An AI orchestration diagram often presents model routing as a performance optimization: send simple tasks to a smaller model and difficult ones to a more capable model. The diagram seems technical and neutral. Yet the distinction between easy and difficult is a business judgment. A short request to alter a customer's billing account may be linguistically simple and operationally high risk. A long document summarization may be technically demanding but reversible.
A responsible router begins with the task's consequences rather than its apparent complexity. What system will act on the result? Can a mistake be detected and reversed? Is sensitive data involved? Does a regulatory or contractual requirement constrain the provider? Only then should the system consider latency, cost, model capability and available context. The routing decision should be explainable to operators and testable across representative failures.
The economics need a denominator. Low cost per token is not low cost per accepted business outcome if extra calls, retries and human correction erase the savings. The simplest useful architecture may combine deterministic rules, small specialized models and expensive reasoning only where justified. But every fallback introduces a new contract: how to recover when a model provider times out, returns a refusal, changes output format or has different retention terms.
Treat routing policy like versioned product behavior. Define evaluation sets around real tasks, measure acceptance and harm, preserve the policy version with consequential decisions and keep human override available. A routing layer becomes a competitive advantage when it makes intelligence economical without making responsibility invisible. The product owns the result regardless of which model produced it.