Stop Duplicate & Fraudulent Expense Claims with ChatGPT

Build a Policy-Aware Review Copilot Prompt

The core of a successful ChatGPT expense review copilot is a carefully designed system prompt that embeds your policies and desired outputs. Instead of asking vague questions like “Is this claim OK?”, give the model a clear role, step-by-step instructions and structured outputs. This makes results more consistent and easier to integrate into your approval process.

Example system prompt for ChatGPT:
You are an AI assistant supporting the finance team in reviewing employee expenses.

Goals:
- Detect potential duplicate claims.
- Identify potential fraud indicators.
- Check claims against the company's travel & expense policy.

Instructions:
1) Carefully read the expense details, receipt text, and card transaction data.
2) Compare this claim against the list of past claims provided.
3) Identify any of the following:
   - Possible duplicates (same vendor, amount, date range, or very similar description).
   - Policy violations (over limits, missing justification, non-reimbursable items).
   - Suspicious patterns (rounded amounts, unusual vendors, mismatched locations/dates).
4) Return a JSON object with:
   - risk_level: low / medium / high
   - reasons: list of specific findings
   - duplicate_candidates: list with claim IDs and similarity explanations
   - suggested_questions: questions for the employee or approver
   - recommendation: approve / approve_with_comment / escalate

Expected outcome: a standardized assessment that reviewers can scan in seconds, while audit and analytics teams can use the structured fields for reporting and trend analysis.

Automate Duplicate Detection Across Historical Claims

To catch reused receipts and split bills, you need to compare each new claim against a broad history of expenses. Tactically, combine your existing expense or ERP system with ChatGPT-based similarity checks. Pre-select candidate duplicates using simple rules (same amount + vendor within a date range), then let ChatGPT analyze the receipt text and narratives to confirm or discard the match.

Example prompt for duplicate verification:
You receive:
- current_claim: {id, employee, date, amount, currency, vendor, description, receipt_text}
- candidate_claims: [{id, employee, date, amount, currency, vendor, description, receipt_text}, ...]

Task:
For each candidate, assess how likely it is to be a duplicate of current_claim.
Consider:
- Same or very similar vendor names (including spelling variants).
- Same or very similar amounts (after currency conversion).
- Overlapping dates (e.g., same hotel nights).
- Very similar descriptions or receipt text.

Return JSON:
[
  {
    "candidate_id": <id>,
    "duplicate_likelihood": 0-100,
    "evidence": ["Same hotel name, same nights", "Receipt number identical"]
  }
]

Expected outcome: far fewer “false duplicates” reaching reviewers and a high catch rate for genuine double claims, even when descriptions or PDFs are slightly modified.

Use ChatGPT to Generate Reviewer-Ready Risk Summaries

Raw model outputs are rarely ready for busy finance teams. A tactical best practice is to let ChatGPT convert technical checks into concise risk summaries for expense approvers. These can be embedded directly into your expense tool or emailed as part of the approval flow, so managers see clear, human-readable reasons to approve or question a claim.

Example prompt for approver summary:
You are preparing a short summary for the line manager who must approve this expense.

Input:
- employee_role and department
- expense details (category, amount, date, vendor, description)
- AI risk assessment (risk_level, reasons, duplicate_candidates, suggested_questions)
- Relevant policy excerpts

Task:
Write a 3-5 sentence summary covering:
- Overall risk level in plain language.
- The 2-3 most important reasons for this risk level.
- 1-3 specific questions the manager should ask before approving.
Use neutral, professional language.

Expected outcome: managers spend less time understanding “what’s going on” and more time exercising judgment, leading to faster, better-informed approvals.

Flag Suspicious Narratives and Inconsistent Stories

Many fraudulent or non-compliant claims hide in the description field: vague wording, mismatched purpose, or narratives that don’t fit the employee’s role. ChatGPT is strong at natural language inconsistency detection. Use it to cross-check narratives against roles, travel calendars, meeting data (where permissible), and policy wording.

Example prompt for narrative checks:
You get:
- employee_role: "Senior Sales Manager, DACH"
- expense_description: free-text from the claim
- receipt_text: OCR from the receipt
- policy_summary: key rules for this expense category

Question:
1) Is the stated purpose plausible for this role and category?
2) Does the description align with the receipt details (location, date, participants)?
3) Are there any phrases that commonly indicate personal rather than business use?

Return:
- narrative_risk: low/medium/high
- explanation: bullet points
- suggested_follow_up_questions.

Expected outcome: more subtle cases – such as personal meals claimed as client entertainment – are flagged systematically without adding manual review work on every single claim.

Generate Clear, Audit-Ready Justifications and Logs

Every AI-supported decision in finance should be explainable. A practical step is to use ChatGPT to generate standardized explanations for approvals, partial approvals or rejections, based on the detected issues and the policy references. Store these explanations with the claim to build an audit trail that is easy to review months or years later.

Example prompt for justification generation:
Input:
- claim_id and basic claim details
- decision: approve / approve_with_comment / reject
- AI findings: list of issues and policy references
- human notes: optional comments from reviewer

Task:
Create a short justification (100-200 words) suitable for internal audit, covering:
- Key facts of the claim.
- Main reasons for the decision.
- Policy clauses that are relevant.
- Any mitigating factors.
Use clear, non-accusatory language.

Expected outcome: consistent decision rationales, less time spent writing explanations, and a much stronger position in internal or external audits.

Monitor Performance and Continuously Tune Prompts

Once your ChatGPT-based expense control is live, treat it as a system that needs ongoing tuning. Tactically, track KPIs such as percentage of claims flagged, confirmed duplicate/fraud rate, false positive rate, average review time, and employee dispute rate. Regularly review a sample of AI decisions with finance and internal audit to identify patterns.

Use those insights to refine prompts, thresholds and training examples. If certain vendors or categories generate too many false positives, adjust rules or add examples of what “good” looks like. If investigators find new fraud patterns (e.g., common fake merchant names), update the prompts so ChatGPT can look for them proactively. Over time, you’ll move from a static rule set to a living control system that improves with every case.

Expected outcomes: with a well-implemented setup, finance teams commonly see 30–60% reduction in manual review time per claim, significantly higher detection of duplicate and fraudulent expense claims, and a more predictable, audit-ready control environment – without adding headcount.