Evaluation method
Financial Data Quality Checks Before AI Analysis
This checklist concerns the input before a report is generated. It complements a separate review of the finished report. A calculation can be internally consistent and still answer the wrong question when its source range or meaning is wrong.
Write a small source contract
Define required columns, row meaning, period, currency, amount basis and accepted date/number format. Keep the contract beside the source extract. Map columns through a human-approved profile instead of relying on a model to reinterpret headers each time.
A blank cell, a zero and an excluded row are different states. A missing actuals dataset must not become zero expenses. Keep an explicit reason when a row cannot be used.
Distinguish duplicates from repeated business events
Check stable record identifiers and relevant fields before removing anything. Two equal payments can be legitimate. Two rows with the same source identifier may be a duplicate export, but a person must confirm the intended source record.
Do not deduplicate solely on amount and date. Preserve an original file and a cleaned revision with the changed rows listed. Silent deletion can make a tidy forecast less accurate than the source.
Refuse silent currency mixing
Check a source currency column when present. A profile currency alone does not authenticate every source row. Mixed currencies require an explicit conversion policy, rate source and date before amounts can be compared.
Do not infer a rate from a currency symbol or ask a model to choose an undocumented conversion. The dedicated source-bound process refuses incompatible source currency rather than silently aggregating it.
| Input issue | Required decision | Record |
|---|---|---|
| Mixed currencies | Separate or explicitly convert outside the unsupported aggregation | Currency/rate/basis |
| Ambiguous date | Confirm source format | Approved mapping and example |
| Duplicate identifier | Verify original record | Kept/removed row and reason |
| Unknown category | Approve mapping or leave unresolved | Mapping revision |
| Missing period | Obtain source or retain the gap | Completeness limitation |
Check signs and dates with representative rows
Test one receipt, one payment, a boundary date and an overdue item before selecting the full range. Confirm whether a payment is a positive magnitude with a separate direction or a negative signed amount. Mixing both conventions can reverse cash movement.
A locale affects how strings are parsed; a time zone affects period boundaries. Confirm both in the profile. Use small explicit examples rather than trusting visual formatting in the workbook.
Retain negative test cases
The public bad-data-cases.json lists synthetic cases and the observed boundary of the exercised checks. Inspect the expected and actual outcomes. A refused input is a useful result when the alternative would be an unverified report.
This is not a claim that every possible error is detected. Review the input contract again after a source-system change. Then verify the generated result using the separate report-checking method.
Sources
Continue with your workflow
- How to Verify Numbers in an AI Financial Report
- What Happens to Cash When a Customer Pays Late?
- Financial Agent: Supported Files, Formulas and Refusal Limits
- 13-Week Cash Flow Template and Verifiable Example | Braio
- Braio help - first workflow, QA and SEO/GEO reports
- AI agent security - access, approvals and costs | Braio

