Test the revised template against approved examples before returning it to normal processing. Compare required fields, field meanings, layout assumptions and version identifiers, then check the resulting documents and destination records. Include cases that should stop for review. A successful run means the expected information reached the expected place, with exceptions handled correctly.
The following is a proposed testing method for a document processing workflow. It applies when a client changes an incoming form, an outgoing document template, or both. The practical decision is whether that specific revision is ready to resume.
Establish exactly what the client changed
Ask for the previous template, the revised template and completed examples of each. A blank template shows available spaces; a completed example reveals wrapping text, populated tables and page breaks. Keep the original files so that later corrections do not overwrite your test evidence.
Write a short change record covering renamed labels, moved fields, added sections, removed fields and changed instructions. Confirm whether a renamed field still means the same thing. “Delivery address” becoming “Site address” might be cosmetic, or it might change which location the operator must use.
Separate appearance changes from business changes. Moving an address box needs a mapping check. Changing which address belongs there needs a client decision before testing can establish correctness.
Record when the revised template takes effect and whether older documents remain acceptable. During a transition, the workflow may need to recognise both versions. An unfamiliar version should enter review under your proposed operating rule.
Compare fields, layout assumptions and identifiers
Create a field comparison showing the old label, new label, meaning, required status, expected format and destination. Include fields that remain unchanged: moving a table can affect surrounding content even when its labels stay identical.
For each required field, specify what counts as usable evidence. A customer name appearing somewhere on the page does not establish whether it identifies the buyer, recipient or contact person. Check the relationship between labels and values.
Microsoft documents extraction of text, tables and document structure, as well as custom template models for static layouts. These capabilities explain why layout deserves explicit testing; they do not establish that a particular revised form will work. Source: Azure Document Intelligence overview
Check assumptions such as “the reference appears in the header” or “the final table row contains the total”. For generated documents, inspect long names, multiline addresses, repeating rows, signatures and footer placement.
Keep the client’s template identifier separate from the extraction configuration and workflow revision. If the client supplies no identifier, assign an internal revision linked to the exact file. Do not rely solely on a filename that someone can reuse.
Build samples with expected answers
Select samples around the changes. Include an ordinary completed document, a long or multipage example, a required value missing from its new location, conflicting values, a duplicate submission and an older accepted template. These are coverage categories, not a universal sample-size requirement.
Before running anything, have the process owner write the expected values and expected handling for each sample. Otherwise, reviewers can accidentally accept whatever the automation produces. Store the expected result beside the source document, including the page or section supporting each important value.
If AI produces structured data, test content separately from structure. OpenAI documents schema-constrained output and describes exceptions such as refusals and incomplete responses. Matching the expected structure is therefore only one check; compare values against the approved source answers as well. Source: OpenAI Structured Outputs
Use authorised examples or fabricated documents with realistic layout variation. If you remove sensitive content, preserve the lengths and page structure needed for the test.
Follow the result through the whole workflow
Run samples in an isolated test destination with external sending and operational updates disabled. Check the extracted values, any generated document, the exported data and the receiving record. Opening a correct-looking PDF does not verify a separate spreadsheet export.
Where the workflow uses tool calling, OpenAI describes the application executing code in response to the model’s request. Consequently, checking a proposed action is insufficient: inspect what the test application actually did. Source: OpenAI function calling
For generated documents, compare content and appearance separately. Confirm that the correct client details and template revision were used, then inspect clipping, overlapping text, missing pages and misplaced labels. Do not fail an otherwise correct document merely because an approved design change moved a heading.
For receiving systems, read back the stored test record. Check field destinations, attachments and review status. Also verify that a stopped document leaves no partial operational record that an operator might mistake for completed work.
Work through ordinary and exception cases
Consider a hypothetical maintenance business whose client moves the site address into a table, renames “Job number” to “Work reference” and adds a document revision. All identifiers and quantities below are hypothetical; handling rules are proposed.
Ordinary case: The revised form contains work reference JOB-042, an unambiguous site address and three task rows. Expected handling is to preserve the reference, extract the address and retain every row. The human reviewer compares the generated job sheet and destination record with the approved answers before marking the sample passed.
Missing case: The new site-address cell is blank, but the client’s head-office address remains in the footer. Expected handling is a review hold for the missing site address. The operator requests clarification from the client; the workflow must not quietly substitute the footer address.
Ambiguous case: The cover page and task table show different work references. Expected handling is to retain both candidates with their locations and pause the affected document. The operator confirms the intended reference with the client, records the decision and reruns the corrected sample.
Duplicate case: The same completed form arrives again under a different filename. Expected handling is to identify it as a possible repeat before creating another job sheet or record. The operator checks whether it is a resend or a genuine amendment. Changed content with the same reference needs amendment review rather than automatic deletion.
Finally, rerun an older approved form. If support for it continues, it must still reach the correct processing path. If support has ended, it should receive the agreed review outcome rather than being interpreted as the new layout.
Template-change regression checklist
Copy this checklist into the change record. Complete each item with evidence rather than a tick based on expectation.
- Identify the client, document purpose, change owner and intended restart date.
- Preserve the previous and revised template files and record their identifiers.
- Confirm the effective date and whether older versions remain accepted.
- Compare field labels, meanings, required status, formats and destinations.
- Record changed layout assumptions, tables, page breaks and repeated sections.
- Record the extraction configuration, output template and workflow revisions tested.
- Prepare ordinary, long, missing, ambiguous, duplicate and older-version samples.
- Approve expected values and handling for every sample before execution.
- Run against isolated destinations with operational actions disabled.
- Compare extracted values with source evidence and approved answers.
- Inspect generated documents visually and read back destination records.
- Verify review holds, duplicate handling and absence of unintended partial records.
- Record failures, corrections and rerun results for affected and previously passing cases.
- Name the reviewer, record the restart decision and document the recovery procedure.
Decide whether to resume and hand over
A useful proposed release rule is that every required-field and routing case must match its approved outcome. An exception sample passes when it reaches the intended hold, not when the workflow forces it through. List any accepted cosmetic differences explicitly.
Give the operator the accepted template identifiers, examples of review holds, unresolved limitations and the person to contact. The recovery procedure should explain how to pause the changed path, locate affected documents and reconcile existing records before replaying anything. Restoring an old configuration does not make a new layout compatible.
Check current product, account, plan and region eligibility before deployment if the change requires a different service capability. For broader design choices, our AI agents versus automation comparison helps distinguish fixed checks from tasks needing interpretation. A custom AI agent still needs this evidence.
FAQ: testing a revised client template
Must we retest if only the logo changed?
Run a focused comparison to confirm that fields, pagination and version recognition remain intact. Expand testing if the new logo shifts nearby content or changes the document’s processing path.
Can a successful extraction count as approval?
No. Check the resulting document and receiving record, plus the expected review handling. Extraction success does not establish that the address reached the correct field or that duplicate records were prevented.
What if we cannot distinguish old and new templates?
Hold uncertain documents for review while agreeing a reliable version marker with the client. Test that marker against both layouts. Our custom AI agent workflow guide provides broader workflow context.
If your business needs help making template changes reviewable within its AI automation, get in touch with the previous template, revised template and representative samples.