AI Agent Development with Microsoft Foundry
AI Agent Development with Microsoft Foundry designs, builds, evaluates, and deploys production pro-code AI agents on Microsoft Foundry — the platform Microsoft renamed from Azure AI Foundry. Scope covers retrieval-augmented generation (RAG) over your company data, function and tool calling into Microsoft Graph and line-of-business systems, multi-agent orchestration where the workflow justifies it, automated evaluations, content-safety guardrails, and telemetry — with the finished agent deployed to Microsoft Teams, a web experience, or Microsoft 365 Copilot. Work is time and materials at $175 per hour, typically from $12,500, and a typical build runs about six weeks; final scope depends on data readiness, integrations, security requirements, and the number of environments and channels.
What this engagement is
Microsoft renamed Azure AI Foundry to Microsoft Foundry at Ignite in November 2025, and the new name became official in Microsoft's Product Terms in January 2026. The platform, SDKs, and your existing Azure resources are the same — this page uses the current name throughout. This service is for organizations whose agent use case has outgrown low-code tooling: you need control over model selection, the retrieval pipeline, orchestration logic, source control, CI/CD, and measurable answer quality. We build on Foundry Agent Service and the Foundry SDKs — grounding the agent in your data through a retrieval index, wiring tool calls into Microsoft Graph and your business systems, adding evaluation runs and content-safety filters before launch, and instrumenting telemetry so you can see what the agent actually does in production. The engagement is engineering-led: everything lives in a repository you own, and the agent ships only after it passes the evaluation set you approved. If your scenario fits a low-code build, our Custom Agent Development with Microsoft Copilot Studio service is the right starting point instead — we will tell you which one applies during scoping.
Success criteria
What you receive
How the work unfolds
Discovery and use-case framing — confirm the business goal, target users, success measures, risk profile, and the level of autonomy the agent is allowed; validate that pro-code Foundry work is warranted over a low-code build.
Architecture and data review — assess grounding sources, permission models, model availability in your Azure region, integration targets, compliance constraints, and environment strategy; produce the solution design.
Environment setup — provision or validate the Foundry project, model deployments, retrieval index, source repository, and CI/CD path in your Azure subscription.
Grounding build — index the approved data sources, implement the retrieval pipeline, and verify that answers respect existing permissions.
Tools and orchestration — implement function/tool calling into Microsoft Graph and scoped business systems; add multi-agent orchestration only where the design calls for it.
Evaluation and safety — build the evaluation set with your subject-matter experts, run it against the agent, configure content-safety filters and guardrails, and iterate until agreed thresholds pass.
Pilot and deployment — publish to the agreed channel for a limited pilot group, review telemetry and feedback, fix what the pilot surfaces, then release to the approved audience.
Handoff — walk your team through the repository, runbook, evaluation assets, and telemetry; deliver the improvement backlog and the operations recommendation.
Prerequisites
Who does what
IT Partner
- Use-case feasibility, solution architecture, and the honest pro-code-versus-low-code recommendation.
- Foundry project setup, model deployment, and retrieval pipeline implementation in your Azure subscription.
- Agent build: instructions, grounding, function/tool calling, and orchestration code in a repository you own.
- Evaluation set construction, content-safety and guardrail configuration, and pre-launch test runs.
- Deployment to the agreed channel, telemetry wiring, pilot support, and fixes for what the pilot surfaces.
- Documentation, runbook, handoff, and the written operations recommendation.
Your team
- Provide the business goal, priority scenarios, acceptance criteria, and realistic test questions with expected answers.
- Assign business, technical, and security owners with authority to make timely decisions and approvals.
- Provide or approve Azure subscription access, tenant roles, service accounts, app registrations, and integration credentials.
- Prepare and permission the grounding data sources; own the quality of the underlying content.
- Purchase Microsoft licensing and carry Azure consumption costs for the agent's resources.
- Review evaluation results, participate in the pilot, and approve production release.
- Own internal communication, adoption, and business-process decisions after handoff unless separately contracted.
What's not included
Limitations & technical notes
Frequently asked questions
Is Microsoft Foundry the same thing as Azure AI Foundry?
