---
name: expense-policy-audit
description: Check every expense report against policy, approve what's clean, and queue only genuine exceptions. Use this skill when: New expense report submitted — Fires as reports land in the expense sheet or inbox.
---

# Audit expense reports against policy

Goal: check every expense line against policy and queue only the genuine exceptions.

Connected tools this skill uses: googlesheets, gmail, drive, slack.

## When to use

- New expense report submitted — Fires as reports land in the expense sheet or inbox.
- Or when the user asks directly to check every expense line against policy and queue only the genuine exceptions.

## Method

Audit expense reports submitted since the last run.

1. Pull new expense reports from {{expense_sheet}}, with receipts from
   {{drive_folder}}.
2. For each line item, check that the receipt matches the claimed vendor,
   amount, and date exactly.
3. Check each line against {{policy_doc}}; cite the exact clause behind
   every approve or flag decision.
4. Before you approve any line, confirm both checks passed, a policy clause
   was cited, and the receipt matches the claim exactly. If either check
   fails, move the line to the exception queue instead of approving it.
5. Mark clean, in-policy, receipt-matched lines as approved in
   {{expense_sheet}}.
6. Post the exception queue to {{review_channel}}: each line, the policy
   clause it triggered, and why it wasn't auto-approved.

Do not approve a line the policy doesn't clearly cover, and do not issue
reimbursement — approval marks the line ready, payment still needs
sign-off.

## Expense policy reader

Apply the written expense policy line by line, citing the exact
clause behind every approve or flag decision.

## Method
1. Re-read `{{policy_doc}}` fresh each run, not a cached summary, so
   policy updates take effect immediately.
2. For each expense line, identify the policy clause that governs its
   category and amount.
3. Compare the line's amount and category against that clause's limit
   and conditions.
4. Record the clause reference alongside the decision, not just the
   decision itself.
5. If no clause clearly applies, do not approve by default — route to
   ambiguity escalation instead.

## Rules
- Never approve a line without a cited policy clause attached to the
  decision.
- Never issue or authorize reimbursement — approval marks the line
  ready for payment sign-off, nothing more.
- Approval limit above which a line always escalates regardless of
  policy fit (set on first run, then stored in memory).
- Re-read the policy document every run; never rely on a memorized
  version from a prior run.

## Output
A decision per line: `{line_id, decision, policy_clause, reasoning}`.

## Ambiguity escalation

Escalate any expense line the written policy doesn't clearly cover,
instead of guessing.

## Method
1. Take lines where the policy reader found no clearly applicable
   clause.
2. Check whether the ambiguity is about category, amount, or missing
   context.
3. Draft a short note describing exactly what's ambiguous and which
   clauses were considered.
4. Route the line to the exception queue with that note attached;
   never auto-approve it.
5. Track recurring ambiguity types over time so gaps in the written
   policy surface.

## Rules
- Never resolve an ambiguous line by inferring intent — escalate
  every time, no exceptions.
- Never treat silence in the policy as implicit approval.
- What counts as "clearly covered" versus ambiguous (set on first
  run, then stored in memory).
- Log every escalation reason so recurring policy gaps become
  visible to the controller.

## Output
An escalation entry per line: `{line_id, ambiguity_reason,
clauses_considered}`, added to the exception queue.

## Receipt-to-claim match

Confirm the receipt in Drive actually supports the claimed vendor,
amount, and date before any approval.

## Method
1. Locate the receipt file in `{{drive_folder}}` linked to the
   expense line.
2. Extract vendor, amount, and date from the receipt image or PDF.
3. Compare each field exactly against the claimed value in
   `{{expense_sheet}}`, recomputing the comparison independently
   rather than trusting a prior match flag.
4. Treat any mismatch, however small, as a failed match.
5. If no receipt is found, treat the line as unmatched, not as a
   pass.

## Rules
- Never approve a line where the receipt amount, vendor, or date
  differs from the claim, even slightly.
- A missing or unreadable receipt always fails the match — never
  assume it would have matched.
- Acceptable rounding tolerance for amount matching, if any (set on
  first run, then stored in memory).
- Never edit the claimed amount to match the receipt — flag the
  discrepancy instead.

## Output
A match result per line: `{line_id, receipt_found, matched,
mismatch_fields}`.

## Recurring jobs

This skill pairs with scheduled jobs. Offer to set each one up; only create the ones the user approves:

- Expense report sweep (Every 30 minutes on weekdays) — https://eluu.ai/cookbook/expense-policy-audit/jobs-setup/expense-report-sweep.md