Microsoft Copilot, Agent Builder, Copilot Studio & Foundry — When to Use What

Almost every Microsoft AI conversation in 2026 starts the same way: "We want Copilot." Two meetings later, half the room means Microsoft 365 Copilot, a quarter mean Copilot Studio, and someone in the back is quietly building on Azure AI Foundry. They are not the same product, they are not priced the same, and they are not governed the same.

Here is the practical map we use with enterprise customers — four tiers, each with a clear "use when" rule, plus one tiny worked example so you can see the choice in action.

The four tiers of Microsoft's Copilot stack

How they fit together

These tiers are not alternatives — they layer:

A simple worked example: an HR Policy Assistant

Imagine HR wants an internal assistant that answers questions like "How many days of paternity leave do I get in Germany?" from the company policy library on SharePoint, and can also open a leave-request ticket in ServiceNow.

Option A — Microsoft 365 Copilot (zero build)

If your policies already live in SharePoint and users are licensed for M365 Copilot, they can ask Copilot directly today. Copilot will retrieve from the Graph and cite the SharePoint document. Cost: license only. Build effort: none. Limitation: no ServiceNow action.

Option B — M365 Agent Builder (15 minutes)

HR opens Copilot Chat → Create agent → points it at the "Policies" SharePoint site, adds instructions ("Always cite the policy URL and the country it applies to"), and publishes to the HR Team in Teams. No code. Still no ServiceNow action.

Option C — Copilot Studio (1–2 weeks)

IT builds the agent in Copilot Studio: SharePoint as knowledge, a ServiceNow connector for the action, a Power Automate flow for approval routing, and Entra ID auth so the agent only sees policies the user is entitled to. Channels: Teams + intranet web chat. This is the right tier for most enterprise use cases.

Option D — Azure AI Foundry (when you need more)

Same assistant, but HR is in a regulated business that requires the model to run in a specific region, with full evaluation pipelines, content-safety filters, and fine-tuning on internal terminology. The agent is orchestrated in Foundry and surfaced back into Teams via a Copilot Studio agent that calls the Foundry endpoint.

# Pseudocode for a Copilot Studio "Call Foundry Agent" action
from azure.ai.foundry import AgentsClient

agent = AgentsClient(endpoint=os.environ["FOUNDRY_ENDPOINT"]).agents.get("hr-policy-eu")

def ask(user_question: str, user_country: str) -> str:
    response = agent.invoke(
        input=user_question,
        metadata={"country": user_country, "policy_scope": "hr"},
    )
    return response.output_text

Choosing in 60 seconds