Cross-document validation compares agreed fields across related records and explains any disagreement. It should preserve the original values, show the comparison rule and ask an authorised person to resolve genuine conflicts. The software should not choose whichever value looks most plausible.
Before building, define the relationship between the records. An invoice may cover several visits, while a service report covers one. Comparing every field as if the documents described an identical object will create misleading alerts.
A fictional pack with two different problems
For job J-418, an office receives purchase order PO-76, a service report and a draft invoice. The purchase order names “Depot North”, the report names “Depot South”, and the invoice uses site code DN-02. The report’s job reference is written as “J 418”.
Removing the space from the job reference may be a safe formatting rule if the business confirms that convention. Deciding that South must mean North is not. The first comparison can match; the second needs review.
| Field | Observed values | Proposed result |
|---|---|---|
| Job reference | J-418 / J 418 | Match under approved formatting rule |
| Site | Depot North / Depot South | Conflict requiring confirmation |
| Site code | DN-02 / no code on report | Incomplete comparison |
Keep incomplete and conflicting results separate. An absent site code does not prove that the report belongs elsewhere. Equally, a matching job reference should not suppress a conflicting site name.
Define comparisons before using a model
Start with exact identifiers and approved reference tables. A maintained site list could link DN-02 to Depot North. Record which version of that list was used. Use flexible text matching only to propose candidates when exact matching cannot resolve the relationship.
For every field, specify whether comparison is exact, normalised, numeric within an agreed tolerance, or manual. Avoid applying one permissive similarity threshold to everything. A slightly different trading name and a slightly different bank account deserve very different treatment.
- Preserve both source values and their locations.
- Record the relationship being tested, such as report-to-job.
- Show the normalisation applied before comparison.
- Separate missing, unreadable, conflicting and matching states.
- Require a reason when staff override a finding.
Make resolution a recorded action
The reviewer of J-418 might determine that the report was attached to the wrong job. The next action is to obtain the correct report, not edit its site field in a database and treat the document as corrected. Preserve that distinction in the interface.
A useful resolution record says what was wrong, which replacement or clarification was accepted, who accepted it and which downstream status changed. If the evidence changes again, the original decision remains visible but may no longer establish current readiness.
NIST’s AI framework includes responsibilities, measurement and ongoing management. Our recommendation here is to make those responsibilities concrete at the individual exception: someone owns the decision and can explain its evidence.
Test legitimate differences
A customer may use a central billing address and a separate delivery address. A consolidated invoice may list a parent company while reports name its sites. These are not necessarily errors. Model the roles explicitly instead of comparing a generic “address” or “company” field.
Also test a correct value copied into the wrong record. Matching text is weak evidence if the document-to-job association is wrong. The comparison should begin with record identity, then examine the fields inside the matched set.
Start with one relationship
Choose a recurring pair of documents and a short list of fields. Label representative matches, true conflicts and acceptable differences with an operational reviewer. Run the proposed checks without changing live records, and inspect both the alerts and the cases passed quietly.
A custom review workflow earns its place when those specific comparisons are difficult in the existing tools. Measure false alerts, missed conflicts and resolution effort before expanding to another document family.
Sources and further reading
AI-assisted research and writing, checked against the linked sources. Worked examples are fictional. Product features can change; confirm the details in your own account.