Learned classification
Haiku handles 70% of volume (312 W-2s + 108 co-signer docs). Opus handles 20% complex (84 self-employed + 48 non-US + 48 other).
Fintech · regulated docs
Regulatory accuracy at startup economics.
Monthly savings
$53K
63% cut · $84K → $31K
Compliance accuracy
99.8%
maintained through rollout
Audit trail
100%
routing decisions logged
Problem
Paying Opus for routine W-2s.
600 loan applications per day, all routed to Opus at $84K/month to hit 99.8% accuracy for OCC compliance. 80% of documents — standard W-2s, paystubs, co-signer forms — didn't need premium reasoning.
Solution
Complexity-aware routing with an audit trail.
Thompson sampling learns document patterns. Routine applications route to Haiku. Complex ones (self-employed, non-US income) route to Opus. Router builds classification from document metadata — no manual triage.
Guardian Intelligence
Logs model, confidence, cost, and any guardrails fired (PII, PCI) for every decision.
Compliance trail
Audit trail proves complex cases always route to the premium model. Statistical basis for every decision satisfies OCC requirements.
“Annualized savings exceeded $600K. The compliance team reported that the audit trail actually improved regulatory posture — proving AI routing was statistically sound, not arbitrary.”