Model policy
Clients ask two questions about models, usually in this order: can our people use whichever model they want, and can we stop them. Both have answers, and the second one is a differentiator worth demonstrating rather than describing.
What you can control
Smart routing
The default. Each prompt is evaluated and routed to the best model for the task, balancing quality, speed, and credit cost. Nobody has to choose a model.
Manual pinning
Lock a specific model to a conversation or project when the work demands a known, fixed behaviour rather than the best available answer.
Department defaults
On Enterprise, all projects in a given department can default to a particular model family. Applied at admin level, automatically.
Credit budgets
Limit frontier model usage to approved users, so the most expensive models stay available for the work that justifies them.
Model routing policies are managed at the admin level and applied automatically, so employees do not have to understand or comply with the policy for it to hold. Custom routing policies are an Enterprise capability, arranged with Skynet as part of the deployment. See Enterprise: on your infrastructure.
Start from smart routing
The instinct on a locked-down deployment is to pin everything immediately. Resist it for the first wave.
Smart routing is the cost control as much as it is the quality control. Most everyday prompts do not need a frontier reasoning model, and routing them to a fast model is what keeps the bill down. Pinning a single expensive model organization-wide usually costs more than leaving routing on, and delivers worse answers on the tasks that model is poorly suited to.
Pin deliberately, where there is a reason: a regulated workflow that requires consistent behaviour, or a department whose work genuinely favours one model family.
Context survives the model
Every model draws on the same collective memory. Switching model mid-conversation preserves full context, because the workspace owns the context rather than the model provider.
This is the point to make when a client worries about committing to a policy: changing the model later does not cost them their accumulated context. Model choice is reversible in a way that most of their other AI decisions are not.
Next: data handling.