Dynamic model routing under a cost ceiling

Model selection is the highest-leverage decision in an agentic stack, and in most enterprises it is not a decision at all — it is a default. This piece specifies routing as a constrained optimisation across stakes, data class, jurisdiction, latency and an enforced budget envelope.

In an agentic stack, the single decision that most determines cost, compliance posture and user-perceived quality is which model executes a given task. In most enterprises that decision is made once, at integration time, by a developer choosing a default — and then inherited by every task the workflow will ever run, regardless of stakes.

Treating it as a live decision changes the economics and the risk profile at the same time, because the constraints that should govern model choice are exactly the constraints that govern exposure: what the task is worth getting wrong, what class of data it carries, which jurisdictions may process it, how fast the answer is needed, and how much authority remains in the budget envelope.

The routing decision, stated

1. Stakes — the cost of an incorrect action, expressed by directive class rather than inferred from prompt length.
2. Data class — determined before routing, restricting eligible models, regions and execution modes.
3. Jurisdiction — which legal regime may process the payload, and therefore which providers are eligible at all.
4. Latency budget — an interactive directive and an overnight batch do not warrant the same tier.
5. Spend envelope — the remaining authority attached to the role and the period, enforced rather than reported.

Given those inputs, routing selects the cheapest tier that satisfies every constraint. Where no tier satisfies them, the correct behaviour is refusal with the failing constraint named — not silent selection of the closest available option, which converts a policy failure into an invisible exposure.

Why the ceiling belongs in the same step

Cost control implemented downstream of routing can only report or interrupt. Implemented inside routing, the envelope becomes another constraint in the same evaluation, which means the system can degrade along a path the accountable role has already approved. That is the difference between a cost control that saves money and one that quietly damages outcomes to hit a number nobody agreed to.

Measuring the router

Escalation rate is the honest test. A router that never escalates is over-provisioning; one that escalates constantly is under-tiering and paying twice. The target is a stable, low escalation rate at a falling cost per completed outcome — and that target only becomes measurable once every action carries its directive class, its tier and its realised cost in the same record.

“Choosing the strongest model for every task is not a quality strategy. It is the absence of one.”