Model routing

Purpose

Model routing policy defines where supervised AI traffic may go. This is useful when an organization has approved specific providers, blocked others, or limited certain models to specific teams.

What to control

Common routing controls include:

• Provider allow-lists.
• Model allow-lists.
• Workspace or team restrictions.
• Request-size limits.
• Destination host restrictions.

Example decisions

A policy can allow a request to an approved provider:

ALLOW provider=anthropic model=claude-sonnet team=engineering

It can block an unapproved destination:

BLOCK provider=unknown host=unapproved-llm.example

It can also require lower-risk models for specific directories or repositories.

Audit fields

Routing decisions should record the provider, model, destination host, agent process, user, device, workspace, and matched policy rule. These fields make the difference between a rough traffic log and a security audit trail.