Visual Framework

Interactive decision trees to guide Microsoft AI technology selection.

Use these diagrams after working through the Decision Framework and Evaluation Criteria. They reinforce the nine critical questions and are designed to support workshops, architecture reviews, and executive readouts.

Table of contents

  1. Diagram Index
  2. Complete Decision Flow
  3. Persona-Based Flow
  4. Data Grounding Decision
  5. Complexity Assessment Flow
  6. Budget & Timeline Tradeoffs
  7. Governance & Compliance Path
  8. Multi-Agent Orchestration
  9. Upgrade Paths
  10. Lifecycle Check
  11. IQ Layer Selection

Diagram Index

Diagram Purpose Maps To Framework
1. Complete Decision Flow End-to-end technology selection: agent checkpoint, Q0 buy-before-build gate, the 3 AI Buckets, and platform routing Phase 2: Q0 to Q9 - All critical questions
2. Persona-Based Flow Role on-ramps decided by runtime ownership (Maker harnesses, Business Builder apps, Developer loop) Phase 2: Q2 and Q8 + Scenarios
3. Data Grounding Decision Question shape, trust boundary, retrieval store, and vector engine Phase 2: Q3 - Data grounding patterns
4. Complexity Assessment The four behavior-and-authority tiers (Informational, Transactional, Reasoning, Autonomous) Evaluation Criteria: Complexity
5. Budget & Timeline The four funding lanes: Seats (USL), Copilot Credits (UBB), Azure meters, and Prepaid commitments Evaluation Criteria: Budget
6. Governance & Compliance Five ANDed gates: risk tier, trust boundary, identity & authorization, control layer, and network isolation Evaluation Criteria: Governance
7. Multi-Agent Orchestration Boundary gate (Soloist vs. Orchestra), connected agents, and workflow frameworks Phase 2: Q4 and Quick Reference: Orchestration
8. Upgrade Paths Progressive enhancement of low-code agents and promotion of business-built apps Implementation Patterns: Progressive Enhancement and Scenario 11
9. Lifecycle Check Pre-flight readiness gate for retirement horizons, maturity statuses, and financial commitments Quick Reference: Lifecycle Status and The Migration Calendar
10. IQ Layer Selection Which Microsoft IQ layer answers the question (Foundry IQ, Work IQ, Fabric IQ, Web IQ) Microsoft AI Stack: Microsoft IQ

Complete Decision Flow

Interactive version available. Pan, zoom, and click nodes in the embedded explorer below, or open the full explorer.

View as static Mermaid diagram
%%{init: {'theme':'dark'}}%%
flowchart TD
    Start([Start: a business problem]) --> G0{Do you need an agent<br/>at all?}
    G0 -->|Structured, predictable| StopCode([Stop: deterministic code<br/>or non-generative AI])
    G0 -->|Static retrieval only| StopRAG([Stop: classic RAG,<br/>no agent])
    G0 -->|Reasoning or tool use| Q0{Q0: Has someone<br/>already built this?}
    Q0 -->|Yes: use, browse, configure| DoneAdopt([Adopt it: use, browse,<br/>or configure])
    Q0 -->|No, and I can say why| Q1{Q1: Who benefits?}

    Q1 -->|My workforce| B1{AI for You:<br/>what do people need?}
    Q1 -->|Customers or systems| B2Gates[AI as a Product:<br/>set constraints first]
    Q1 -->|Building software| B3{AI for Your Codebase:<br/>who owns the result?}

    B1 -->|Answers in apps they use| B1Use[Microsoft 365 Copilot<br/>Chat first]
    B1 -->|Hand off a multi-step job| B1Delegate[Copilot Cowork:<br/>hand-off with checkpoints]
    B1 -->|Our knowledge inside Copilot| B1Extend[Copilot Studio:<br/>Copilot chat harness<br/><i>GA not stated</i>]

    B3 -->|Engineers, in a repo| B3Dev[GitHub Copilot:<br/>editor to cloud agent to SDK]
    B3 -->|A team, by description| B3Buy{Does a list, workbook,<br/>or SaaS already solve it?}
    B3 -->|Details| XPersona[[Persona flow:<br/>builder and developer paths]]
    B3Buy -->|Yes| B3BuyOut([Use it.<br/>Skip the app.])
    B3Buy -->|No| B3Biz[Describe-it builders<br/><i>Preview</i>]

    B2Gates -->|Open data flow| XData[[Q3: Data Grounding flow]]
    B2Gates -->|Open multi-agent flow| XMulti[[Q4: Multi-Agent flow]]
    B2Gates -->|Open governance flow| XGov[[Q5, Q7: Governance flow]]
    B2Gates -->|Open budget flow| XBudget[[Q6: Budget flow]]
    B2Gates -->|Open lifecycle flow| XLife[[Lifecycle Check flow]]
    B2Gates -->|Constraints written down| B2Start{How does work start?<br/>Q9}

    B2Start -->|A person asks, chat UI| B2Control{Q2, Q8: Configure,<br/>or code and own it?}
    B2Start -->|An event or schedule| AutoQ{What fires it?}
    B2Start -->|A system calls it, API| HostQ{Who hosts it?}

