal or financial data that must stay within approved services and regions.
  • Downstream systems that reject incomplete or invalid records.
  • Auditors need to see the source, the extracted value and who approved it.

Architecture

Processing pipelineStraight-through where confidence is high, human review where it is not.
Intake to system of recordEvent-driven, one document at a time
  1. IntakeEmail, upload, API or scanned post
  2. ClassifyDocument type, language, quality
  3. ExtractFields and tables with a confidence score per field
  4. ValidateBusiness rules, reference data, duplicates
  5. ReviewLow-confidence fields shown to a person beside the source
  6. PostValidated record written to the system of record
Every field keeps its source location, confidence and reviewer
Measured continuously
  • Straight-through rate
  • Field accuracy
  • Review time per document
  • Cost per document
  • Backlog and turnaround
  • Exceptions by type

Key decisions and trade-offs

  • Confidence per field, not per document: people review only the uncertain fields, which is where most of the time saving comes from.
  • Rules after the model: totals, dates, reference data and duplicate checks catch errors that confidence scores miss.
  • Thresholds set by risk: stricter thresholds for payments and identity data, looser ones for low-impact fields.
  • Feedback loop: reviewer corrections become test cases and tuning data for the next release.
  • Event-driven and idempotent: each document is processed once, retried safely and traceable end to end.

Measures of success

  • Straight-through processing rate
  • Field-level accuracy against a labelled sample
  • Average handling time per document
  • Cost per document
  • Turnaround time at peak volumes

Common pitfalls

  • Aiming for 100% automation and stalling, instead of starting with high-volume, high-confidence types.
  • Review screens without the source image beside the field.
  • No labelled sample, so accuracy is never measured.
  • Sending sensitive documents to services outside approved regions.