AI-extracted text can look tidy while still misreading a digit, dropping a footnote, or attaching the right value to the wrong row. Before you reuse it, compare the output with the original image—not just with what you expect the document to say.
This guide offers a manual checking workflow for screenshots, scans, and photographed documents. The example below is entirely fictional: it is not a product test, a real receipt, or a measured error rate.
1. Keep the original and the raw extraction
Save the source image unchanged, then keep the first extracted text separately from your corrections. Give both files a matching reference, such as sample-page-01. If there are several pages, confirm their order and count before reviewing individual words.
Place the image and text side by side. Zoom enough to inspect small characters, but return to the full page regularly so you do not miss a heading or note outside the crop. A tighter crop can help you inspect one area; it should not replace the full original.
Before uploading anything, check whether the service and account are approved for the document. Remove unnecessary sensitive details from a separate copy, and check that redactions are actually applied. Do not upload passwords, private identity records, confidential work material, or someone else’s personal information merely to try an OCR tool.
2. Check coverage and reading order first
Start with the page’s structure: title, columns, paragraphs, table, labels, and notes. Tick off each visible region against the extraction. A perfectly spelled paragraph is not a complete transcription if the second column disappeared.
Google’s Cloud Vision OCR documentation distinguishes general text detection from document-oriented extraction. Its responses can include word bounding boxes or a page-to-word structure, depending on the feature. If your tool exposes location information, use it to trace an extracted word back to its position in the image.
Read columns in their intended order. For tables, compare each row label with its values rather than reading a flattened list of numbers. If your app only provides plain text, manually add row labels to a review copy; do not assume the line breaks preserve the original layout.
3. Work through a fictional example
Imagine a photographed workshop order sheet whose visible text is specified in the middle column below. The extraction in the next column is deliberately wrong. These invented entries illustrate four different checks, not the output of any named AI system.
| Check | Original image says | Hypothetical extraction | Verified correction |
|---|---|---|---|
| Character | Order B108 | Order B1O8 | Replace letter O with digit 0. |
| Decimal | Unit price: $12.50 | Unit price: $1250 | Restore the visible decimal: $12.50. |
| Row alignment | Blue folders: 6 Green folders: 9 |
Blue folders: 9 Green folders: 6 |
Attach 6 to blue and 9 to green. |
| Missing note | *Delivery not included. | No footnote extracted. | Add the footnote exactly as shown. |
The row-alignment example matters because both numbers are present and readable, yet the meaning is wrong. The missing note matters because it changes how a reader interprets the price. Checking spelling alone would not catch either issue.
4. Inspect meaning-changing characters
Review identifiers, dates, quantities, decimal separators, minus signs, percentages, currency symbols, and units character by character. Check look-alike characters such as O/0 and I/1 against the actual image. Do not change an unusual identifier simply because another version looks more familiar.
Preserve the displayed date or number format in the transcription. For example, do not silently convert 03/04 into a named month when the source does not establish the convention. Any later normalization should be a separate, documented step.
Arithmetic can flag a possible inconsistency, but it cannot prove what the image says. If a total looks surprising, recheck the digits and labels; do not rewrite the source to make the calculation balance.
5. Mark uncertainty instead of filling gaps
If the image is blurry, record a clear marker such as [unclear: final digit] in your review copy. Keep that marker separate from the source wording. Ask for a clearer scan or consult the original document when the uncertain field matters.
A prompt can help organize another pass: “Transcribe the visible text without correcting it. Preserve row labels and footnotes. Mark unreadable characters rather than guessing.” Treat the response as another draft to inspect, not independent confirmation of the first answer.
Amazon Textract’s best-practice guidance recommends considering confidence scores alongside the sensitivity of the use case and flags irregular or merged table cells as potential difficulties. Use available confidence information to prioritize review, not to skip checking a critical field against its source.
6. Save a reviewable final version
Keep a short correction log: source page or region, raw extraction, corrected text, and any unresolved issue. Separate faithful transcription from edited prose. Fixing the author’s spelling or expanding an abbreviation may be useful later, but it is not the same task as recording what is visible.
Before sharing, confirm that every page and note is accounted for, important values match their labels, and uncertainty remains visible. If a crucial field is unreadable, label the extraction incomplete rather than calling it verified. For spoken rather than printed material, use the separate AI transcript review guide.
Official documentation checked October 29, 2026. The workflow and fictional examples are editorial guidance, not a benchmark or an endorsement of a particular OCR product.
