Microsoft Foundry: Your AI Platform
Understand Microsoft Foundry — the unified AI platform for building, deploying, and managing custom AI applications with enterprise security and responsible AI built in.
The platform for custom AI
If Copilot for M365 is a ready-made meal, Microsoft Foundry is a professional kitchen where you cook whatever you want.
Foundry gives you access to hundreds of AI models — from OpenAI’s GPT to open-source models like Llama and Phi. You pick the model, connect your data, build your application, and deploy it — all within Microsoft’s secure cloud.
Why not just use Copilot? Because some problems need custom solutions. A factory that needs AI to inspect steel quality can’t use Copilot for that. They need a custom vision model — and Foundry is where they build it.
What Foundry provides
| Capability | What It Does | Business Value |
|---|---|---|
| Model catalogue | Access 1,800+ AI models from multiple providers | Choose the best model for your task — no vendor lock-in |
| Playground | Test models with different prompts side-by-side | Compare quality and cost before investing in development |
| Prompt flow | Design multi-step AI workflows visually | Orchestrate complex AI applications without deep coding |
| RAG pattern | Connect models to your own data sources | AI answers grounded in YOUR data, not just general knowledge |
| Fine-tuning | Customise models with your own training data | Specialise a model for your industry or domain |
| Deployment | Host models as scalable API endpoints | Production-ready AI with autoscaling and monitoring |
| Evaluation | Test model quality, safety, and groundedness | Ensure outputs meet your standards before going live |
| Content safety | Built-in filters for harmful content | Protect your brand and users from inappropriate AI outputs |
When to use Foundry vs Copilot
| Scenario | Use Copilot | Use Foundry |
|---|---|---|
| Help employees write emails and documents | Yes — Copilot in M365 apps | No — over-engineering |
| Build a customer-facing AI chatbot with custom personality | No — Copilot is employee-facing | Yes — full control over UX and model |
| Analyse spreadsheets for trends | Yes — Copilot in Excel or Analyst | No — Copilot handles this |
| Create an AI that classifies support tickets by urgency | No — requires custom model | Yes — train on your ticket history |
| Summarise meeting recordings | Yes — Copilot in Teams | No — built-in capability |
| Build an AI-powered recommendation engine for your product | No — outside Copilot’s scope | Yes — custom model + deployment |
Exam tip: Foundry is for 'Build' scenarios
In the Build, Buy, or Extend framework from the previous module:
- Buy = Copilot for M365
- Extend = Copilot + Studio + Graph connectors
- Build = Microsoft Foundry (or Azure OpenAI Service)
If the exam describes a scenario where an organisation needs a custom AI model, a customer-facing AI application, or AI that processes specialised data (images, documents, industry-specific language), the answer is likely Foundry.
Enterprise benefits of Foundry
The exam specifically tests on Foundry’s business benefits — especially scalability and security.
Security and compliance
| Security Feature | What It Means |
|---|---|
| Azure infrastructure | Runs on Microsoft’s enterprise-grade cloud with certifications (ISO 27001, SOC 2, HIPAA, etc.) |
| Network isolation | Deploy models in your own virtual network — data never leaves your environment |
| Managed identity | No passwords or API keys stored — Azure handles authentication |
| Data encryption | Data encrypted at rest and in transit |
| Content safety | Built-in filters block harmful, biased, or inappropriate outputs |
| Audit logging | Full audit trail of model usage, prompts, and responses |
Scalability
| Scalability Feature | What It Means |
|---|---|
| Autoscaling | Model endpoints scale up and down based on demand |
| Global deployment | Deploy to Azure regions worldwide for low latency |
| Provisioned throughput | Reserve capacity for predictable, high-volume workloads |
| Pay-as-you-go | Start small and scale costs with usage |
| Multiple model sizes | Use smaller, cheaper models for simple tasks; larger models for complex ones |
Model choice and flexibility
| Benefit | Why It Matters |
|---|---|
| 1,800+ models | Choose the best model for your specific task |
| No vendor lock-in | Switch models without rebuilding your application |
| Open-source options | Use Llama, Phi, Mistral alongside commercial models |
| Comparison tools | Test multiple models on the same data before choosing |
| Fine-tuning | Customise any supported model with your own data |
🏗️ Ravi builds a proof of concept in Foundry
Ravi, CTO of TechVantage Solutions, needs a custom AI application that analyses code pull requests and suggests quality improvements. No off-the-shelf product does this for their specific technology stack.
Step 1: Model selection Ravi’s team opens the Foundry model catalogue. They test GPT-4o and Llama 3.1 on sample code reviews. GPT-4o produces better suggestions, but costs 5x more. They settle on GPT-4o for complex reviews and Llama for routine checks.
Step 2: Data integration They connect Foundry to TechVantage’s code repository and historical review data using RAG. The model now understands their coding standards and common patterns.
Step 3: Prompt engineering The team designs prompts that produce structured code review feedback — severity rating, suggested fix, and explanation. They iterate in the playground until quality is consistent.
Step 4: Deployment The model is deployed as an API endpoint within TechVantage’s Azure virtual network. Network isolation ensures source code never leaves their environment.
Step 5: Monitoring Foundry’s monitoring tracks response quality, latency, and cost. Ravi sets up alerts for anomalies.
Why Ravi didn't use Copilot for this
This is a clear “Build” scenario because:
- The use case is highly specialised (code quality analysis for a specific tech stack)
- It requires a custom data pipeline (connecting to the code repository)
- It needs model comparison and selection (GPT-4o vs Llama for different tasks)
- It’s a competitive differentiator for TechVantage
- It requires network isolation for source code security
Copilot for M365 doesn’t analyse code repositories. Copilot Studio builds chatbots, not code review systems. Foundry is the right platform.
Ravi is evaluating platforms for TechVantage. A healthcare division needs to build an AI application that analyses medical images to detect anomalies. The application must comply with HIPAA and keep all data within their Azure environment. Which platform should they use?
Ravi is presenting the benefits of Microsoft Foundry to his leadership team. They ask about handling traffic spikes during product launches. Which of the following is a key scalability benefit of Microsoft Foundry?
🎬 Video coming soon
Next up: Azure AI Services: Vision, Search and Beyond — explore the specialised AI services for vision, speech, language, and search.