AI-assisted forensic AP review for healthcare finance teams.
AuditCopilot ingests accounts payable transaction data, runs a forensic accounting rule engine to flag anomalies, and generates auditor-ready workpaper memos using Claude, citing AU-C 240 (fraud), AU-C 315 (risk assessment), and IAS 21 (FX) standards.
Built by Bhoomika Bothra. CA Finalist (ICAI, Group 1 cleared), MS Accounting and Analytics (Seattle University), CPA Candidate.
📂 Application source code: artifacts/audit-copilot/
In a previous Group Financial Accountant role at a multinational pharma distributor, I uncovered a $1.2M ghost vendor scheme through manual reconciliation, and a $180K FX misallocation under IAS 21. Both reviews took weeks of manual work.
AuditCopilot is the AI-assisted version of that workflow. On 199 synthetic healthcare AP transactions with planted anomalies, it:
- Reviews 100% of the population in seconds (vs. statistical sampling)
- Identifies $1.22M of risk-flagged value across 32 flagged transactions
- Generates 32 workpaper-ready audit memos citing the relevant standards
- Exports a 3-tab Excel workpaper ready for senior reviewer sign-off
| Rule | What it catches | Standard |
|---|---|---|
| Duplicate Vendor (Fuzzy Match) | Ghost vendors via name similarity (e.g. MedSupply Inc vs Med Supply Inc.) |
AU-C 240.A25 |
| Round-Dollar Payments | Unusual round amounts above materiality | AU-C 240.A32 |
| Weekend / Holiday Postings | Transactions outside normal business hours | AU-C 315 |
| Split-PO Patterns | Multiple invoices just under approval threshold | AU-C 240 |
| FX Misallocation | USD invoices posted to EUR-designated GL accounts | IAS 21 |
| Duplicate Invoice Numbers | Same invoice number across vendors | AU-C 240 |
| Benford's Law Analysis | Dataset-level first-digit distribution test | Forensic accounting |
- Python 3.11, Streamlit — web framework
- pandas, numpy, scipy, rapidfuzz — data + forensic rules
- plotly — interactive charts
- openpyxl — Excel workpaper export
- Anthropic Claude API (
claude-sonnet-4-5) — audit memo generation
- Upload an AP CSV (or use the built-in 199-row synthetic dataset)
- The forensic rule engine flags anomalies across 7 categories
- Risk Dashboard surfaces KPIs, flag counts, dollar exposure, and a sortable transaction table with red/amber/green risk badges
- Toggle "Generate AI Memos" → Claude drafts audit memos citing the relevant standard and recommending a specific testing procedure for each flag
- Export a 3-tab Excel workpaper (Summary, Flagged Transactions, AI Memos) ready for senior reviewer sign-off
git clone https://github.com/bhoomika122/auditcopilot.git
cd auditcopilot/artifacts/audit-copilot
pip install -r requirements.txt
export ANTHROPIC_API_KEY=sk-ant-... # optional, for AI memos
streamlit run main.pyApp runs at http://localhost:8501. Without an API key, the rule engine still works — only the AI memo step is disabled.
- Forensic accounting domain expertise — rules cite the standards a senior auditor would reach for in the field
- AI integration that adds judgment, not just text — memos recommend procedures, not generic disclaimers
- End-to-end deliverable thinking — UI, rule engine, AI layer, and downloadable workpaper, all in one application
- Healthcare AP context — McKesson, Cardinal Health, AmerisourceBergen, Henry Schein, Medline, Becton Dickinson, and Stryker as a realistic vendor universe