A supplier sheet says 0.8 millimeters. The extracted text says 8 millimeters.
The sentence reads cleanly. The number fits the destination field. An import can succeed while carrying the wrong specification into the next decision.
Optical character recognition, or OCR, needs a review boundary before its output becomes business data. Put the page beside the extracted text. Identify the fields that control the next action. Keep every correction attached to its source location.
The workflow below is a proposed review method. Its supplier document, errors, and decisions are fictional. It defines evidence an operator can inspect without treating a vendor score as approval.
Identify Which Representation You Are Checking
Start by naming the representations in your own document workflow.
A scanned page is an image of a page. An existing text layer is text already carried inside the PDF. Recognized text is the output your selected OCR process produces. Layout describes where text and other elements sit. Business interpretation assigns a meaning to a value.
Adobe's scanned-document guidance distinguishes image-only scans from the searchable text layer that Acrobat creates through OCR.
Keep those layers separate in the review record. Selecting a word in a viewer tests the available text representation. It does not establish that the selected word matches the visible page or the intended specification.
For each input, record whether your extraction used an existing text layer, image recognition, or a combination. If the tool does not expose that distinction, mark it unknown. Save the original file and the unedited extraction together.
Review recognition against the visible source. Review interpretation against the document's labels and the receiving process's requirements. A correctly recognized “0.8” still needs its unit, field label, and applicable revision.
This boundary precedes AI data entry automation. The downstream workflow needs a known input. A plausible value does not supply that evidence.
Freeze the Input Before Reviewing the Output
Give the fictional document a stable review identity: supplier sheet S-17, revision B, three pages.
Record its file fingerprint or another identifier your team can reproduce. Keep a separate identifier for the extraction run. Include the tool, processor, API version, selected pages, and relevant options when available.
Do not overwrite the first result with a corrected copy. Preserve three separate records: the received source, the raw extraction, and the reviewed output. Link them through the same document identity.
This makes a later disagreement answerable. Someone can inspect whether the source changed, recognition changed, or a reviewer changed the value.
Define the intended use before checking fields. In this example, the output fills a draft internal comparison worksheet. It does not authorize an order or a product substitution. That narrow use determines what the reviewer must verify.
If a later process wants to use the same data for another decision, require a new acceptance check. Approval for a comparison worksheet does not establish suitability for every business action.
Inspect the Page Set Before Reading Individual Fields
Count the source pages yourself. Compare that count with the pages requested and returned by the extraction run.
For sheet S-17, page one contains identification and revision information. Page two contains material dimensions. Page three contains handling notes. The fictional scan rotates page two sideways and leaves part of page three faint.
Create one coverage row for each source page. Mark whether the image opens, whether its edges are present, whether recognition returned text, and whether the requested page range included it.
A returned page with no recognized text needs a reason. It could be an intentionally blank divider in your source. It could also be an unreadable page. Inspect it before assigning a status.
Check both ends of each page. A clear heading does not establish that the bottom note survived. Include narrow margins, rotated labels, and faint sections in the coverage review.
When a page is missing from the input itself, stop the dependent work. Running recognition again cannot recover a page you never received.
Use Vendor Scores as Review Signals
Google's Document AI response guide separates recognized text, text anchors, and layout coordinates. Enterprise Document OCR can return page image-quality scores when processOptions.ocrConfig.enableImageQualityScores is enabled. That score describes readability assessment, not approval of your business fields.
For Azure Document Intelligence v4.0 confidence guidance, confidence is an estimated likelihood of correct detection. Availability varies by field and model. The 2024-11-30 GA notes identify table, row, and cell confidence for custom models. Do not assume every layout result includes those scores.
Keep each score's level and meaning beside it. A page-quality score and a word-confidence score answer different questions. A missing score remains missing.
Use the signals to choose where review starts. Then check the critical values directly. A confident transcription can still conflict with the visible source. An uncertain transcription can be correct after inspection.
Do not adopt a universal acceptance cutoff from this worksheet. Define any local routing threshold against checked examples from the specific document type and configuration. Keep the human acceptance rule separate.
Select Critical Fields by Their Downstream Effect
Ask the receiving process owner which errors would change the next decision.
For the fictional comparison worksheet, critical fields are sheet identity, revision, part identifier, thickness, unit, and the handling restriction. Decorative text is outside the agreed extraction scope.
Write that scope before review. Otherwise a reviewer can spend the session polishing headings while the decimal that controls comparison remains unchecked.
For each critical field, capture the label and nearby context. A source location such as “page two, dimensions block, thickness row” helps another reader find the evidence. Add a bounding region or crop reference if your tool preserves one.