    B2Control -->|Configure, low-code| Studio[Copilot Studio<br/>pick a harness]
    B2Control -->|Code it, pro-code| PlatQ{Primary platform?}
    PlatQ -->|Microsoft 365-centric| SDK[M365 Agents SDK]
    PlatQ -->|Azure-centric| UIProto{Need a custom<br/>UI protocol?}
    UIProto -->|Yes| AGUI[Agent Framework + AG-UI<br/><i>GA not stated</i>]
    UIProto -->|No| Foundry[Microsoft Foundry - Azure]

    AutoQ -->|Low-code triggers| StudioAuto[Copilot Studio triggers<br/><i>GA not stated</i>]
    AutoQ -->|Schedule, GitHub, Teams| Routines[Foundry routines]
    AutoQ -->|Many business systems| LogicApps[Logic Apps agentic workflows<br/><i>Preview</i>]
    AutoQ -->|Code-defined workflow| AgentFW[Microsoft Agent Framework]

    HostQ -->|Managed, PaaS| AgentService[Foundry Agent Service]
    HostQ -->|Self-hosted| AgentFW
    HostQ -->|On-device or edge| FoundryLocal[Foundry Local:<br/>on-device, edge]

    Studio --> XPublish[[Q8: Publish targets<br/>by platform]]
    SDK --> XPublish
    AGUI --> XPublish
    Foundry --> XPublish
    XPublish --> Monitor[Monitor and govern<br/>Phase 4]
    StudioAuto --> Monitor
    Routines --> Monitor
    LogicApps --> Monitor
    AgentFW --> Monitor
    AgentService --> Monitor
    FoundryLocal --> Monitor
    Monitor --> End([Production])

    style B1Use fill:#004578,color:#fff
    style B1Delegate fill:#004578,color:#fff
    style B1Extend fill:#004578,color:#fff
    style B3Dev fill:#4b2070,color:#fff
    style B3Biz fill:#004578,color:#fff
    style B2Gates fill:#8c5e00,color:#fff
    style Studio fill:#004578,color:#fff
    style SDK fill:#4b2070,color:#fff
    style AGUI fill:#4b2070,color:#fff
    style Foundry fill:#a52617,color:#fff
    style StudioAuto fill:#004578,color:#fff
    style Routines fill:#a52617,color:#fff
    style LogicApps fill:#004578,color:#fff
    style AgentFW fill:#4b2070,color:#fff
    style AgentService fill:#a52617,color:#fff
    style FoundryLocal fill:#004578,color:#fff
    style DoneAdopt fill:#0b6a0b,color:#fff
    style End fill:#0b6a0b,color:#fff

Validation Summary

Last Validated: October 14, 2026

Read the status column, not the product name. Copilot Studio’s Microsoft Copilot harness is GA as of 2026-08-03, as are generative orchestration (within the standard harness), computer use, A2A (on the standard harness), and the Foundry IQ integration (per the 2026-10-07 blog). Its autonomous/triggered agents and child/inline agents carry no explicit GA statement, connected agents on the Microsoft Copilot harness (Studio, Foundry, A2A, and M365 Agents SDK, updated 2026-10-14) carry Preview in the What’s New feed and no banner on the how-to pages, and memory, apps, and hooks remain preview. (Scheduled prompts and Copilot Tuning are Microsoft Copilot features, not Copilot Studio features.) (Copilot Studio what’s new)

UI-Based Agents (GA unless noted)

Technology Action Safety Proactive Description
M365 Copilot 🔒 User-in-the-loop always 🔄 Reactive (or delegated via Copilot Cowork and scheduled prompts) Conversational chat and delegated work in M365 apps (docs)
Copilot Studio ⚠️ Actions can execute (add approval workflows) 🔄 Reactive (conversational) or ✅ Autonomous (event triggers / workflows) Three harnesses fixed at creation: Copilot chat, Standard, and GitHub Copilot (docs)
M365 Agents SDK ⚠️ Custom action safety design ✅ Proactive capable Pro-code, 10+ channels, C#/JS/Python, BYO orchestrator (docs)
Microsoft Foundry (Azure) ⚠️ Autonomous planning loops ✅ Proactive capable Custom UI and GA publish path to Microsoft Copilot and Teams (docs)
Agent Framework + AG-UI ⚠️ Approvals via AG-UI middleware ✅ Proactive capable (inherits host orchestration) Python agent-framework-ag-ui 1.5.1 is stable on PyPI (2026-10-08); .NET Hosting.AGUI.AspNetCore is Preview; Learn states no overall status banner (docs)

Autonomous & Event-Driven Agents

Technology Action Safety Proactive Description
Copilot Studio (event triggers / workflows) ⚠️ Actions can execute ⚠️ Triggered execution, not documented as GA Event triggers on the standard harness; manual, schedule, and event triggers in Workflows on the Microsoft Copilot harness (docs)
Foundry routines (GA) ⚠️ Autonomous execution ✅ Timer, recurring, and GitHub/Teams event triggers Listed as GA in Learn’s GA readiness table (2026-09-09) and Foundry blog (2026-09-24). Reminder tool and .NET SDK packages remain preview; prompt and hosted agents only (docs)
Logic Apps agentic workflows ⚠️ Autonomous execution ✅ Proactive (event-driven) Official terms are “agentic workflows” and “agent loop”. Consumption is explicitly in preview. Standard carries no preview banner on the agent loop, but Microsoft never states Standard is GA and specific Standard model sources are marked preview (docs)
Microsoft Agent Framework (GA core) ⚠️ Code-defined middleware and checkpoints ✅ Proactive capable GA for Python (1.21.0) and C#/.NET (1.25.0); Go is Public Preview (0.1.1) (docs)

