Table extraction is not just reading text. The workflow must preserve which value belongs to which item, column, unit and page. Ask an agency to demonstrate those relationships on representative documents before using extracted rows to update inventory, purchasing or reporting.
Related service: AI automation development- 01Recover layout
- 02Map headers and units
- 03Validate row relationships
- 04Review and import
SECTION 01
Specify what a row means
Bring examples with merged headers, wrapped descriptions, repeated headings and tables spanning pages. Define whether a blank cell inherits a value from the previous row or genuinely means missing information. These are document-specific rules, not safe assumptions for every table.
Microsoft's layout documentation describes extracting table structure as well as text. Ask the agency to preserve cell or page references where the selected tool supports them. JSON Schema can validate the resulting row shape, but it cannot decide whether an amount belongs to the quantity column. Structural and business checks need separate acceptance examples.
SECTION 02
Illustrative example: two packaging units
Suppose a supplier sheet lists cartons and individual items under a shared quantity heading. The unit appears in a second header line, while one long item description wraps onto the next visual row. A naive import could shift the unit or turn the continuation into a new item.
The proposed workflow shows each extracted row beside its source region. It distinguishes a continuation from a separate product, keeps the source unit and checks any conversion against the approved catalog rule. A reviewer resolves the ambiguous row before import. The example does not assume AI knows the pack size or that the supplier's total is itself correct.
SECTION 03
Make discrepancies visible to the operator
Request checks for missing rows, duplicated repeated headers, misplaced decimal separators, multi-page breaks and subtotal lines. If a document provides a total, compare it with the interpreted rows under the agreed arithmetic rules. A discrepancy should explain which values differ rather than showing a generic extraction failure.
At handoff, receive the field map, sample documents, normalization rules and a correction history. Demonstrate a partial import being blocked or clearly managed according to the agreed policy. Where suppliers can provide reliable structured exports, compare that route with PDF extraction: fewer interpretation steps may be more useful than a visually impressive but fragile document pipeline.
SECTION 04
Frequently asked questions
Is a matching document total sufficient?
No. Incorrect rows can offset one another or use the wrong units. Check row identity and relationships as well as arithmetic.
Should all blank cells become zero?
Only where a documented rule makes that interpretation valid. Otherwise preserve missing values for review rather than inventing numerical meaning.
Can staff correct an extracted row?
Yes, if the workflow records the original value, the correction and its source context. Corrections should not obscure what the document originally contained.
PRIMARY REFERENCES
Sources and further reading
These references cover the standards, platforms, or published prices discussed in the guide. Worked examples and checklists are our editorial guidance.
- Document layout analysisMicrosoft Learn
- JSON Schema object validationJSON Schema
EDITORIAL METHOD
About this guide
We use AI to assist with drafting and editing. Catapult AI Work is responsible for the published content. Examples illustrate possible approaches; they are not client case studies unless identified as such.
Budget examples are not Catapult package prices. Check linked provider pages for current fees and plan limits before making a purchase.
Read the editorial policy