Guide Microsoft AI/SI · Chapter 5

    Which Microsoft platform fits which task.

    Where the work happens, which sources it needs and who maintains it decide between Copilot, Copilot Studio and Foundry. The platform finder gets you there.

    As of: October 2026 · Editor: Marcel Haas

    • Microsoft Copilot
    • Copilot Studio
    • Microsoft Foundry
    • Agent flows
    • Hosted agents
    • Entra Agent ID

    Platform finder

    The questions that lead to the layer.

    Every answer points at one or two layers. The layer with the most hits is the starting point; what to check there is listed below.

    1. 01

      Where does the work happen?

      • In Word, Outlook, Excel or Teams, triggered by a personleads toMicrosoft Copilot
      • In Teams or Copilot, as an agent for manyleads toCopilot Studio
      • In a workflow of your own, on a website or in a line-of-business systemleads toMicrosoft Foundry
      • In the repository or with an agent that OpenAI runsleads toOpenAI directly
    2. 02

      Which sources and actions does it need?

      • Only what the person has open or may searchleads toMicrosoft Copilot
      • Fixed sources in SharePoint or Dataverse and actions through connectorsleads toCopilot Studio
      • Interfaces of your own, databases, batch processingleads toMicrosoft Foundry
    3. 03

      Does it start on its own?

      • No, a person asksleads toMicrosoft Copilot
      • Yes, on an email, an entry or at a timeleads toCopilot StudioMicrosoft Foundry
      • Yes, as a workflow in code that waits for approvalsleads toMicrosoft Foundry
    4. 04

      Does it need a specific model?

      • No, the Copilot router is enoughleads toMicrosoft CopilotCopilot Studio
      • Yes, chosen on the test set, with deployment type and regionleads toMicrosoft Foundry
      • Yes, and Codex or the Agents API are part of itleads toOpenAI directly
    5. 05

      Who maintains it after handover?

      • Nobody has to maintain anythingleads toMicrosoft Copilot
      • The business team in the interfaceleads toCopilot Studio
      • A development team with a pipelineleads toMicrosoft Foundry

    Where the answers lead.

    Microsoft Copilot

    What that means: The task stays in the programs the person already uses; Copilot works with their rights.

    What you check: Permissions, labels and sharing in SharePoint, the licence tier per role, a pilot with measurable tasks.

    Open page

    Copilot Studio

    What that means: An agent with fixed sources, actions and triggers, reachable in Teams and Copilot, maintained by the business team.

    What you check: Sources and connectors, the Entra Agent ID and its permissions, Copilot Credits and a cap per agent, the test set.

    Open page

    Microsoft Foundry

    What that means: An application or agent of your own in your Azure subscription, with model, deployment type and network of your choice.

    What you check: Region and deployment type per model, the agent's identity, guardrails, budget and alerts, handover as infrastructure code.

    Open page

    OpenAI directly

    What that means: Codex for software work, the ChatGPT workspace or the Agents API; the OpenAI guide takes it from here.

    What you check: Sign-in through Entra ID, data region and retention per project, usage limits; see platform choice in the OpenAI guide.

    Open page
    01

    The task determines the layer.

    An employee submits expense receipts, and someone checks them against the policy. If that happens in Outlook and Excel, Copilot helps by summarising the policy next to the receipt. If every submitted receipt should be checked on its own and passed on for approval when it deviates, it takes an agent with fixed sources and actions. If receipts should be read from photos in large numbers and posted to the accounting system, it becomes an application of your own with a model that understands images. These questions separate the cases: Where does the work happen? Does it need fixed sources and actions in line-of-business systems? Must it start without being asked? Does it need a specific model? Who maintains it after handover?

    02

    In Microsoft 365: Copilot first, then an agent.

    Copilot helps where people search for information, summarise it and turn it into drafts, and it sees only what the person may see; as long as a person triggers the task and uses the result themselves, nothing more is needed. An agent in Copilot Studio takes over as soon as the task needs fixed sources, actions through connectors or a trigger without being asked, and should be reachable for many in Teams or Copilot. After handover the business team maintains it in the interface, and because every agent carries an Entra Agent ID, permissions, conditions and data rules apply per agent.

    03

    Foundry, when the task leads out of Microsoft 365.

    If the task needs a specific model, a workflow in your own code, processing without a chat or an interface outside Microsoft 365, we build it as an application on Microsoft Foundry, with GPT, Claude or open models from the catalogue and hosted agents in the Swiss regions. Prompt agents are created there without code from instructions, a model and tools; hosted agents bring your code in a container and wait at the point where a person decides, for days if need be. Both can be combined with Copilot Studio when employees should stay in Teams while the work behind it happens in the application.

    04

    What stays the same on all three layers.

    Sign-in runs through Entra ID, for people and for agents alike; labels, loss prevention and the log of every prompt and answer sit in Purview; and every application has a test set it is measured against. A change of layer therefore costs interface and connectors, while the security stays: what began with Copilot moves to Copilot Studio as an agent and to Foundry as an application, with the same identities and the same rules.

    05

    The test set decides.

    Two candidates work through the same set of real cases: correct, incomplete and prohibited requests, missing sources, failures. We compare answer quality, access boundaries, failure handling, the review effort for people and the cost per case, and the proposal records why one layer won. A pilot with one group then shows what holds up in daily work; after every change to model, sources or instructions, the test set runs again.

    What you decide.

    Copilot or an agent?

    Options

    1. 1Copilot in the programs, with clean permissions
    2. 2An agent in Copilot Studio with sources, actions and triggers

    Our recommendation: Copilot, as long as a person triggers the task and uses the result themselves, for instance when summarising, drafting and analysing. An agent, as soon as fixed sources, actions in line-of-business systems or a start without being asked come into play, or several people should do the same task in the same way.