API/Headless & On-Device Services (GA unless noted)

Technology Action Safety Proactive Description
Foundry Agent Service ⚠️ Autonomous planning loops ✅ Proactive capable REST Responses API, managed PaaS for Prompt Agents (GA) and Hosted Agents (GA service, Python and C# only) (docs)
Foundry Local 🔒 Local containment 🔄 Host-driven On-device and edge SLM execution with cloud-parity API (docs)

Sources: Complete Decision Flow

Persona-Based Flow

Interactive version available. Pan, zoom, and click nodes below, or open the full explorer.

View as static Mermaid diagram
%%{init: {'theme':'dark'}}%%
flowchart TD
    Start([Start: where are you<br/>starting from?]) --> P1{Your role?<br/>an on-ramp, not a gate}

    P1 -->|End user| User[Microsoft 365 Copilot<br/>plus Cowork for hand-offs]
    P1 -->|Business maker| Maker{What are you making?}
    P1 -->|Business builder| BBuy{Does a list, workbook,<br/>or SaaS already solve it?}
    P1 -->|Developer| Dev{What are you building?}
    P1 -->|Data or analytics| DS{Primary workload?}
    P1 -->|Integration| IntSpec[Logic Apps agentic workflows<br/><i>Preview</i>]

    Maker -->|An agent| MkHarness{Which Studio harness?<br/>fixed at creation}
    Maker -->|A workflow| MkWF[Copilot Studio workflows<br/>or agent flows<br/><i>GA not stated</i>]
    Maker -->|An app| BBuy

    MkHarness -->|Employee answers in Copilot| MkChat[Copilot chat harness:<br/>extend Microsoft Copilot<br/><i>GA not stated</i>]
    MkHarness -->|Short, bounded, explicit| MkStd[Standard harness:<br/>rule-based, predictable]
    MkHarness -->|Long-running, reasoning| MkGH[Microsoft Copilot harness:<br/>reasoning-heavy]
    MkHarness -->|Customer support: voice and chat| MkCustomer[Customer Conversation harness<br/><i>Public Preview</i>]

    BBuy -->|Yes| BBuyOut([Use it.<br/>Skip the app.])
    BBuy -->|No| BOwner[Name an owner first:<br/>before any share link]
    BOwner --> BWho{Who will use it?}
    BWho -->|Just me| BCode[Copilot Code<br/><i>Preview</i>]
    BWho -->|My team, over M365 content| BCowork[Cowork App skill<br/><i>Preview</i>]
    BWho -->|One step in a process| BStudio[Copilot Studio apps<br/><i>Preview</i>]
    BWho -->|The department depends on it| BHand[Same repo: GitHub Copilot<br/>plus Managed Runtime CLI<br/><i>Preview</i>]
    BWho -->|On the Power Platform| BVibe[Power Apps vibe<br/>forks on handoff<br/><i>Preview</i>]
    BWho -->|Customers| BCust[Customers? Settle hosting<br/>and sign-in first]

    Dev -->|Software, with AI helping| DevLoop{Which rung of the<br/>developer loop?}
    Dev -->|An agent people talk to| DevOwn{Own the runtime<br/>and the on-call?}
    Dev -->|An agent that acts on events| DevAuto{What fires it?}
    Dev -->|Several cooperating agents| DevOrch[Microsoft Agent Framework<br/>boundary-driven only]
    Dev -->|A custom streaming UI| DevAGUI[Agent Framework + AG-UI<br/><i>GA not stated</i>]

    DevLoop -->|Stay in the editor| DevIDE[GitHub Copilot agent<br/>mode in the IDE]
    DevLoop -->|Delegate a whole issue| DevCloud[Copilot cloud agent:<br/>issue in, PR out]
    DevLoop -->|Specialize the agent| DevCustom[Custom agents<br/>plus AGENTS.md]
    DevLoop -->|Build on the harness| DevGHSDK[GitHub Copilot SDK:<br/>Foundry hosts Python, C sharp]

    DevOwn -->|No: configure, extend| DevStudio[Copilot Studio<br/>plus custom code]
    DevOwn -->|Yes, Microsoft 365-centric| DevM365[M365 Agents SDK]
    DevOwn -->|Yes, Azure-centric| DevAzure[Microsoft Foundry - Azure]

    DevAuto -->|Schedule, GitHub, Teams| DevRoutines[Foundry routines]
    DevAuto -->|Many business systems| DevLogic[Logic Apps agentic workflows<br/><i>Preview</i>]
    DevAuto -->|Code-defined workflow| DevOrch
    DevAuto -->|Low-code triggers| DevStudio