Check small characters explicitly. Read the decimal, sign, separator, suffix, and unit as separate parts of the value. For an identifier, compare character by character rather than recognizing the general shape.
Do not repair a suspicious value from memory. If the visible source says 0.8, preserve 0.8. If the visible source is unclear, preserve the uncertainty and assign the question.
Work Through the Fictional Supplier Sheet
The fictional raw extraction returns “Sheet S-17, revision B” correctly on page one.
Page two returns a thickness of “8 mm.” The visible page clearly shows “0.8 mm.” Record a transcription error. Store the raw value and the corrected value in separate fields. Identify the reviewer and the source location.
The rotation belongs in the page record. It explains a condition worth checking across that page. It does not prove that rotation caused this particular error. Keep the cause unknown unless you test it.
Page three returns “Keep dry” from the handling block. A faint final line remains unreadable on the source image. Do not complete that line from a familiar supplier phrase.
The page coverage review is complete. The handling restriction is unresolved. Because that restriction is critical to the intended worksheet, the document cannot yet pass acceptance.
Ask the document owner for a clearer source or replacement scan. Keep the comparison worksheet in draft. A corrected decimal does not resolve an unrelated missing instruction.
Copy This OCR Review Worksheet
Keep a short document heading above the worksheet: source identity, revision, page count, extraction run, intended use, reviewer, and review date.
Use a separate row for each coverage question or field check. The following entries describe only the fictional sheet.
| Page coverage | Critical field | Source location | Observed error | Correction | Unresolved text |
|---|---|---|---|---|---|
| Page 1 present and checked | Sheet identity and revision | Identification block | None observed | None | None |
| Page 2 present, rotated | Thickness and unit | Dimensions block, thickness row | Raw text says 8 mm | Reviewed text: 0.8 mm | None at this location |
| Page 3 present, faint section | Handling restriction | Final line of handling block | Source cannot be read confidently | No value supplied | Final instruction needs clearer source |
| All 3 pages accounted for | Document acceptance | Whole review | Critical instruction remains open | Hold draft output | Document owner must supply evidence |
Add your own decision, reviewer, evidence reference, and resolution date columns when the record needs them. Keep the raw extraction available even if the worksheet displays a shorter value.
“None observed” means the stated check found no discrepancy. It does not claim that every character on the page was reviewed. Record the review scope beside the result.
For a replacement scan, open a new extraction run and link it to the original review. Verify document identity and revision again. A clearer file can still be the wrong edition.
Assign Accept, Review, Rescan, or Stop
These four outcomes are proposed operating rules. Adapt them to the receiving process and name who can make each decision.
Accept means page coverage is complete and all in-scope critical fields have evidence sufficient for the stated use. Record any permitted noncritical exception explicitly.
Review means the source is readable but the extracted value, context, or relationship needs a person to resolve it. Keep the affected output out of the receiving process until that review is complete.
Rescan means source quality prevents the required check. Ask for a clearer capture or an authoritative replacement file. Preserve the earlier source and explain which region needs attention.
Stop means the required document, authority, identity, or meaning is unavailable. A missing page, disputed revision, or unresolved critical instruction can stop release. Name the condition that allows work to resume.
Apply outcomes at both field and document levels. One accepted field can belong to a held document. Do not let a green cell silently release the entire sheet.
For S-17, the thickness field passes review after correction. The handling line requires a clearer source. Document acceptance remains stopped pending that evidence.
Release a Reviewed Record With Its Limits
Before releasing output, compare the final reviewed values with the worksheet. Confirm that every critical field has a resolved status and a source reference.
Name the acceptance owner. Have that person confirm the intended use and any exclusions. The extraction operator can correct text without owning the business decision that follows.
Retain the source reference when the text enters another system. A later business knowledge retrieval workflow needs to distinguish reviewed content from unresolved recognition output.
Use the same evidence boundary in an operations automation audit. Ask where extracted text first becomes actionable data and who accepts it there.
Close the review with a short receipt: document identity, checked page set, corrected fields, unresolved exceptions, decision owner, and permitted use. Keep the original extraction attached.
The page remains the evidence.
Sources
Adobe Acrobat: Recognize text in scanned documents. Image-only scans and the searchable text layer created through OCR. Checked 2026.10.28.
Google Cloud: Handle processing response. Recognized text, anchors, layout coordinates, and configured Enterprise Document OCR image-quality assessment. Checked 2026.10.28.
Microsoft Learn: Interpret and improve accuracy and confidence scores. Confidence meaning and custom-model field availability in the v4.0 documentation. Checked 2026.10.28.