Model Routing Policy: Tiers, Budgets, and Degradation
Explore how model routing policies manage AI request classification by tiers budgets and fallback strategies. Learn to enforce routing rules that balance safety latency and cost constraints with fallback modes and auditability, preparing you to design compliant and efficient AI systems.
Deployment topology settles where each sensitivity tier is allowed to run. It says nothing about how that boundary actually gets enforced on any single request that comes in. Model routing is the policy layer that turns product intent into enforceable per-request decisions: which model may be called, with what options, under what budgets, and when the correct outcome is to refuse or degrade instead.
Routing sits downstream of the model selection matrix that identifies viable candidates, and downstream of the deployment topology that constrains where those candidates can run. It resolves the last-mile decision for the copilot’s three request classes: low-risk Q&A, restricted drafting with citations, and tool-eligible read-only lookups. One model for everything rarely meets minimum sufficient cost, latency, and sensitivity constraints all at once.
A routing decision consumes a small set of loggable, auditable inputs: task type, sensitivity tier, ...