    DS -->|Analytics and BI| DSFabric[Fabric data agents]
    DS -->|ML and custom models| DSFoundry[Microsoft Foundry - Azure:<br/>ML and custom models]

    style User fill:#0b6a0b,color:#fff
    style MkChat fill:#004578,color:#fff
    style MkStd fill:#004578,color:#fff
    style MkGH fill:#004578,color:#fff
    style MkCustomer fill:#004578,color:#fff
    style MkWF fill:#004578,color:#fff
    style BOwner fill:#8c5e00,color:#fff
    style BCode fill:#004578,color:#fff
    style BCowork fill:#004578,color:#fff
    style BStudio fill:#004578,color:#fff
    style BHand fill:#4b2070,color:#fff
    style BVibe fill:#004578,color:#fff
    style BCust fill:#8c5e00,color:#fff
    style DevIDE fill:#4b2070,color:#fff
    style DevCloud fill:#4b2070,color:#fff
    style DevCustom fill:#4b2070,color:#fff
    style DevGHSDK fill:#8c5e00,color:#fff
    style DevStudio fill:#004578,color:#fff
    style DevM365 fill:#4b2070,color:#fff
    style DevAzure fill:#a52617,color:#fff
    style DevRoutines fill:#a52617,color:#fff
    style DevLogic fill:#004578,color:#fff
    style DevOrch fill:#4b2070,color:#fff
    style DevAGUI fill:#4b2070,color:#fff
    style DSFabric fill:#8c5e00,color:#fff
    style DSFoundry fill:#a52617,color:#fff
    style IntSpec fill:#004578,color:#fff

Validation Summary - Persona-Based Flow

P1 is a node id, not a taxonomy. Microsoft does not use “1P/3P” as documented terminology. The documented ownership taxonomy is the Agent Registry four types: Microsoft agents (“built and maintained by Microsoft”) · External partner-built agents (“built by trusted non-Microsoft developers”) · Published by your org (“custom agents approved and published by your organization… might be referred to as LOB agents”) · Shared by creator (“created and shared by individual users or developers at your organization”). Learn those four; map the slang once if you must. (Agent Registry)

Last Validated: October 14, 2026

End User (GA)

Technology Description Documentation
M365 Copilot & Cowork Built-in AI in M365 apps and long-running delegated work via Copilot Cowork (GA for work/school accounts; Preview for personal accounts) M365 Copilot | Cowork

Business Maker (GA unless noted)

Technology Description Documentation
Copilot Studio harnesses Choose by interaction and control: Customer Conversation harness (Public Preview) for customer voice/chat support; Copilot chat harness (extends Microsoft Copilot, internal publishing only, status not stated), Standard harness (rule-based and generative orchestration, GA, billed after publish), and Microsoft Copilot harness (reasoning-heavy, GA 2026-08-03, billed from first build/test/evaluation) Harnesses in Copilot Studio | Compare harnesses
Copilot Studio workflows & agent flows Deterministic process automation inside Studio. Agent flows are GA; Workflows on the Microsoft Copilot harness carry no page-level preview banner (updated 2026-10-12), while specific nodes such as adding a Copilot node are Preview Workflows overview

Developer (GA unless noted)

Technology Description Documentation
M365 Agents SDK Pro-code for M365-centric solutions, C#/JavaScript/Python, 10+ channels, BYO orchestrator M365 Agents SDK
Microsoft Foundry (Azure) Pro-code for Azure-centric solutions, custom models, full control Microsoft Foundry
Microsoft Agent Framework Microsoft’s investment direction - Microsoft describes it as the “direct successor” and “next generation of both” Semantic Kernel and AutoGen. Orchestration patterns: Sequential, Concurrent, Handoff, Group Chat, Magentic. Languages: Python (1.21.0 GA), C#/.NET (1.25.0 GA), and Go (0.1.1 Public Preview). No sunset date has been announced for Semantic Kernel or AutoGen. Agent Framework
Agent Framework + AG-UI Protocol layer for web/mobile clients, supports SSE streaming, backend tool rendering, human approvals, shared/predictive state, and CopilotKit components. Python (1.5.1) is stable; .NET is Preview; Learn states no overall status banner. AG-UI Integration
Copilot Studio + custom code Chosen when you want managed hosting and Power Platform governance rather than owning the runtime and on-call Add tools to custom agents
Foundry routines (GA) Timer, recurring, and GitHub/Teams event triggers on prompt and hosted agents. Listed as GA on Learn (2026-09-09); reminder tool and .NET SDK packages remain preview Automate agents with routines
Logic Apps agentic workflows Event-driven agent loop, 1,400+ connectors. Consumption is explicitly in preview. Standard shows no preview banner on the agent loop, but Microsoft never states Standard is GA and some Standard capabilities are marked preview. Logic Apps agentic workflows

Developer: The Codebase Loop (Bucket 3)

The Trade-off: Speed of the inner loop vs. ownership of the harness. The four rungs below are the developer-loop journey from the Capability Model. Each rung buys asynchrony and costs governance.

Rung Technology Status and limits Documentation
1. Assist in the editor GitHub Copilot agent mode GA in VS Code and Visual Studio Copilot agents in VS Code
2. Delegate the issue Copilot cloud agent (renamed from “coding agent”) GA on all paid plans including Student; NOT available on Copilot Free About Copilot cloud agent
3. Specialize the agent Custom agents (.md + YAML frontmatter) and AGENTS.md, which GitHub calls “agent instructions” Custom agents are GA for the cloud agent, VS Code, and Visual Studio; Public Preview for JetBrains, Eclipse, and Xcode. GitHub’s own caveat on agent instructions: “currently not supported by all Copilot features.” Custom agents
4. Build on the harness GitHub Copilot SDK GA. Python, TypeScript, Go, .NET, Java, Rust. Wraps the Copilot CLI engine over JSON-RPC. BYOK: OpenAI, Microsoft Foundry, Anthropic. Copilot SDK

The seam has a hard edge. Microsoft Foundry Agent Service explicitly lists the GitHub Copilot SDK as a supported framework for Hosted agents, but Foundry Hosted agents support Python and C# only. The SDK ships six language bindings; Foundry hosting accepts two. A Go, Rust, Java, or TypeScript Copilot SDK agent is not directly hostable as a Foundry Hosted agent. Pick the language at rung 4 with rung 5 already in mind.

