13.2 Foundry Model Catalog for Custom Prompts
Key Takeaways
- Custom prompts in Copilot Studio can use managed platform models or bring-your-own models from the Azure AI Foundry model catalog for frontier and specialized capabilities.
- Connect a Foundry model from the prompt builder Model dropdown using exact Model deployment name and Base model name from Foundry; wrong names break the connection.
- Bring-your-own Foundry models for prompts require a chat completions endpoint ending in /chat/completions; Responses API endpoints fail with resource-not-found style errors.
- Choose models by capability, latency, cost/licensing rate, multimodal needs, and release stage—not by picking the largest name on the list.
- Governance for Foundry models in prompts uses the Azure AI Foundry connector and Power Platform data policies; apply Responsible AI controls in Foundry as well.
13.2 Foundry Model Catalog for Custom Prompts
Quick Answer: In Copilot Studio prompt tools, open Model, use + to connect a deployment from the Azure AI Foundry / Model catalog, enter deployment name and base model name exactly, and point the endpoint at chat completions. Pick models for task fit, latency, cost, and multimodal needs—not brand hype.
AB-620 measures configure custom prompts to use the Foundry model catalog. Earlier topics covered custom prompts as advanced response tools; this leaf is specifically which model powers the prompt and how Azure model estate attaches to Copilot Studio.
Why prompts need a model choice
A prompt (prompt tool / prompt action) is a reusable generative unit: instructions, optional knowledge or inputs, and a model that executes the completion. Prompts appear as agent tools or inside topics. The model choice changes:
- Reasoning depth and instruction following
- Context window behavior
- Multimodal input (images/documents) support
- Latency and token/credit cost
- Content moderation controls available in the prompt UI
- Data residency and external hosting terms (for some catalog models)
Microsoft’s Azure AI Foundry models for prompts integration brings frontier and open models into Copilot Studio, Power Apps, and Power Automate so makers are not limited to a short fixed list of managed defaults.
Managed models vs Foundry (bring-your-own) models
| Aspect | Managed models in prompt builder | Foundry catalog / BYO models |
|---|---|---|
| How you get them | Appear in the Model dropdown by default | Connect via + to Azure AI Foundry deployment |
| Examples (illustrative) | Platform GPT mini/general/deep managed options | Deployments of Llama, DeepSeek, Phi, GPT variants, 1,800+ catalog entries as offered |
| Setup burden | Low | Requires Foundry project, deployment, exact names, correct endpoint |
| Content moderation slider in prompt settings | Available for managed models | Microsoft documents the moderation slider as unavailable for Azure AI Foundry (and some external) models—manage RAI in Foundry |
| Best when | Standard enterprise prompts, least ops overhead | Need a specific catalog model, specialty capability, or org-standard deployment |
Exam trap: “Bring your own model” does not mean paste an arbitrary public API key into a message node. It means the supported Foundry connection experience on the prompt’s Model field.
Configuration walkthrough (maker path)
- Create or open an agent in Copilot Studio.
- Add a Prompt tool (Tools → Add a tool → New tool → Prompt, or insert a prompt from Topics).
- Name the prompt and author Instructions (or use Copilot-assisted prompt drafting).
- On the panel beside Instructions, open the Model dropdown.
- Select the plus (+) control to Connect a model from Azure AI Foundry.
- Enter Foundry details: Model deployment name and Base model name exactly as shown in Azure AI Foundry.
- Complete Connect. The model appears in the dropdown for that prompt.
- Select the model so every run of the prompt uses that deployment.
- Test the prompt with representative inputs; verify latency and output quality before publishing.
Endpoint rules that fail real projects
Microsoft documents a hard requirement for BYO Foundry models in prompts:
- Supported type: chat completion models.
- Azure model endpoint URL must target the chat completions route ending in
/chat/completions. - Example shape:
https://<your-resource>.services.ai.azure.com/openai/deployments/<deployment-name>/chat/completions. - Using a Responses API path ending in
/openai/v1/responsestypically fails with a Resource not found style error—fix the URL rather than regenerating the whole prompt.
Also note documented product limits such as GPT-5 family and later models not being supported for bring-your-own in prompts at the time of Microsoft’s article—exam scenarios may describe “connection fails for newest family” and want the supported chat completion deployment answer, not “delete Copilot Studio.”
