Measure source-citation coverage at the finding level, not by counting links in an overall report. Every material AI-assisted AR finding should identify the authorized source row or document fragment that supports its amount, status, timing evidence, and recommended internal action. Findings that cannot be retrieved to evidence should be labeled unsupported and routed to human review rather than blended into a confident forecast.
NIST's AI RMF core emphasizes documented testing, evaluation, verification, and validation processes and notes that documentation can improve review and accountability. NIST's discussion of validity and reliability centers objective evidence, ongoing testing, monitoring, and human intervention when a system cannot detect or correct errors.
Define the unit of coverage
Treat each material assertion as a reviewable unit. Examples include:
- invoice open balance;
- customer identity and invoice number;
- dispute or promise status;
- evidence date;
- proposed collection range;
- internal owner; and
- next action.
A paragraph with one invoice link is not fully covered if it makes several claims that the source does not support. Store citations beside the exact fields or assertions they support.
The source-lineage guide defines the row-to-finding chain. Citation coverage tests whether that chain is present and usable across the output population.
Track four practical measures
Use measures that can be reproduced from the review record:
- **Finding coverage:** Share of material findings with at least one retrievable supporting source.
- **Amount coverage:** Share of the total reviewed invoice amount whose balance and allocation cite the relevant source rows.
- **Citation validity:** Share of sampled citations that open the claimed source and support the stated field.
- **Unsupported-finding rate:** Share of material findings that lack adequate evidence and should be withheld or escalated.
Do not invent a universal pass threshold. Set the required coverage by use case and risk, document it before evaluation, and treat balance-changing or timing-changing outputs as higher consequence than descriptive summaries.
Test retrieval and support separately
A citation can be retrievable but wrong. Test whether the referenced row exists, matches the invoice and customer, carries the cited field, uses the correct snapshot, and supports the conclusion. A note that mentions a dispute may support an exception label without supporting an exact collection date.
Use the AI AR evaluation-set guide to include clean, missing, conflicting, stale, and adversarial cases. The deterministic-controls boundary keeps arithmetic validation separate from judgment review.
Preserve the reviewer disposition
For each sampled finding, record pass, partial support, unsupported, or source unavailable; reviewer; review date; model or rule version; and corrective action. The finance AI review record provides a durable place for that evidence.
Decision rule
An AI-assisted finding is reviewable only when a human can retrieve the cited source and see that it supports the claimed amount, status, or timing evidence. Missing coverage should trigger abstention or review, not an invented citation.
AR Pulse's approved product boundary emphasizes source-row traceability and human review for an authorized QuickBooks receivables export. It does not claim universal AI accuracy or promise collection outcomes.