Route AI tasks by tested requirements instead of model prestige
Define easy and hard task classes, evaluate candidates on the same examples and use a controlled escalation path.
In brief
Model routing selects a suitable model for a task rather than sending every request to the largest option. It only helps when the smaller or faster candidate still meets the task’s requirements.
Simple formatting and evidence-heavy reasoning can deserve different routes, but classify them using tested outcomes.
Separate requirements before models
A request to reshape verified text differs from a request requiring multi-source judgment. List format, factual and tool-use requirements first. Do not label a task “easy” just because its prompt is short; a one-line question can require difficult evidence.
Design escalation deliberately
Escalate on a defined validation failure or missing capability, not every uncertain-looking word. Preserve the source and record why escalation occurred. For consequential actions, enforce permission checks regardless of which model was chosen.
Evaluate the whole route
Track error, retry and escalation rates alongside latency and billing. Review samples from both branches so good high-tier answers do not conceal poor default answers. Update rules after meaningful model or task changes rather than assuming one evaluation remains valid indefinitely.
Before you act
A smaller model is not automatically cheaper per successful task if it creates more errors, retries or escalations. Routing can also make behaviour harder to trace.
What to do next
- Define task classes and success criteria using real permitted examples.
- Compare candidate models with the same inputs, evidence and validation.
- Set a documented routing and escalation rule, then monitor quality and cost per completed task.
