The enterprise question for two years was which model to standardise on, and it was always a slightly odd question. Nobody asks which database to standardise on across every workload, because the answer obviously depends on the workload.
What is replacing it is routing: a layer that receives a request, decides what kind of request it is, and sends it to whichever model handles that kind well and cheaply. Classification and extraction go to something small and fast. Long-context reasoning goes somewhere expensive. Anything touching regulated data goes to the deployment that satisfies the residency requirement.
The saving is real and the layer is not free
The economics are straightforward and substantial. Most enterprise volume is boring — summarise this, classify that, pull three fields out of a document — and paying frontier prices for boring work is the largest avoidable line in most AI budgets. Routing addresses precisely the problem that emerged when inference costs moved from footnote to line item.
It also loosens a dependency. An organisation that can route can substitute, which is the only practical answer to everybody diversifying their model suppliers and nobody diversifying the layer underneath — the diversification is meaningless unless something can actually switch.
What that system costs is where deployments get optimistic. The router is a piece of software on the critical path of everything, with its own latency, its own failure modes and its own need for evaluation — because a routing rule that sends the wrong class of request to a small model produces quiet, plausible, cheap wrong answers rather than an error anyone notices.
Evaluation is therefore harder rather than easier. Instead of assessing one model against a task, an organisation is assessing a routing policy against a distribution of tasks, and most cannot yet do the simpler version. This is the evaluation gap with an extra dimension added.
There is also a version problem that compounds. Each model underneath has its own deprecation schedule, so a router is a component whose dependencies expire independently and without coordination — which is model deprecation as an enterprise risk multiplied by the number of providers in the pool.
The organisations doing this well treat the router as part of the platform rather than as a clever script inside one application — which is the same instinct that produced the internal developer platform, and carries the same warning about who maintains it in three years.
