Building AI Agents in the Microsoft 365 / Copilot Stack for SMBs (2026)

How the Microsoft 365 / Copilot agent stack fits together for SMBs in 2026: six layers from Copilot Studio to Azure AI Foundry — roles, pricing, adoption.

Updated: May 29, 2026
by Michael Kerkhoff

TL;DR

The short version: these six Microsoft building blocks don't compete with each other — they stack on top of each other. You don't pick one; you snap them together like Lego to build your own AI agent. In order: Microsoft 365 Copilot (now in its new "Wave 3," live since early May 2026) is the surface your team already works in all day — Word, Outlook, Teams. It already includes Copilot Studio: the builder where an IT lead creates a custom agent in about 30 minutes, with no coding. You add the free Microsoft Learn MCP server there as a knowledge source, so the agent draws on current, accurate information. Power Automate is the next step: now the agent doesn't just answer, it gets things done — open a ticket, update the CRM. Azure AI Foundry is the pro tier for developer teams, for when things get more complex later. And Langdock, out of Berlin, is the option for anyone who needs their data to stay in the EU. Most mid-sized companies start small — Copilot Studio plus the Learn MCP server — and add the rest only once their first agents are running. The one genuine "either/or" here isn't "which Microsoft tool wins," it's: all-Microsoft, or an EU-owned layer like Langdock alongside it.

The 6 Layers of the Microsoft 365 Agent Stack

1

The surface your team already lives in all day — Word, Outlook, Teams — and the foundation everything else builds on. The new "Wave 3" (live since early May 2026) adds three things: Cowork (Copilot thinks ahead and proactively suggests the next steps), Agent Mode right inside the document, and Agent 365 — a single place to roll out your own agents and keep an eye on them. The practical kicker: if you have a Copilot license, you already own the builder for the next layer (Copilot Studio) and just switch it on.

AI right inside the Office apps, Agent Mode, several agents working together (Agent 365), access to your Microsoft data~$30/user/month (M365 Copilot); requires an underlying M365 business plan
2

This is where you build the agent — a kit-of-parts tool, no coding required. An IT lead has a working agent ready in about 30 minutes. The latest update (2026 Release Wave 1) adds finer control over what the agent is allowed to do, a new editor for workflows, agents that can operate a computer on their own, and real-time voice. Copilot Studio sits directly on top of M365 Copilot (free for internal agents; you only need an extra license for external channels like your own website) — and it's where the knowledge, action and pro-code layers below come together into one finished agent.

Agents without coding, connection to your own knowledge, automated workflows, publishing to external channelsIncluded in M365 Copilot for internal agents; standalone $200/month for 25,000 credits (or ~$0.01/credit pay-as-you-go via Azure)
3

A building block, not a rival platform — a free, public service that makes Microsoft's complete, up-to-the-day documentation available to your agent. You add it in Copilot Studio (layer 2) as a knowledge source, and in about 15 minutes you have an agent that bases its answers on the real, current docs instead of guessing. Think of it as the ideal first step: no code, no real risk, and the only cost is the Copilot Studio credits the agent spends while it works.

Docs as a knowledge source, up-to-the-day access to Microsoft Learn, connecting further toolsFree server; you only pay Copilot Studio credits per query
4

The layer that turns "answering" into "getting it done." The latest update (2026 Release Wave 1) adds so-called agent flows plus guardrails for processes that run on their own — so your Copilot Studio agent can read a CRM, update a SharePoint list, open a ticket or kick off multi-step processes across more than 1,000 ready-made connections. It extends the agents you built in layer 2; it doesn't replace them.

Automating processes, agent flows, actions across multiple apps, guardrails (DLP)Power Automate Premium add-on; consumption/connector-based pricing
5

The pro tier for developer teams — you move here only once the simple kit hits its limits, not as a first choice. The Foundry Agent Service (generally available) gives developers full control: several agents working together, free choice of model, and deep integration with Azure search. Many companies run the simple layers (2–4) and this pro layer side by side — for the tasks where you need full control over the model, its memory and the underlying tech.

Several agents working together (for developers), free choice of model, custom memory & tools, enterprise scaleConsumption-based (token + agent/orchestration costs) via Azure
6

The sovereignty layer — and the one point where a real "either/or" exists. A Berlin-based platform that can connect to any AI model and is hosted entirely in the EU. Many companies in the DACH region use M365 Copilot for everyday work in the Office apps and add Langdock as a company-wide, EU-owned AI hub for especially privacy-sensitive topics — deliberately both at once. So the real strategic question isn't "which Microsoft tool," it's: all-Microsoft, or an EU-owned layer alongside it.

AI hub with free choice of model, data stays in the EU, assistants & workflows, answers from your own company knowledgePer-user / enterprise (see langdock.com/pricing)