Business Builder: The Natural-Language Build Path (Bucket 3)

The Trade-off: Speed of the build vs. ownership of the app. The first three rows are alternative build paths, not steps. Cowork and Copilot Studio apps run on Copilot Managed Runtime, and Microsoft says Code’s shared apps use the same foundation; the Run it and Hand it over rows describe that host. The last row is Power Platform’s own path, which hands over by fork instead. The Copilot Studio and Power Apps vibe pages both warn: “Preview features aren’t meant for production use and might have restricted functionality.” How to choose among them is worked through in Scenario 11.

Step Technology Status and limits Documentation
Describe it: just you, or a quick team tool Copilot Code (Code in the Copilot app) Frontier preview on Learn. Every page in the 10-page Copilot Code docset (updated 2026-10-14) states “This is a preview feature” and “You need to be part of the Frontier preview program to get early access.” Separate tab in the Microsoft Copilot app; billed in Copilot Credits under UBB. Learn’s Managed Runtime overview still lists only three creation paths (Cowork, Studio, SDK/CLI). Copilot Code overview | Requirements | USL and UBB
Describe it: a team app over Microsoft 365 content App skill in Copilot Cowork Frontier program only. “Cowork requires a Microsoft Copilot license, but the subscription doesn’t include Cowork usage.” Shared by link. Cowork’s page: “Anyone with the link can open and use the app, including all its data”. The runtime FAQ: sharing “doesn’t grant access to the app’s underlying data”. The pages conflict; test with an account that lacks source access before sharing. Use Copilot Cowork | Admin overview
Build it: one piece of a business process Copilot Studio apps Preview. “Microsoft is gradually rolling out apps in Copilot Studio (Preview). This feature might not be available in your environment yet.” App creation is “powered by the GitHub Copilot harness”, and using, building, previewing, and testing apps can consume Copilot Credits. “The code view is read-only.” Sharing an app “might not automatically share its underlying data”. Apps overview | FAQ
Run it Copilot Managed Runtime Preview. A governed host, not a completed governance review. Microsoft Entra sign-in. Connector governance depends on the environment group: with an advanced connector policy and Advanced connector policies only off, “The most restrictive result applies”; with that setting on and no ACP, “Data policies are ignored. Without an ACP, connectors aren’t restricted.” Sharing limits per environment group. One Apps inventory in the Microsoft 365 admin center, whose Data & tools view “doesn’t provide a complete view of exact destinations, dynamically resolved endpoints, actions that the app executed, or each user’s permissions to the underlying data”. Apps must be TypeScript or JavaScript single-page apps. Running an app takes Copilot Credits or a Power Apps Premium license; with Premium, credits still apply to separately billed services such as Work IQ APIs, or past the plan’s API request limits. Git pushes and code changes are not audited in Purview. Overview | Governance | Security and compliance | FAQ
Hand it over GitHub Copilot with the Managed Runtime plugin, or the ms CLI Preview (SDK, CLI, and plugin). Same Git repository, same runtime. The plugin “also works with Claude Code and other coding agents through Open Plugins”. Repository type is fixed when the app is created: platform-managed Git, or your own repository on GitHub.com or GitHub Enterprise Cloud (GitHub Enterprise Server and Azure DevOps aren’t supported). With your own repository, “your existing branch policies and pull request reviews continue to apply”. SDK overview | GitHub Copilot quickstart
Or describe it on the Power Platform side Power Apps vibe Preview. Builds “requirements, data, and full apps with generated code”. Not available in a default environment; US, Australia, Asia, and India regions only, English only. Its own page says of Cowork and Copilot Studio app building: “We recommend using these experiences moving forward.” Handoff is a fork: “Converting a vibe app creates a new code app. The code app is disconnected from the original plan.” Overview | Convert to a code app

The handoff has a blind spot. The Hand it over row continues the same repository; the Power Apps vibe row creates a separate code app. On the runtime, Purview lists Git pushes and code changes as “Not audited”. Bind the app to your own GitHub repository at creation and your branch policies and pull-request reviews keep applying. For the platform-managed default, the documentation we reviewed doesn’t establish a review gate, so define one on purpose before the handoff.

Data Scientist/Analyst

Technology Description Documentation
Fabric data agents GA core (formerly “AI skill”). Analytics/BI focus, Python SDK, evaluation capabilities, Power BI/semantic models. Requires F2+ or P1+ capacity. Cross-platform consumption routes (Copilot Studio, M365 Copilot, Foundry via Fabric IQ) are Preview, and the older Foundry “Microsoft Fabric” tool retires 2026-11-06. Fabric data agent | Python SDK
Microsoft Foundry (Azure) ML/custom models, full AI/ML pipeline control Microsoft Foundry

Integration Specialist

Technology Description Documentation
Logic Apps agentic workflows Enterprise integration focus, 1,400+ connectors, workflow automation. Consumption agentic workflows are explicitly in preview; Standard has no preview banner on the agent loop, but Microsoft never states Standard is GA. Verify the exact capability you plan to use. Logic Apps Overview | Agentic workflows

Sources: Persona-Based Flow

Data Grounding Decision

Interactive version available. Pan, zoom, and click nodes in the embedded explorer below, or open the full explorer.

