A neat spreadsheet can hide a broken table.
The first page ends with a parts row. The second begins with another header. A shared unit cell spans two part rows. An extraction returns rectangular cells, but one quantity belongs to the next part.
Check the table while its source pages are still beside it. Establish which cells form each record before exporting a CSV or importing values elsewhere.
The method below separates structural checks, arithmetic checks, and business acceptance. Its parts schedule and extraction defects are fictional. The worksheet defines a proposed review process, not a vendor guarantee.
Define What One Source Row Means
Start with the destination's intended use. In this example, an operator prepares a draft parts comparison for an internal reviewer. The import does not place an order or approve a substitution.
Define a record as one identified part line with a quantity, a unit, and a reference back to the document. Keep section labels, headers, subtotals, and notes as distinct source elements.
Write the record definition before counting rows. A source table can contain ten visible lines without containing ten part records. Your count depends on the role each line plays.
Preserve the original PDF and raw extraction. Add a reviewed working table separately. Give each working record a source page and visible row reference.
When a description wraps, count the part identifier and its associated fields together. Do not assume every line of text is a new record. When a cell is merged, record its extent before filling blank cells below it.
This work happens before AI data entry automation. An import needs records whose boundaries and meanings have already been checked.
Read the Layout Result for Your Verified Model
Azure Document Intelligence v4.0 layout documentation identifies prebuilt-layout with API version 2024-11-30 GA. Table results describe row and column counts, cell indexes, row and column spans, bounding regions, and text spans. Text spans reference the response's content. Cell spans describe the grid. Those are different relationships.
The same documentation says table bounding regions exclude associated captions and footnotes in this version. Inspect the surrounding page. A returned table region is not the complete business context.
Google's Document AI response guide describes page tables with header and body rows. Its Form Parser guidance limits extraction to conventional tables without merged cells; the documented rowSpan and colSpan values are always one. Do not transfer Azure capabilities to that processor.
Save the actual processor identity and version from your run. The Google guide's stated limitation does not verify every current processor version. Inspect your own configuration before relying on its fields.
No vendor call runs in this example. The defects below are invented review cases. They illustrate questions to ask of an extraction, rather than measured performance of either product.
Build a Small Fictional Source Table
Imagine schedule P-24, revision C, spread over two pages.
Page one begins with a merged heading across the quantity and unit columns: “Required quantity and unit.” Below it, individual column labels read “Part,” “Description,” “Quantity,” and “Unit.” The unit cell “each” spans the A-10 and A-11 rows vertically.
Page two repeats the column header before the remaining parts. A footnote below the final subtotal explains that lengths are per part line and exclude spare material.
The following table transcribes the fictional source after a reviewer has inspected those merged cells. It is an interpretation of the stated source layout, not raw extraction output.
| Source page and row | Part | Description | Quantity | Unit |
|---|---|---|---|---|
| Page 1, row 1 | A-10 | Corner bracket | 12 | each, shared source cell |
| Page 1, row 2 | A-11 | Mounting plate | 8 | each, shared source cell |
| Page 1, row 3 | L-20 | Spacer stock | 2.5 | m |
| Page 2, row 1 | L-21 | Trim stock | 1.5 | m |
| Page 2, row 2 | A-12 | End cap | 4 | each |
| Page 2, subtotal | None | Counted parts subtotal | 24 | each |
| Page 2, subtotal | None | Stock length subtotal | 4.0 | m |
Every identifier and quantity is synthetic. The two subtotals summarize different units. The footnote applies to the two stock lines. Keep that relationship visible in the working record.
Record the merged heading separately if it helps explain the columns. It does not become a part description. Preserve the shared unit cell's source extent so another reviewer can inspect why both parts inherit “each.”
Check Structure Before Calculating Anything
Compare the first and last source record on each page with the working table.
For P-24, page one ends at L-20. Page two begins at L-21 after its repeated header. Verify both identities and their values before joining the page results.
In the fictional raw extraction, the page-two header appears as a data row. Classify it as a header using the source location and labels. Exclude it from part counts while preserving it in the raw result.
Next, compare source records with working records one by one. The expected inventory is five part records, two subtotal records, one relevant footnote, and the source headers. Record those counts by role.
A wrapped description must stay attached to its part. A merged unit must keep its source relationship. Do not fill every empty unit with the preceding value. That shortcut would assign “each” to stock length when the unit changes.
Check empty cells explicitly. Distinguish a source blank from an extraction omission and from a value inherited through a verified merged cell. Use a status field when the difference remains unresolved.
Keep subtotal rows outside the part dataset. Otherwise a later count can add the summary back into the underlying quantities.