    What we advise against: An agent for a task nobody has yet described cleanly by hand. Without sources, permitted actions and a point where a person decides, the test set measures nothing.

    Copilot Studio or Microsoft Foundry?

    Options

    1. 1Copilot Studio with connectors, agent flows and MCP
    2. 2Foundry with model catalogue, Agent Service and your own code
    3. 3Both, with the agent in Teams as the way into the Foundry application

    Our recommendation: Copilot Studio when the agent lives in Teams or Copilot, answers from SharePoint and Dataverse and your business team maintains it itself. Foundry when a specific model, a workflow in code, processing without a chat, an interface outside Microsoft 365 or private networks are required. Both together when employees should stay in Teams while the work behind it happens in an application.

    What we advise against: Foundry for an agent the business team wants to adjust itself next week, or Copilot Studio for a pipeline that processes documents in batches overnight.

    Prompt agent or hosted agent in Foundry?

    Options

    1. 1A prompt agent from instructions, model and tools, without code of your own
    2. 2A hosted agent with your code in a container, with Agent Framework or LangGraph
    3. 3A pipeline without an agent that runs in batches

    Our recommendation: A prompt agent when instructions, a model and tools from the toolbox carry the task and your team should maintain it in the portal. A hosted agent when the workflow has several stages, waits for approvals or needs libraries that exist only in code. A pipeline without an agent when nobody talks to the system and documents run through in batches overnight.

    What we advise against: A hosted agent for a task a prompt agent solves in the portal; every line of code wants to be operated.

    Who builds, and who maintains?

    Options

    1. 1The business team in Copilot Studio
    2. 2Your developers or we on Foundry
    3. 3Both: the business team in front, development behind it

    Our recommendation: The business team maintains what lives in the Copilot Studio interface: sources, topics, connectors, agent flows. Developers maintain code, pipelines and hosted agents on Foundry. In combined solutions the agent in Teams belongs to the business team and the application behind it to development, with one shared test set.

    What we advise against: An agent nobody maintains after handover. Sources go stale, instructions stay as they were, and the test set no longer runs.

    Status and sources.

    As of: October 2026

    • Since June 2026

      New authoring experience in Copilot Studio, new primary models.

      The new authoring experience on the GitHub Copilot harness is available as a production-ready preview: Microsoft IQ for emails, calendar, files and Teams messages, reusable skills and memory per user. Claude Sonnet 5 and GPT-5.5 Chat are generally available as the primary model.

      Source: Microsoft Learn: What's new in Copilot Studio
    • Billing as of 14 September 2026

      Copilot Credits, included for licensed employees.

      Agents are billed in Copilot Credits, through prepaid packs or pay-as-you-go. Internal use by people with a Microsoft Copilot licence is included within fair-use limits, computer use excepted. In the Power Platform admin centre a monthly cap can be set per agent. This applies to the standard harness.

      Source: Microsoft Learn: Billing rates and management
    • As of 11 September 2026

      Hosted agents run in Switzerland North and Switzerland West.

      Your own agent code, built with Agent Framework, LangGraph or no framework at all, runs as a container in Foundry Agent Service: with its own Entra ID, its own endpoint, an isolated sandbox per session and traces in Application Insights. Both Swiss regions are on the list of supported regions.

      Source: Microsoft Learn: Hosted agents in Foundry Agent Service
    • Since September 2026

      Copilot Cowork is generally available.

      Cowork takes a described task from start to finish: it gathers context from emails, meetings, chats and files, works across several tools and delivers a finished result. Who reviews the result before it is passed on is set by your directive.

      Source: Microsoft Learn: Release notes for Microsoft Copilot

    Frequently asked questions.

    Can we start with Copilot and build an agent later?

    Yes, and that is the usual route. The permissions, labels and order in SharePoint that Copilot needs, an agent needs just the same. The tasks where employees keep asking Copilot the same thing are the first candidates for an agent, and the cases from that time become its test set.

    What happens to a Copilot Studio agent if we need Foundry later?

    Sources, instructions and the test set move with it; interface and connectors are rebuilt in the application. Often the agent in Copilot Studio remains as the way in from Teams and calls the Foundry application as a tool, so employees relearn nothing.

    Do we have to program for Copilot Studio?

    Not for sources, topics, connectors and agent flows; that happens in the interface, and the business team maintains it itself after handover. Code comes into play when a tool is built as an MCP server or an action should reach a line-of-business system for which there is no connector.

    Can agents from Copilot Studio and Foundry work together?

    Yes. An agent in Copilot Studio calls a Foundry application as an action or through an MCP server, a Foundry agent appears in Teams and Copilot itself once published, and for agents talking to each other Microsoft offers the A2A protocol. Every agent keeps its Entra Agent ID and its log.

    When is Copilot Chat without a licence enough?

    For work with web data and with files the person uploads or has open; chat with internet results belongs to eligible Microsoft 365 subscriptions. As soon as Copilot should fetch emails, chats, meetings and files itself, the person needs the Copilot licence. In the pilot we show both on the same tasks.

    Who decides on the platform in the end: IT or the business?

    Both, with distinct roles. The business team describes the task, sources, permitted actions and the point where a person decides; your IT sets the frame of approved layers, identities, data rules and cost. The proposal names the layer with the reasoning, and the test set checks it before anything is opened to everyone.

    Next chapter · Chapter 6

    Data boundaries: where Copilot and Foundry process.

    Where do Copilot and Foundry process our data, and what may we send there?