AI Explanations
When an invoice is flagged for review, VaultAP generates a plain-language explanation summarizing why. These explanations are produced by AI to help reviewers quickly understand the risk signals — but they are never used to calculate scores or make approval decisions.
AI explanations are supplementary. The risk score is calculated entirely by deterministic rules. AI never influences the score or decides whether an invoice is approved or rejected.
When Explanations Are Generated
AI explanations are generated only for flagged invoices — those routed to standard review or high-alert. Invoices that are auto-cleared do not receive AI explanations, since no human review is expected.
What an Explanation Includes
Each AI-generated explanation provides:
- Which rules fired — A summary of the specific rules that contributed to the score, written in plain language rather than rule codes.
- Context — Relevant background such as the vendor's history, how the current invoice compares to past invoices, and what changed.
- Suggested actions — Practical next steps the reviewer might take, such as confirming bank details with the vendor or cross-referencing a purchase order.
Example
This invoice was flagged because the vendor's bank account details differ from the last three payments (BANK_CHANGE). The invoice also arrived without a purchase order reference (NO_PO_MATCH). Consider contacting the vendor through a known phone number to verify the bank change before approving.
Limitations
AI-generated explanations have inherent limitations that reviewers should be aware of:
AI explanations can be inaccurate or misleading. Always verify the explanation against the raw rule flags and invoice data before acting on it.
- AI can be wrong. The explanation may mischaracterize why a rule fired or provide incomplete context.
- Explanations are not exhaustive. The summary may omit nuances that are visible in the raw data.
- Context is limited. The AI works from the data available in VaultAP. It does not have access to external systems, email threads, or verbal agreements.
- Suggested actions are guidance, not policy. The AI does not know your organization's specific approval workflows or escalation procedures.
Best Practices
- Treat AI explanations as a starting point, not a conclusion.
- Always review the raw flags panel alongside the explanation. The raw flags show exactly which rules triggered and their individual weights.
- If an explanation seems incorrect, rely on the rule data. The deterministic score is the source of truth.
- Provide feedback on poor explanations using the thumbs-down button — this helps improve future explanation quality.