Reconcile the Numeric Meaning
Now compare each value as text before converting it into a number.
The fictional extraction reads L-20 as “25 m.” The source shows “2.5 m.” Record the decimal discrepancy and the reviewed correction. Keep the original characters available for inspection.
Check signs even when the example contains only positive quantities. If a future source contains a minus sign or parentheses, establish their meaning from that document. Do not normalize punctuation before review.
Check unit labels with the quantities. Preserve “each” and “m” as distinct fields. Conversion requires an explicit rule and an accepted reason. Nothing in this example authorizes converting a length into a part count.
Calculate the two stated relationships independently: 12 + 8 + 4 = 24 each, and 2.5 + 1.5 = 4.0 m. Compare those results with the source subtotals.
Agreement supports these arithmetic checks. It does not prove the identifiers or descriptions are correct. Two swapped quantities can preserve a subtotal while changing the record meaning.
If the printed subtotal disagrees with the inspected rows, preserve both. Ask the document owner to resolve the difference. Changing a row to make the sum fit would manufacture evidence.
Use a Reconciliation Worksheet
Give the worksheet a heading with document identity, revision, extraction run, expected page set, intended use, and reviewer.
Use one row per discrepancy or checked relationship. The examples below refer only to fictional schedule P-24.
| Source page and row | Extracted value | Expected relationship | Review result | Unresolved difference |
|---|---|---|---|---|
| Page 1, A-10 and A-11 unit cell | A-11 unit blank | Shared source cell covers both part rows | Assign each to both with source reference | None after visual check |
| Page 1, L-20 | 25 m | Source text reads 2.5 m | Decimal corrected in reviewed copy | None after visual check |
| Page 2, repeated header | Part, Description, Quantity, Unit | Header precedes L-21 | Exclude header from part records | None |
| Page 2, counted parts subtotal | 24 each | 12 + 8 + 4 = 24 | Arithmetic agrees | Business acceptance pending |
| Page 2, stock subtotal | 4.0 m | 2.5 + 1.5 = 4.0 | Arithmetic agrees | Business acceptance pending |
| Page 2, footnote | Footnote absent from table output | Applies to stock lengths | Attach inspected note to those lines | Owner must accept intended use |
A reviewed correction needs its author, date, and source reference. Store those fields beside the worksheet when the table becomes too wide.
Keep unresolved differences specific. “Needs checking” hides the next action. “Document owner must confirm whether revision C applies to this comparison” assigns a concrete question.
Do not replace an unknown value with zero. Zero makes a numeric claim. Keep a blank plus an explicit unresolved status when the source cannot support a value.
Ask for Business Acceptance Separately
A structural pass means the working records reflect the inspected grid and page boundaries. An arithmetic pass means the stated calculations agree under the recorded units and rules.
Business acceptance answers whether those records are suitable for the named use. Assign that decision to the receiving process owner.
For P-24, the reviewer has accounted for five parts and both subtotals. The decimal is corrected. The footnote is linked to the stock lines. The owner still needs to accept revision C as the appropriate basis for the draft comparison.
If revision applicability is unknown, hold release. Accurate extraction of the wrong revision does not meet the task.
Permit partial release only when the owner defines an independent subset and accepts its exclusions. Record which rows are released and which remain held. A filtered export needs its own row count and source references.
For a later Google Sheets workflow, carry those references into the receiving worksheet. Keep review statuses out of numeric quantity columns. A person should be able to follow a cell back to its source without reconstructing the entire extraction.
Inspect the Export Before the Import
Create the tabular export only after the review decision. Open the exported file separately and compare its records with the accepted working table.
Check identifiers, decimal characters, unit fields, and text containing separators. Confirm that headers remain headers. Verify that excluded subtotals and held rows did not reappear through an export filter.
Record the final part count and the accepted scope. For the complete fictional example, that count is five. The two subtotals remain reconciliation evidence outside the imported parts dataset.
Keep the footnote relationship in a field or linked review record the receiving process can preserve. Do not assume every import destination supports the same notes or cell structure.
If the result feeds business reporting, state which units each metric includes. Adding every quantity column together would mix counted parts and lengths.
Close with the document revision, accepted rows, unresolved differences, owner, and destination. Keep the PDF and raw extraction available beside that receipt.
Reconcile the rows before they move.
Sources
- Microsoft Learn: Document layout analysis. The prebuilt-layout model, API 2024-11-30 GA, cell indexes and spans, source positions, and table-region exclusions. Checked 2026.10.28.
- Google Cloud: Handle processing response. Page table structure and the documented Form Parser conventional-table limitation. Checked 2026.10.28.