Microsoft AI Foundry

Tietotalo builds AI agents and apps on Microsoft Foundry, formerly Azure AI Foundry, with retrieval, evaluation and guardrails inside your own Azure tenant.

Picking a model is the easy part. Everything around the model is where AI projects succeed or stall.

Most organisations that try generative AI get a promising demo within a week, then spend months trying to turn it into something staff can trust. The model answers well in testing, then gives a confident wrong answer about a contract clause, or surfaces a document to someone who should never have seen it. The problem is rarely the model. It is the retrieval, permissions, evaluation and monitoring that nobody planned for.

Microsoft Foundry, known until late 2025 as Azure AI Foundry, is Microsoft's platform for building that missing layer. It brings models, agents, search indexes, evaluations and safety controls together in one place, inside the Azure subscription where your data, identities and security policies already live. We use it when a client has outgrown what Copilot Studio can do, or needs an AI feature that runs inside its own software rather than inside Microsoft 365.

How we help

  • Use case and model selection: We start from the task, not the model. Foundry gives access to models from OpenAI, Anthropic, Meta, Mistral and Microsoft, and we test candidates against your own examples before recommending one.
  • Retrieval over your own data: Grounding answers in SharePoint, Blob Storage, SQL or document libraries through Azure AI Search and Foundry IQ, with permissions carried through so people only get answers from content they are allowed to see.
  • Agents with real tools: Building agents in Foundry Agent Service that call your APIs, look up records and complete multi-step tasks, with clear limits on what each agent is allowed to do.
  • Evaluation before launch: Test sets built from real questions, scored for accuracy, groundedness and safety, so you know how the system performs before staff start relying on it.
  • Guardrails and governance: Content safety filters, logging, cost limits and role-based access configured from day one rather than retrofitted after the first incident.
  • Integration into your applications: Exposing Foundry agents and models through APIs to your web apps, booking systems, Teams or Copilot, so AI becomes part of the workflow instead of yet another separate tool.

Technology

  • Microsoft Foundry (formerly Azure AI Foundry and Azure AI Studio)
  • Foundry Agent Service and Microsoft Agent Framework
  • Foundry Tools (formerly Azure AI Services): Document Intelligence, Language, Speech, Translator, Vision
  • Azure AI Search and Foundry IQ
  • OpenAI GPT models and Anthropic Claude models in the Foundry model catalogue
  • Foundry evaluations and content safety
  • Azure Monitor and Application Insights
  • Microsoft Entra ID, Azure Key Vault and private endpoints
  • .NET, Python and TypeScript SDKs
  • Copilot Studio, Microsoft Teams and Microsoft 365 Copilot as delivery channels

Why it matters

Our own development runs on Azure, and we have built Copilot Studio agents for clients such as Skýra Retreat and Arctic SnowHotel, including an agent that works over a SharePoint based partner CRM. Those projects showed us where low-code agents stop being the right tool: when you need precise control over retrieval, your own evaluation process, or an agent embedded in custom software.

Foundry is the next step up on the same Microsoft stack. Your identities, security policies and Microsoft agreements carry over, instead of starting again on a new platform with a new set of contracts to review.

The most common failure in retrieval-based AI is not hallucination in the usual sense. It is the system finding the wrong document, an outdated price list or a superseded policy, and then answering faithfully from it. The model did its job; the index did not. That is why we spend more time on content hygiene, chunking and metadata (which version is current, who owns it, when it expires) than on prompt wording. A well-kept index with an average model beats an excellent model searching a messy SharePoint.

Frequently asked questions

Yes. Microsoft renamed Azure AI Foundry to Microsoft Foundry at Ignite in November 2025, and before that it was called Azure AI Studio. Existing projects carry over. What changed, besides the name, is a stronger focus on agents and on connecting AI to company knowledge.
For many internal assistants Copilot Studio is the faster and cheaper choice, and we will tell you if that is the case. Foundry makes sense when you need custom retrieval, your own evaluation, a specific model, or an AI feature inside your own application or a customer-facing service.
Foundry resources are deployed in an Azure region you choose, and many models can be deployed with processing kept within the EU. Not every model is available in every region, so we confirm availability and data processing terms for each model during planning.
Yes. The Foundry catalogue includes models from Anthropic, Meta, Mistral, Cohere and Microsoft alongside OpenAI. We choose per task, and often use a smaller, cheaper model for routine steps and a stronger one only where reasoning quality matters.
Foundry is pay as you go: you pay for model usage, search and the Azure resources behind them. We estimate running costs from your expected volumes during the proof of concept, so there are no surprises when usage grows.