Regulatory accuracy at
startup economics

The Problem

Fintech loans require precision. Regulatory bodies (OCC, Federal Reserve) audit AI decisions on mortgages and personal loans. A misclassified risk factor can trigger compliance violations. This company was using Claude Opus on every loan application — 600 documents daily — because they couldn't risk a mistake.

Opus accuracy on financial document analysis was 99.8%. But at $30/million tokens, processing 600 loan applications monthly cost $84K. The team needed a way to reduce costs without compromising the accuracy that regulators demanded.

The Insight

80% of loan applications are straightforward. W-2s, paystubs, bank statements — standard documents, standard income verification. Only 20% are complex: self-employed borrowers, recent immigrants with non-US income, co-signers with unusual financial structures.

Running Opus on routine applications was regulatory overkill. But the team couldn't manually triage — the complexity wasn't apparent until after analysis. They needed automatic classification with no risk of routing a complex case to the wrong model.

The Architecture

BrainstormRouter automated the distinction. The router analyzed document metadata — document types present, income source indicators, credit profile flags — and routed accordingly:

// Every loan application goes through BR
response = client.chat.completions.create(
    model="auto",
    messages=[
        {"role": "system", "content": LOAN_ANALYSIS_PROMPT},
        {"role": "user", "content": document_payload}
    ]
)

// Response headers show the decision trail
X-BR-Routed-Model: anthropic/claude-haiku-4-5
X-BR-Route-Reason: thompson-sampling
X-BR-Route-Confidence: 0.93
X-BR-Actual-Cost: $0.0048

// For complex applications:
X-BR-Routed-Model: anthropic/claude-opus-4
X-BR-Route-Reason: thompson-sampling
X-BR-Route-Confidence: 0.91
X-BR-Actual-Cost: $0.1240

Document Routing Distribution

The Compliance Trail

Financial regulators don't just want accurate results — they want to know how you got there. BrainstormRouter maintained a complete compliance log for every decision:

• Which model processed the application
• Why the router selected that model (Thompson Sampling confidence)
• The routing confidence score at time of decision
• Cost and latency of the request
• Whether any guardrails fired (PII detection, PCI patterns)

This audit trail satisfied the OCC's requirements for AI decision transparency. The compliance team could demonstrate that complex cases always routed to the premium model and explain the statistical basis for routing decisions.

Results

Costs dropped 63% — from $84K/month to $31K/month. Compliance accuracy stayed at 99.8% because Opus handled every complex case. Loan approval time fell 40% because Haiku processes documents 8x faster than Opus, and it now handled 70% of volume.

The annualized savings exceeded $600K. The compliance team reported that the audit trail actually improved their regulatory posture — they could now prove that AI routing decisions were statistically sound, not arbitrary.