An AI system can organise and interpret information, but it cannot reliably recover business facts that were never recorded. When records disagree, the application needs a source of authority and a way to surface uncertainty.
Data quality is part of the workflow design. It affects retrieval, calculations, model interpretation and the decisions people make from the result.
Start with the meaning of a field
Two systems may use the same label for different concepts. A sales record’s “customer” might be the person requesting a quote, while the accounting system expects the legal billing entity. Copying between them without a defined mapping can create a valid-looking but wrong record.
Document important fields, allowed values and ownership. Decide which source wins when values conflict and when a person must resolve the discrepancy.
A model may help identify a likely relationship in unstructured material. Its interpretation should remain distinguishable from a verified fact.
Trace a result through the pipeline
Consider an illustrative job-preparation workflow. A quote contains one address, an attachment contains another and the scheduling record has an incomplete unit number. A generated customer message may read smoothly while directing a crew to the wrong place.
The useful correction is to identify the conflict, stop the affected action and ask the responsible person to verify the address. Asking the model for a more confident answer does not create the missing fact.
This is a hypothetical example, not a client story or a measured result. Its purpose is to show how an upstream ambiguity can survive into a polished output.
Check extraction before interpretation
Documents can lose structure during parsing. A number can be separated from its heading, a checkbox can disappear or a scanned page can be read incorrectly.
Preserve page references and inspect representative extracted records. Validate required fields and formats before asking the model to interpret them. If the extraction failed, improving the final prompt may only disguise the problem.
For tables, test that rows, columns, units and exceptions remain associated correctly. Where exact amounts matter, keep source inputs and use an appropriate calculation path that can be rerun.
Keep freshness and permissions together
A source can be accurate when indexed and obsolete when retrieved. Establish how updates, deletions and access changes propagate into search indexes, caches and derived records.
A source owner should be able to identify the current version and its scope. Do not let several nearly identical policy copies compete without a way to distinguish the authoritative one.
Access is part of quality for the application. A correct answer based on a document the user is not authorised to see is still an unacceptable result.
Evaluate failures by stage
Use test cases that separate source errors, extraction problems, retrieval misses and generation mistakes. If everything is recorded as “the AI was wrong”, the team cannot choose an effective correction.
Keep a representative set of difficult inputs and the expected handling. Some should result in clarification or refusal rather than a completed answer.
Measure review and correction effort alongside output quality. A system that produces more material but creates more checking work may not improve the process.
Make data work an owned responsibility
Agree who maintains source definitions, approves corrections and handles unresolved records. The implementation team can build validation and reconciliation paths, but it cannot decide every business meaning without the organisation’s input.
Avoid a blanket claim that better data eliminates AI error. Models, tools and interfaces can still fail when the inputs are good. Data quality and application verification are complementary responsibilities.
The NIST Generative AI Profile includes data and measurement risks within a wider lifecycle view. Our Knowledge Systems and AI Operations Systems pages explain how source handling and workflow controls fit together.
Written by Anshul Kaushik, Managing Director at Ryzolv.
Back to Insights