View as static Mermaid diagram
%%{init: {'theme':'dark'}}%%
flowchart TD
    Start([Need to ground an agent]) --> Shape{What must the<br/>agent retrieve?}

    Shape -->|Documents and files| Docs{Who asks, and where<br/>does the content live?}
    Shape -->|Records in a<br/>system of record| Rows{Must the answer<br/>be live?}
    Shape -->|Work signals, numbers,<br/>or the live web| Managed[Managed context layers<br/>Work IQ, Fabric IQ, Web IQ<br/>See IQ Layer Selection]

    Docs -->|Employees, content<br/>in Microsoft 365| Inherited[Built-in grounding<br/>SharePoint agent or<br/>Copilot Studio knowledge<br/>Permissions inherited]
    Docs -->|Employees, plus a few<br/>outside sources| External[Copilot connectors<br/>Synced or federated MCP<br/>Source permissions respected]
    Docs -->|Content outside M365,<br/>or customers ask| Implemented{How much content,<br/>how messy?}

    Implemented -->|Under 10,000 files<br/>Simple setup| FileSearch[Agent Service File Search<br/>Up to 10,000 files]
    Implemented -->|Production scale<br/>Your chunking and ACLs| Search[Azure AI Search<br/>Hybrid and vector index]
    Implemented -->|Scans, images,<br/>audio, video| ContentUnderstanding[Content Understanding<br/>with AI Search]

    ContentUnderstanding --> CUPreview[Semantic chunking, image<br/>descriptions, agentic mode<br/><i>Preview</i>]
    Search -->|Shared knowledge base<br/>or query planning| FoundryIQ[Foundry IQ knowledge base<br/>See IQ Layer Selection]

    Rows -->|Yes, current state<br/>of one record| Live[Call the API or MCP tool<br/>No copy to keep fresh]
    Rows -->|No, find similar<br/>records beside the data| VecDB{Which engine<br/>already holds it?}

    VecDB -->|Relational, open source| Postgres[PostgreSQL with pgvector]
    VecDB -->|JSON at global scale| Cosmos[Cosmos DB vector search<br/>flat, quantizedFlat, DiskANN<br/><i>GA not stated</i>]
    VecDB -->|SQL on Azure or Fabric| SQLGA[VECTOR_SEARCH, vector index<br/>Azure SQL, SQL in Fabric,<br/>SQL MI Always-up-to-date]
    VecDB -->|SQL Server 2025| SQLPreview[SQL Server 2025 or MI on<br/>the 2025 update policy<br/>PREVIEW_FEATURES required<br/><i>Preview</i>]

    style Inherited fill:#004578,color:#fff
    style External fill:#004578,color:#fff
    style FileSearch fill:#004578,color:#fff
    style Search fill:#004578,color:#fff
    style ContentUnderstanding fill:#006192,color:#fff
    style CUPreview fill:#8c5e00,color:#fff
    style FoundryIQ fill:#264f78,color:#fff,stroke-dasharray: 5 5
    style Managed fill:#4a3278,color:#fff,stroke-dasharray: 5 5
    style Live fill:#004578,color:#fff
    style Postgres fill:#004578,color:#fff
    style Cosmos fill:#004578,color:#fff
    style SQLGA fill:#004578,color:#fff
    style SQLPreview fill:#8c5e00,color:#fff

How this diagram divides the work. It asks who may see the answer before it asks which store holds the data, and it stops where a managed layer takes over. Work signals, business numbers, and the live web hand off to IQ Layer Selection, whose summary owns their status, billing, and consumer routes so each fact has one home. Choosing the agent platform is a separate decision, made in the Complete Decision Flow, and this diagram no longer re-asks it.

Validation Summary - Data Grounding Decision

Last Validated: October 14, 2026

Read the status column, not the product name. A grounding product can be GA while one feature on its page is Preview, and an integration can be Preview while the product it connects is GA. Where the docs state no GA status, the row says so instead of inferring one.

