How model choice works inside your tenant.
Microsoft-native where that is right, the best model where it is better, inside your Azure tenant under your identity and policies.
The stack
Seven parts. One tenant.
- Microsoft Foundry
- Azure OpenAI
- Anthropic Claude
- Open-source models
- Copilot Studio
- LangChain
- Semantic Kernel
-
Runs inside Microsoft Foundry
Entra ID, Azure RBAC, zero data retention. No new trust boundary.
-
Draws down your Azure commitment
Billed against your MACC, not a new vendor line.
-
One governed publishing pipeline
Built once in Foundry, surfaced in Copilot and Teams.
-
Model choice, not lock-in
Per-agent selection in Copilot Studio, models side by side in Foundry.
Per task
Chosen per task, on your data.
| Task | Choice | Why |
|---|---|---|
| Structured extraction over documents | Azure OpenAI or Claude | Highest score on your documents wins. |
| Long-context reasoning and code | Anthropic Claude in Foundry | Behind our code modernisation work. |
| High-volume classification | A small open-source model in Foundry | Cost per call over fluency, checked against the frontier. |
| Copilots inside Microsoft 365 | Copilot Studio | Native surface, native identity, one publishing pipeline. |
| Sovereign or restricted tenants | Microsoft-native path, in region | Availability and residency decide; nothing is redesigned. |
Where a model is unavailable in region, the Microsoft-native path on the same architecture; when it arrives, a configuration change, not a programme.
Let's connect
Bring us the problem AI keeps almost solving.
A 30-minute call with the people who'd do the work. We'll pressure-test the opportunity, show you what good looks like, and tell you plainly whether it's worth building.