80% of Requests Never Hit the Expensive Model
May 7, 2026
I built an AI-native Mexican Fiscal Compliance-as-a-Service platform where 80% of requests never touch the expensive model, without degrading quality. How?
Every message in DISAI_Conta's chat interface passes through a three-tier router before hitting any fiscal logic.
Tier 1: classify
Haiku runs a classification pass: domain, query type, confidence score. Max 60 tokens. ~100ms. That's it. Ten fiscal domains: facturacion, nomina, carta porte, pagos, compliance, and more. One fast model call. Done.
Tier 2: the domain expert
A domain expert takes over. But not a generic prompt - a scoped module with its own allowed MCP tools, preloaded resources, and a system prompt tuned for that specific domain. Ten experts. Each one knows its lane.
Tier 3: full orchestration
Only action-type queries spin up the full agent: Sonnet doing real tool orchestration against the SAT-MCP server. Informational queries get answered at Tier 2 and never go further.
The result: a platform that costs a fraction of "Sonnet-for-everything," without degrading the experience on the queries that actually matter.
The tradeoffs, explicit and deliberate
- +100ms on action paths from the Tier 1 hop. Accepted.
- Added complexity from maintaining 10 expert modules. Accepted - even that is easier to maintain than a ~2k-line single orchestrator.
Because the alternative is paying Sonnet-level costs for someone asking "Cuales son las formas de pago aceptadas por el SAT?" and burning reasoning tokens on a query that has a known, scoped answer.
Most teams building AI products default to one model for everything. It's simpler. But it's also expensive and lazy.
So if your queries have structure (and fiscal/accounting queries absolutely do), you can route by intent, then by domain expertise, before you burn tokens on reasoning.
Prompts are architecture decisions. Routing is product design. Expertise is a layer, not an afterthought.
Originally published on LinkedIn.