Documents and Files

Technology Status What decides the choice Documentation
Built-in grounding (SharePoint agent, Copilot Studio knowledge) No status label on the pages checked Permissions are inherited and user-scoped. Copilot Studio allows 500 knowledge sources and 500 uploaded files per agent; each SharePoint or OneDrive source holds up to 1,000 files, 50 folders, and 10 levels; generative orchestration allows 25 SharePoint site URLs. Knowledge answers bill as generative answers in Copilot Credits, with no charge for Microsoft 365 Copilot licensed users. Custom agents that need Microsoft 365 data start from Work IQ, which Microsoft documents as the foundation for new agentic apps on Microsoft 365 data Knowledge sources | Quotas and limits | Billing rates
Copilot connectors (formerly Graph connectors) Synced: GA. Federated: GA per the release notes, see Technologies Connectors reach data beyond Microsoft 365. SharePoint, OneDrive, and Teams content is already in Microsoft 365, so a connector adds cost and nothing else. Synced connectors index into Microsoft Graph; federated connectors fetch live over MCP and store nothing Copilot connectors overview
Foundry Agent Service File Search GA 10,000 files per vector store, 512 MB and 5,000,000 tokens per file, one vector store per agent and per conversation. File search has charges beyond model tokens. Standard setup keeps files in your storage and in AI Search File Search
Azure AI Search GA You own chunking, indexers, skillsets, and ACLs. Vector search uses HNSW or exhaustive KNN; the docs list no IVF AI Search overview | Vector ranking
Azure AI Content Understanding Service GA (API 2025-11-01). AI Search skill GA in the 2026-04-01 REST API The skill calls Content Understanding 2025-11-01. Preview: image descriptions (modelName, modelDeployment), semantic chunking, and token chunk units need Search 2026-05-01-preview; agentic mode needs Content Understanding 2026-06-01-preview What’s new | Content Understanding skill
Technology Status What decides the choice Documentation
PostgreSQL pgvector GA The azure_ai extension adds in-database embeddings and LLM calls. There is no “PostgreSQL agent” product PostgreSQL vector search
Cosmos DB for NoSQL GA date not stated in the docs. Treat the silence as silence flat (exact kNN, up to 505 dimensions); quantizedFlat and diskANN (up to 4,096 dimensions, at least 1,000 vectors, otherwise a full scan runs). Enable the EnableNoSQLVectorSearch capability first. Higher-throughput vector search is Private Preview Cosmos DB vector search
Azure SQL Database, SQL database in Fabric, SQL MI (Always-up-to-date) VECTOR_SEARCH() and vector indexes GA TOP_N is deprecated; use SELECT TOP (N) WITH APPROXIMATE. SQL MI reaches this status only on the Always-up-to-date update policy VECTOR_SEARCH | SQL MI update policy
SQL Server 2025, SQL MI (SQL Server 2025 update policy) VECTOR_SEARCH() and vector indexes Preview Requires the PREVIEW_FEATURES database scoped configuration Vector data type
vector type and VECTOR_DISTANCE() No preview label on any of the four platforms float16 vectors are GA on Azure SQL Database, SQL MI, and SQL database in Fabric, and Preview on SQL Server 2025 VECTOR_DISTANCE

Cost Shape

Path How it bills Where the bill surprises people
Built-in grounding and Copilot connectors Covered for users with a Microsoft 365 Copilot license. Copilot Chat users and custom agents draw Copilot Credits for shared tenant data Tenant graph grounding is 10 Copilot Credits per message. At $0.01 per credit that is $0.10 per grounded message, so the $30 per user per month license used in the Budget diagram breaks even near 300 such messages per user per month (arithmetic from these figures, not a Microsoft statement)
Azure AI Search, Dedicated Per search unit per hour (replicas times partitions). East US list, checked 2026-10-14: Basic $0.10, S1 $0.34, S2 $1.34, S3 $2.69, L1 $3.84, L2 $7.68 per unit-hour, about $73, $248, $978, $1,964, $2,803, and $5,606 per unit per month The SLA needs two replicas for read-only and three or more for read-write, so multiply. Semantic ranker, agentic retrieval, and AI enrichment bill separately
Azure AI Search, Serverless (Preview) Compute Unit hours plus GB-month, with scale to zero. East US list: $0.24 per CU-hour and $0.20 per GB-month. Billing for the Developer tier began 2026-09-13 No SLA, no migration to or from other tiers, and SU-based calculators do not apply
Agent Service File Search Model tokens plus file search charges Prototype cost is low; the file search meter is not in the token bill

Hand-offs to IQ Layer Selection

Work IQ, Fabric IQ, Web IQ, and Foundry IQ knowledge bases are validated once, in IQ Layer Selection, together with Fabric capacity and the Fabric data agent, so their status is not restated here.

MCP Integration (Preview)

Technology Capabilities Documentation
Logic Apps MCP Server Standard logic apps as remote MCP servers, 1,400+ connectors, OAuth 2.0 auth, Streamable HTTP/SSE transports Logic Apps MCP Server | API Center Integration
Solution Use When Cost
File Search (Agent Service) Up to 10,000 files, simple setup (no indexer or chunking configuration), automatic embedding, rapid prototyping Model tokens plus file search charges
Copilot Studio knowledge SharePoint or OneDrive content, up to 1,000 files per source, permissions inherited Copilot Credits as generative answers; no charge for Microsoft 365 Copilot licensed users
Azure AI Search Production scale, custom chunking, analyzers, scoring profiles, faceting, ACL enforcement, customer-facing scenarios Dedicated tiers from about $73 per month for one Basic unit, and the SLA needs two or three units; Serverless (Preview) bills consumption

Sources: Data Grounding Decision

Complexity Assessment Flow

Interactive version available. Pan, zoom, and click nodes below, or open the full explorer.

