Checking Invoice Rows After OCR

Keep each quantity, unit and amount tied to the right line

Published Updated

Yellow-and-white illustration follows a highlighted invoice row through extraction, magnified cell checks and a review clipboard with accepted and unresolved items.

A repair provider submits a scanned invoice for two packs of replacement seals. Optical character recognition (OCR) reads the quantity and amount correctly, but the outgoing fields describe two individual seals. The line total still adds up. The claim file now contains a plausible amount attached to the wrong description of the work.

Validate the relationships within each row before using extracted values. Check which item, quantity, unit, price and amount belong together, and keep a route back to the source. Recognition of individual characters and a matching total are useful checks; neither establishes that the business row has been interpreted correctly.

Separate extraction, interpretation and acceptance

  1. 1 · Extract

    Read the table

    Collect the relevant headers, cells and page coverage from the source document.

  2. 2 · Check

    Interpret each row

    Associate the item, quantity, unit and amount, then apply the document's validation rules.

  3. 3 · Review

    Accept the business values

    Resolve exceptions and record which values the workflow may use.

A completed OCR action means processing finished. It does not establish that an invoice is correct, a repair is covered or a payment is authorised. Keep those decisions in the responsible claim, finance or connected business process.

Check more than the characters

Four checks for an extracted invoice table.
CheckFailure to detectUseful review
Headers and row boundariesA repeated header becomes a line item, or a wrapped description joins the next row.Compare the billable rows with the relevant source pages and documented continuation rules.
Quantity and unitPacks become individual items, hours become days, or a blank unit is guessed.Retain the printed unit and apply only an approved interpretation or conversion.
Price, amount and totalsA decimal separator is misread, or a subtotal is included twice.Check arithmetic using the invoice's currency, discounts, taxes and rounding rules.
Source and accepted valuesA corrected row cannot be traced to the source, or an old result overwrites a review.Connect the source version, extraction result, mapping and accepted correction.

Keep enough table context for the reviewer

Keep the original cell text alongside the interpreted value where the distinction matters. A blank item code might continue a description or indicate missing information. A new-page header might change a unit. Agree how known layouts are handled and send unresolved cases to review instead of silently carrying values forward.

The FormKiQ OCR API can return table headers and cell values using outputType=TABLES. Amazon Textract's underlying response describes cell relationships, merged cells and geometry. Normalised table or CSV output may omit that richer context. Confirm which output your processing path retains before promising source-cell highlighting. Give reviewers the exact source document and page context even when coordinates are unavailable.

For a claim packet that combines an invoice with other documents, Amazon Textract Insights adds an AWS integration option with automatic document splitting and source locations for extracted fields. Evaluate it separately from FormKiQ’s documented OCR output. Reconcile split pages to the original packet, then use the same row checks on both extraction paths. A field can point to the correct page and still be assigned to the wrong line item.

Worked example: the total is right but the unit is wrong

This fictional repair-provider invoice extends CLA-1042's repair follow-up. Its two billable rows are two packs of five seals at CAD 75 per pack, totalling CAD 150, and 1.5 hours of labour at CAD 80 per hour, totalling CAD 120. The subtotal is CAD 270; taxes and other charges are outside this example.

  1. Retain the invoice with the repair evidence

    The handler links the received scan to CLA-1042 and identifies the source version being processed. The invoice remains distinguishable from the purchase evidence and earlier repair decision already in the claim.

  2. Extract the relevant tables

    The configured OCR action requests Textract table extraction across the required pages. The workflow waits for the result and checks that the two billable rows are represented, without treating headings or the subtotal as extra items.

  3. Catch the misleading quantity

    The proposed mapping says two individual seals. Its arithmetic still passes: two multiplied by CAD 75 is CAD 150. The unit check flags that the source says packs of five. Neither the row count nor the subtotal resolves this mismatch.

  4. Review the source and record the correction

    The reviewer confirms two packs at CAD 75 per pack and retains the pack-size wording. Any conversion to ten individual seals needs an explicit rule for the receiving system, including the corresponding price basis. The original proposal and accepted interpretation remain linked.

  5. Use the accepted rows in the continuing workflow

    The handler records the checked invoice data and outstanding questions with the repair follow-up. A connected system receives the agreed accepted values and reports its outcome. That handoff does not itself approve payment or replace the claim adjudication.

Make exceptions visible before data is used

Define which checks can accept a known layout automatically and which need a person to inspect the source. Prioritise ambiguous rows that affect quantities, costs or decisions. A high confidence score for recognised text does not settle a disputed column mapping.

If the process permits only part of an invoice to proceed, identify the accepted subset and unresolved rows explicitly. Otherwise keep the invoice awaiting review. Preserve later source replacements and revised mappings without silently changing the values used for an earlier decision. The extraction audit guide covers acceptance history, while the OCR rebuild guide covers changed sources and processing dependencies.

Test a representative invoice before scaling

  • Test wrapped descriptions, merged headers, repeated headings and page breaks.
  • Match each billable row to its item, quantity, unit, price and amount.
  • Check currency, decimal separators, discounts, taxes and rounding against the source.
  • Retain source text and the reason for a reviewed interpretation or correction.
  • Keep uncertain rows visible and apply the agreed partial-acceptance rule.
  • Verify that the receiving workflow uses the accepted values and that late results cannot replace newer review.

Make the row mean what the document says

Useful table extraction preserves the association between the values on a line. Check those relationships, retain the source context and record the accepted interpretation before the data supports the next business action. A matching total is one check in that process.

      Start with FormKiQ Core

      The open-source foundation — API-first, deployable into your own AWS account, and free to use. Right for architecture validation and early implementation.

      Get Started Free

      Deploy FormKiQ Essentials or Advanced

      Production-ready editions for departments and complex workflows. Start with a Proof-of-Value deployment or go straight to production.

      Explore Options

      Plan an Enterprise Rollout

      For governance-heavy environments with residency, sovereignty, assurance, and multi-jurisdiction requirements. Talk to us about the right deployment model.

      Book a Call