Multimodal inputs and model filtering
Prompts can accept image and document inputs. When you add image input, the Model dropdown filters to Foundry/models that actually support images so makers cannot pick a text-only deployment by accident. Microsoft lists examples that work with images (for example Phi vision variants and several GPT-4o-class models). Separately, Copilot Studio does not natively expose text-to-image generation as a first-class catalog pick in the same UI—image generation often needs a custom action/REST call to a deployed generative model.
Latency, cost, and capability tradeoffs
Think in categories, not marketing names:
| Category | Typical strengths | Typical costs / latency | Prefer for |
|---|---|---|---|
| Mini | Fast, cheaper, good enough for classification/summarization | Lower latency and credit impact | High-volume routing, short transforms, simple extraction |
| General | Strong multimodal and complex instruction following | Medium | Content drafting, mixed document work, default business prompts |
| Deep / reasoning | Multi-step reasoning, harder analysis | Higher latency and premium rates | Planning, dense analytics narratives, careful logic |
Additional tradeoff dimensions:
- Release stage: Experimental/preview models may be unstable—avoid production SLAs.
- Region availability: A model in the catalog is useless if not deployable in your region.
- External hosting terms: Some catalog models (for example certain third-party hosts) follow non-Microsoft data handling—legal and compliance must approve.
- Temperature & settings: Lower temperature for deterministic extraction; higher for creative drafting. Some reasoning models disable temperature controls.
- Credits: Prompts in Copilot Studio consume Copilot Credits (Power Apps/Automate may use AI Builder credits first). Larger or premium models burn budget faster—capacity planning is an architecture skill, not only a UX preference.
Scenario — Contoso claims summarizer: Nightly batch summarizes adjuster notes. A mini managed model meets quality with sub-second completions and lower credit spend. When the same prompt is reused for contested multi-document legal narratives, makers switch that prompt’s Model to a deeper Foundry deployment. One agent, two prompts, intentional tradeoffs—exam gold.
Scenario — Vision intake: Field photos of damaged equipment must be described. Adding image input hides text-only models; the maker connects a vision-capable Foundry deployment. Choosing a non-vision model is not a “prompt engineering” problem—it is a catalog capability problem.
Governance and Responsible AI
Foundry models surface through connectors. In the Power Platform admin center, the connector appears under names such as Azure AI Foundry, where admins can apply data policies (DLP) to allow or block. Each connected model can also appear on environment Connections pages for lifecycle management.
Microsoft recommends applying Responsible AI policies in Azure AI Foundry for models you expose to makers. Remember: prompt-builder content moderation level sliders are documented for managed models; for Foundry-connected models, rely on Foundry-side controls and organizational RAI process rather than assuming the same slider still works.
Relationship to other AB-620 skills
- Custom prompts (advanced responses) teach what the prompt does.
- This leaf teaches which Foundry catalog model executes it.
- Foundry agent integration (multi-agent chapter) is about calling agents, not merely selecting a completion model for a prompt.
- Generative answers + Search (13.1) grounds free-form Q&A; prompts are structured generative tools with explicit instructions and I/O.
Pitfalls
- Typos in deployment or base model names.
- Wrong endpoint family (Responses vs chat completions).
- Picking experimental models for SLA-bound customer channels.
- Assuming content moderation UI works identically for Foundry models.
- Using a text-only model after enabling image inputs.
- Ignoring DLP—makers “lose” Foundry models after admin blocks the connector.
- Confusing BYO prompt models with building a full Foundry multi-agent solution.
Success checklist
- Connect Foundry models from the prompt Model dropdown with exact names.
- Require chat completions endpoints for BYO prompts.
- Compare managed vs catalog models on ops burden and RAI controls.
- Select mini / general / deep (or equivalents) using latency–cost–capability reasoning.
- Account for multimodal filtering and governance via the Azure AI Foundry connector.
- Test prompts under realistic load before publishing.
When you can explain why Contoso’s extraction prompt stays on a mini model while its dispute narrative prompt uses a Foundry deep deployment—with correct connection mechanics—you own this blueprint skill.
A maker connects an Azure AI Foundry model to a Copilot Studio prompt but receives a Resource not found error. The deployment exists in Foundry. Which fix aligns with Microsoft’s bring-your-own model guidance for prompts?
Which statement best describes managed prompt models versus models connected from the Azure AI Foundry catalog?
Contoso needs a high-volume prompt that classifies support emails into five categories with low latency and controlled credit use. Which selection strategy best matches AB-620 tradeoff thinking?