When a model picks the model: routing is a control in regulated AI
On 17 November, at Ignite 2026, Microsoft made a model called Decision 1.1 generally available. Microsoft describes it as "a fast, low-cost decision model" for routing, classification, prioritisation, verification and workflow control. Its job is to look at a request and decide whether a larger, more expensive model is needed at all.
The same announcement introduced MAI Licensed Weights: customers can take one of Microsoft's frontier MAI models, optionally adapt it with their own data, and run it in their own Azure environment, with control over who can access it.
Neither is a surprise. Both point in the same direction: the model call is turning into an architecture with several decision points. In a regulated or tightly controlled environment, that changes what you need to design.
Routing is a control, not just an optimisation
Microsoft pitches routing on cost and speed, and its own figures for Decision 1.1 are bold (they are vendor-stated, so test them on your own workload). For most organisations, cost will be the reason routing gets introduced.
In healthcare, finance, the public sector or defence, the more important question is a different one: once a component decides which model handles a request, it also decides where the data in that request goes. A router choosing between a small model in your own environment and a large model behind someone else's API is making a data-flow decision, every time, at runtime.
That makes the router a control. And controls have requirements that optimisations do not.
Five design rules for routing in controlled environments
1. Decide what may leave by rule, before the router. Data classification should be deterministic and testable: this tenant, this data type, this source system never goes to an external endpoint. Let the router choose only among models that the request is already allowed to reach. A model that judges sensitivity can be wrong in ways you cannot predict or explain to an auditor.
2. Fail closed. If the router is unavailable, slow or uncertain, the safe default is the model inside your boundary, or a clear refusal. Never "send it to the big one".
3. Log every routing decision. Which request, which classification, which model, which version of the routing policy. When someone asks months later why a record was sent to a particular service, "the model decided" is not an answer.
4. Test routing like code. Keep a set of known requests with known correct routes, and run it whenever the routing model, the policy or a downstream model changes. A router update is a change to your data flows and belongs in change management.
5. Keep the policy outside the model. The router can be a model; the rule about what is allowed should be configuration you can read, review and version. That separation is also what lets you replace the router without re-approving the whole solution.
Licensed weights move the data path, not the questions
Running a model in your own environment is the pattern many controlled environments have asked for: the data stays where your other data is, and inference happens inside a boundary you already govern.
It solves part of the problem. It does not settle three questions that still need an answer:
- Who operates the environment, and under which jurisdiction? Running the weights in your own subscription changes where inference happens. It does not change who the cloud provider is. For organisations that need a specific sovereignty level, that is still the deciding question.
- What happens to adaptations and logs? Fine-tuning data, evaluation sets, prompts and traces are data too. Decide where they live and how long they are kept.
- What is the exit path? Licensed weights tie you to a specific model family and runtime. Know what it would take to move the workload to another model or another environment, and write it down before you need it.
What we would do now
If you are piloting AI in a regulated environment, treat model choice as a data-flow decision from day one:
- Draw the data flows per model, not per application.
- Put a deterministic classification step in front of anything that can reach an external model.
- Make routing decisions observable and auditable.
- Add a regression set for routing to your pipeline.
- Ask the same sovereignty questions of a "model in your own environment" that you would ask of any other managed service.
Routing models will make AI cheaper. In a controlled environment, they only make it safer if the rules about what may go where are yours, written down, and enforced before the model gets a vote.
Sources
- Microsoft: Microsoft Ignite 2026: Empowering human ambition, intelligence and trust (17 November 2026)
- Microsoft AI: Microsoft Ignite 2026 – MAI models (17 November 2026; vendor claims)