Layer-by-Layer Overview

NameWhat it is forTechWho builds itPriceAI-Native
AI right inside the Office apps, Agent Mode, several agents working together (Agent 365), access to your Microsoft dataM365 apps, Microsoft Graph, Agent 365, Copilot Studio, Entra IDAll employees (per-seat)~$30/user/month (M365 Copilot); requires an underlying M365 business plan
Agents without coding, connection to your own knowledge, automated workflows, publishing to external channelsCopilot Studio, Power Platform connectors, Dataverse, MCP, M365 GraphBusiness users & IT leads (no developers required)Included in M365 Copilot for internal agents; standalone $200/month for 25,000 credits (or ~$0.01/credit pay-as-you-go via Azure)
Docs as a knowledge source, up-to-the-day access to Microsoft Learn, connecting further toolsModel Context Protocol (MCP), Copilot Studio, Microsoft Learn catalogIndividuals & IT leads (no developers required)Free server; you only pay Copilot Studio credits per query
Automating processes, agent flows, actions across multiple apps, guardrails (DLP)Power Automate, Power Platform connectors, Dataverse, Copilot StudioPower users & IT (low-code)Power Automate Premium add-on; consumption/connector-based pricing
Several agents working together (for developers), free choice of model, custom memory & tools, enterprise scaleAzure AI Foundry, Semantic Kernel, Azure AI Search, Azure OpenAI / model catalogDeveloper & ML teamsConsumption-based (token + agent/orchestration costs) via Azure
AI hub with free choice of model, data stays in the EU, assistants & workflows, answers from your own company knowledgeMulti-LLM (frontier models from OpenAI/Anthropic/Google/Mistral), EU hosting, RAG, workflow layerOrg-wide rollout (per-seat / enterprise)Per-user / enterprise (see langdock.com/pricing)

← Scroll horizontally to see all columns

How the Layers Fit Together

  • Read this as a stack, not a shortlist: the six entries are layers you combine, not rivals you pick one of. The numbers are the recommended order, not a quality ranking.
  • Start with the easiest quick win: if you have M365 Copilot (layer 1), build your first agent in Copilot Studio (layer 2, free for internal agents) and add the free Microsoft Learn MCP server (layer 3) as a knowledge source — a 15-to-30-minute project at no extra cost that quickly shows what's possible.
  • Add the action layer when answers aren't enough: with Power Automate (layer 4) the agent does real work — reading systems, updating records and kicking off multi-step processes across more than 1,000 connections, kept safe by guardrails.
  • Move to the pro tier only when the simple kit isn't enough anymore: tasks that need bespoke solutions, their own memory or specific models move to Azure AI Foundry (layer 5) — alongside the simple layers, not in their place.
  • Treat data sovereignty as the one real either/or: if it matters that your data stays in the EU or that you can freely choose the model, put an EU-owned layer like Langdock (layer 6) next to the Microsoft stack. The real decision is "all-Microsoft" vs "an EU-owned layer alongside" — not which Microsoft tool wins.
  • Budget by usage, not by users: Copilot Studio on its own is $200/month for 25,000 credits. A simple, pre-set answer costs about 1 credit, while an independent reasoning step can cost ~100 — work out the usage before you roll out widely.

Frequently Asked Questions

For most mid-sized companies the fastest path is Microsoft Copilot Studio together with the free Microsoft Learn MCP server. If you already have a Microsoft 365 Copilot license, building internal agents in Copilot Studio costs nothing extra — an IT lead has a working agent ready in about 30 minutes, with no coding. Add the public Microsoft Learn MCP server as a knowledge source and the up-to-the-day documentation becomes an agent you can talk to in roughly 15 minutes. You only pay Copilot Studio credits when the agent actually runs.

Copilot Studio is already included with Microsoft 365 Copilot — you build and use internal agents at no extra cost. The standalone license is $200/month for 25,000 credits (or roughly $0.01 per credit pay-as-you-go via Azure) and additionally lets you publish agents to external channels such as a website or app. What it really costs depends on the agent: a simple, pre-set answer costs about 1 credit, while an independent reasoning step costs around 100 — so work out your expected usage before rolling out widely.

Microsoft 365 Copilot Wave 3 has been live since early May 2026 and brings three core building blocks: Cowork (Copilot thinks ahead and proactively suggests the next steps), Agent Mode right inside the document, and Agent 365 — a central place to roll out, manage, secure and govern your own agents across the company. It turns the reactive chat assistant into a helper that works on its own and across multiple apps — based on your Microsoft data.

The Microsoft Learn MCP server is a free, public service that makes Microsoft's complete, up-to-the-day Learn documentation available to any compatible agent, including Copilot Studio. Instead of laboriously training a model on the docs (expensive and quickly out of date), it gives the agent real-time access to the current, reliable documentation. An ideal first step: no coding required, the server is free, and you only pay Copilot Studio credits per query.

Stay on the simple path (Copilot Studio and Power Automate agent flows) as long as it's enough — it's faster to set up and cheaper to run. Move to Azure AI Foundry when you need full developer control: bespoke teamwork between several agents, their own memory and tools, specific models, deep integration with Azure search, or enterprise-scale infrastructure. The Foundry Agent Service is generally available and lets several agents work together openly.

Not necessarily — but many companies in the DACH region and across Europe deliberately combine both. M365 Copilot Wave 3 shines at everyday work in Word, Excel, Outlook and Teams, while Langdock — a Berlin-based platform with free choice of model and EU hosting — serves as a company-wide, EU-owned AI hub. For 2026, many EU companies' answer is "both": Copilot for work inside the Office apps, plus an EU-owned layer for privacy-sensitive topics and to avoid depending on a single vendor.

Yes. The difference between a chatbot and an agent is exactly that — taking action. Power Automate (2026 Release Wave 1) adds automated processes with built-in guardrails — so a Copilot Studio agent can read a CRM, update a SharePoint list, open a ticket or kick off multi-step processes across more than 1,000 ready-made connections — all within one conversation and without a developer writing code. Guardrails (DLP for processes that run on their own) keep everything within the allowed limits.

Related Resources

Sources & Further Reading

Context Studios

Ready to start your AI project?

Book a free 30-minute consultation to discuss your requirements and find the right approach.

Book Consultation