Expense Exception Triage

Goal

Generate an exception log, summary, and policy-gap report from receipts, bank/card exports, policy documents, and approval threads so finance knows what blocks each claim and who must act before payment.

Inputs to gather

• Receipts (images/PDFs – primary proof of spend)
• Bank/card export files (CSV/XLSX – authoritative transaction data)
• Expense report export from ERP (optional but improves matching)
• Expense policy document (rules to test against)
• Approval matrix / authority levels (to validate approvers)
• Email or chat threads with approvals or clarifications (evidence of authorization)
• Employee-to-manager roster (for routing owners)
• Currency and FX policy (for conversions)
• Time window and business units in scope (limits dataset)
• Payment hold status or SLA due dates (sets urgency)

Before doing any work, ask the user for these inputs in ONE message. Skip anything they already provided. If they tell you to decide, choose sensible defaults and say what you chose.

Workflow

1. Confirm scope
   a. Reiterate the time window, business units, and payment hold status.
   b. List any missing inputs; if absent after confirmation, treat gaps as exceptions.

2. Normalize and parse sources
   a. Standardize dates to ISO-8601 and amounts to the base currency per FX policy, noting the rate source.
   b. From receipts extract merchant, date, total, currency, tax, line items, last 4 of card, attendees, and business purpose notes.
   c. From bank/card exports extract transaction_id, post/auth dates, amount, currency, merchant, MCC/category, and employee/cardholder.
   d. From the policy pull thresholds, limits, prohibited items, approval rules, deadlines, and documentation requirements.
   e. From emails/threads capture approver, date, scope, clarifications, and promised follow-ups.

3. Match receipts to transactions
   a. Link each receipt to a transaction using amount, date proximity, merchant similarity, and card last 4.
   b. Detect duplicates and suspicious split transactions.
   c. If multiple matches exist, choose the highest-confidence link and note alternatives with confidence scores.

4. Apply reason code taxonomy
   Use the predefined list (MISSING_RECEIPT_OVER_THRESHOLD, UNREADABLE_RECEIPT, NO_BUSINESS_PURPOSE, etc.). Extend only when essential and record new codes in the summary.

5. Evaluate each claim
   a. Check documentation vs. thresholds.
   b. Test category and amount against limits and per-diem rules.
   c. Verify timeliness of submission.
   d. Confirm approvals and authority.
   e. Assign one or more reason codes with concise details citing evidence and policy clauses.

6. Assign ownership and next actions
   a. Missing or unclear docs → owner: employee; action: supply info.
   b. Approval gaps → owner: correct manager/approver; action: approve or deny.
   c. Policy gaps → owner: finance/policy admin; action: clarify or draft rule.
   d. Duplicates/splits → owner: finance reviewer; action: confirm/de-dupe or request explanation.
   e. Set due-by dates per SLA.

7. Create artifacts
   a. Build exception_log.csv (or .jsonl) with all required fields.
   b. Produce summary.md containing counts by reason code, owner, severity, unblock recommendations, and any new reason codes.
   c. Generate policy_gaps.md listing POLICY_GAP_UNSPECIFIED findings with proposed wording.

8. Summarize for the user
   Provide a brief report of key stats and next steps. Offer to draft reminder messages to owners if requested.

9. Guardrails
   • Do not approve, deny, or schedule payments.
   • If evidence is ambiguous, set severity to “warn” and note clarifications needed.
   • Preserve PII; do not send data to external services.
   • For long or complex datasets, check in with the user after steps 3 and 6 before proceeding.

Output

Deliver a ZIP (or shareable folder) containing:

1. exception_log.csv (or .jsonl) – one row per claim/transaction with reason codes, owners, next actions, and confidence.
2. summary.md – counts by reason, owner, severity, unblock recommendations, and any newly created reason codes.
3. policy_gaps.md – unclear or missing rules with suggested text.
   (Optional) Draft reminder messages to owners if the user requested them.