Yes. Microsoft renamed Azure AI Foundry to Microsoft Foundry at Ignite in November 2025, and the new name became official in Microsoft's Product Terms in January 2026. The platform, SDKs, and existing deployments are unchanged — if your team still says 'Azure AI Foundry,' you are talking about the same thing this service builds on.
When do we need Foundry instead of Copilot Studio?
Choose pro-code Foundry work when you need control that low-code tooling doesn't give you: model selection, a custom retrieval pipeline, orchestration logic in source control, CI/CD, automated evaluations, or a custom user experience. If your scenario is a governed Q&A or workflow agent over Microsoft 365 data, our Copilot Studio service is usually faster and cheaper — and we'll say so during scoping.
Can the agent answer from our company data?
Yes — grounding the agent in your data is the core of the service. We build a retrieval-augmented generation (RAG) pipeline over approved sources such as SharePoint, databases, file stores, or APIs, and verify during evaluation that answers come from those sources and respect existing permissions.
Can the agent take actions, not just answer questions?
Yes. We implement function and tool calling so the agent can call Microsoft Graph and your line-of-business systems — creating records, looking up orders, filing tickets — within the permissions you approve. Every action path is covered by the evaluation set and scoped to least-privilege access.
Do you build multi-agent solutions?
When the workflow justifies it. Foundry supports orchestrating multiple specialized agents, and we design that way when a single agent would become an unmaintainable monolith. We don't add multi-agent complexity for its own sake — it increases cost, latency, and testing surface.
Which models can we use?
Foundry exposes a large model catalog, including OpenAI models and others, with availability varying by Azure region. We recommend a model based on your quality, latency, cost, and data-residency requirements, and validate the choice against what your subscription and region actually offer at project start.
How do you make sure the agent is safe and accurate before launch?
We build an evaluation set with your subject-matter experts — real questions with expected answers, plus failure and abuse cases — and run it against the agent before any user sees it. Content-safety filters and guardrails are configured and tested, and the agent ships only after the evaluation results you approved pass.
Where can the agent be deployed?
To Microsoft Teams, a web front end, or Microsoft 365 Copilot, depending on where your users work and what your licensing supports. The channel is agreed at scoping because it affects authentication, licensing, and the user experience we build.
How much does it cost?
Work is time and materials at $175 per hour, with typical projects starting from $12,500. The estimate is put in writing before work begins. Cost drivers are the number of grounding sources, the integrations the agent must call, evaluation depth, and how many environments and channels you need. Azure consumption is separate and billed by Microsoft to your subscription.
How long does a build take?
A typical build runs about six weeks from kickoff to handoff, assuming prerequisites — Azure access, data sources, and named owners — are ready at the start. Complex integrations, multiple channels, or slow review cycles extend that.
Who owns the code and the Azure resources?
You do. Everything is built in your Azure subscription and delivered in a source repository you own, with infrastructure definitions and an operating runbook. There is no proprietary layer of ours between you and the solution.
Is our data used to train the models?
Per Microsoft's documented data-privacy commitments for its Azure-hosted AI services, your prompts, retrieval data, and outputs are not used to train Microsoft's foundation models. Your data stays in your subscription, and we configure the agent to respect your existing permissions. Your compliance team should review Microsoft's current terms for your specific requirements.
What happens after the agent goes live?
The project ends with handoff: repository, runbook, evaluation assets, telemetry, and an improvement backlog. Agents drift after launch — knowledge goes stale, costs creep, usage shifts — so we recommend our Managed AI Agent Operations and Optimization service for ongoing monthly operations, or your team can run it from the runbook.
Do you train or fine-tune custom models?
No — model training and fine-tuning are outside this service. In our experience most business agent use cases are solved better and cheaper with the right existing model plus good retrieval and evaluation. If scoping shows your case genuinely needs fine-tuning, we'll say so and scope it separately.
What do we need to have ready before starting?
An Azure subscription we can deploy into, identified and accessible grounding data sources, named business and technical owners, access to any systems the agent must call, and your security and compliance requirements stated up front. The six-week typical timeline assumes these are in place at kickoff.
What is not included?
Low-code Copilot Studio builds (separate service), model training, Microsoft licensing and Azure consumption, broad data cleanup, ongoing operations after handoff, and enterprise-wide AI governance programs. Each has a named home in our catalog, and we cross-refer rather than blur the scope.