Case Study · Captive Insurance · Anonymized
An AI accountant that auto-classifies 100,000+ bank statements a year - saving about $100K annually.
A PE-backed alternative risk management and captive insurance platform serving 4,000+ mid-market clients was burning accountant hours on a process AI was already better at: classifying every bank-statement transaction into the right QuickBooks GL code. We replaced the manual workflow with an AI accountant - and ran two models in parallel for accuracy.
The problem
Trust admin was buried in manual classification
The platform's trust administration team manually reviewed and classified thousands of bank statements per year covering custodial accounts holding mutual funds, ETFs, bonds, and mortgage-backed securities.
Every transaction had to land in the right QuickBooks general ledger code. The work was tedious, error-prone, and a bottleneck on accounting close. Headcount didn't scale with client growth - but the statements did.
What we did
Replaced manual classification with a dual-model AI workflow
Step 1
Ingestion
Bank statements arrive via email or Drive in PDF form. The pipeline parses every statement automatically - no human handling, no upload step.
Step 2
Dual-model classification
Every transaction runs through Claude Opus and Gemini 2.5 Pro in parallel, each independently classifying against 18 QuickBooks GL codes. The two outputs are cross-checked so the team can see exactly where they agree - and catch the few they don't before anything touches the books.
Step 3
Validation & structured output
Rollup totals reconcile against transaction sums. Only valid QB account codes pass. Deterministic hashing prevents duplicates and makes re-runs idempotent. Output lands in Google Sheets in import-ready format.
Why the team trusts the output
In a regulated context, "AI did it" isn't a finishing line. We built the validation layer that makes the output ledger-ready before a human ever looks at it.
- ✓ Cross-model agreement check (Claude vs. Gemini per transaction)
- ✓ Total reconciliation (AI totals vs. raw transaction sums)
- ✓ GL whitelisting - only valid QuickBooks account codes accepted
- ✓ Duplicate detection across runs via deterministic hashing
- ✓ Date and amount enforcement on every row
The result
A trust admin team that doesn't touch bank statements anymore
Custodial accounts now reconcile automatically. Output is structured, validated, and ready to import into the accounting system without rework. The team's time goes to higher-leverage work; the statements take care of themselves.
The system is designed to handle 100,000+ statements a year - so as the client roster grows, the workflow scales without proportional headcount.
What's your back office still doing by hand?
If a process is repetitive, document-driven, and slows your close - there's almost always an AI version that pays for itself in a quarter.
Book a free AI sessionClient identity withheld by request. The platform is a PE-backed alternative risk management and captive insurance provider serving 4,000+ mid-market clients across property & casualty, employee benefits, and captive management. Throughput, GL codes, and dollar-savings figures are based on the production deployment, 2025–2026.