View as static Mermaid diagram
%%{init: {'theme':'dark'}}%%
flowchart TD
    Start([Assess Use Case]) --> Q1{Does it change a<br/>system of record?}

    Q1 -->|No: answers, drafts, summarizes| Q2{One knowledge domain,<br/>little durable state?}
    Q1 -->|Yes: writes, submits, moves money| Q3{Steps, API calls and<br/>approvals fixed up front?}

    Q2 -->|Yes| T1[Tier 1 Informational<br/>Read-only]
    Q2 -->|No: planning, tool choice, state| Q4{Long-running, multi-agent<br/>or self-correcting?}
    Q3 -->|Yes| T2[Tier 2 Transactional<br/>Deterministic]
    Q3 -->|No: the model chooses| Q4

    Q4 -->|No| T3[Tier 3 Reasoning<br/>Plan and execute]
    Q4 -->|Yes| T4[Tier 4 Autonomous<br/>Bounded, multi-agent]

    T1 --> P1[M365 Copilot or Declarative Agent<br/>Azure AI Search for custom retrieval<br/>Runway: days]
    T2 --> P2[Logic Apps, Power Automate<br/>or Copilot Studio agent flows<br/>Runway: weeks]
    T3 -->|Platform owns the runtime| P3a[Copilot Studio Microsoft Copilot harness<br/>or Foundry Prompt Agent<br/>Runway: weeks]
    T3 -->|Your team owns the runtime| P3b[Agent Framework<br/>or M365 Agents SDK<br/>Runway: months]
    T4 -->|Code-owned orchestration| P4a[Agent Framework<br/>Sequential, Concurrent, Handoff, Magentic<br/>Runway: months]
    T4 -->|Connect specialist agents| P4b[Copilot Studio connected agents<br/>Foundry A2A endpoint<br/><i>Preview</i>]

    P1 --> RefData([More sources change retrieval design,<br/>not the tier. See Data Grounding])
    P1 --> RefGate([Trust boundary and action safety.<br/>See Governance and Compliance])
    P2 --> RefGate
    P3a --> RefGate
    P3b --> RefGate
    P4a --> RefGate
    P4b --> RefGate

    style T1 fill:#0b6a0b,color:#fff
    style T2 fill:#8c5e00,color:#fff
    style T3 fill:#8c5e00,color:#fff
    style T4 fill:#a52617,color:#fff
    style P1 fill:#004578,color:#fff
    style P2 fill:#004578,color:#fff
    style P3a fill:#004578,color:#fff
    style P3b fill:#4b2070,color:#fff
    style P4a fill:#4b2070,color:#fff
    style P4b fill:#004578,color:#fff

Validation Summary: Complexity Assessment Flow

Last Validated: October 14, 2026

What the tree asks. Three behavior questions: does it change a system of record, who fixes the steps, and how long does it run and how many agents are involved. Then it lands on the four tiers of Evaluation Criteria: Complexity. Source count is not a tier input: more sources change retrieval design (Data Grounding). A private-network requirement is a trust-boundary decision (Governance), not a tier.

Tier-to-Path Check

Tier Path in the diagram Status Evidence to demand
1 Informational M365 Copilot or Declarative Agent; Azure AI Search for custom retrieval GA. AI Search Serverless Developer tier is Preview. Foundry IQ is split: agentic retrieval GA via the 2026-04-01 REST API, portal experiences preview-only, Copilot Studio integration GA per the 2026-10-07 blog Citation quality, access trimming, freshness, data boundary, failure behavior
2 Transactional Logic Apps, Power Automate, Copilot Studio agent flows GA. Copilot Studio workflows on the Microsoft Copilot harness carry no page-level preview banner (updated 2026-10-12), while specific nodes such as the Copilot node are Preview Idempotency, connector identity, rollback, approval ownership, audit trail
3 Reasoning Managed: Copilot Studio Microsoft Copilot harness, Foundry Prompt Agent. Code-owned: Agent Framework, M365 Agents SDK Harness GA per the 2026-08-03 announcement (memory, apps, and hooks carry their own Preview labels). Agent Framework: Python (1.21.0) and C#/.NET (1.25.0) GA; Go (0.1.1) is Public Preview Tool accuracy, state recovery, latency budget, evaluation thresholds, runtime ownership
4 Autonomous Agent Framework; Copilot Studio connected agents; Foundry A2A Agent Framework GA core. Connected agents Preview (release feed 2026-10-06; the Foundry side enables the Activity protocol with Foundry-Features: AgentEndpoints=V1Preview). Foundry A2A v1.0 GA; v0.3 Preview (served by default if no version header is sent). Hosted Agents only when the constraint card fits. Foundry Workflows retires 2026-12-01: do not route here Bounded autonomy, isolation, stop conditions, escalation, protocol direction, end-to-end observability

Harness is the Copilot Studio lever. The extension guide points the Copilot chat harness at grounded answers; the standard harness fits “well-defined and rule-based” agents; the Microsoft Copilot harness fits “reasoning-heavy” work. The Customer Conversation harness (Public Preview) adds the customer voice/chat path. The harness also sets the billing model (see Budget & Timeline).

Runway is an anchor, not a clock. The labels say days, weeks, or months. Evaluation Criteria §4 lists what must already be true.

Sources: Complexity Assessment Flow

Budget & Timeline Tradeoffs

Interactive version available. Pan, zoom, and click nodes below, or open the full explorer.

View as static Mermaid diagram
%%{init: {'theme':'dark'}}%%
flowchart TD
    Start([Project Constraints]) --> Q1{How is it<br/>paid for?}

    Q1 -->|Seats| SeatQ{Who needs it?}
    Q1 -->|Copilot Credits| MeterQ{Which surface<br/>meters it?}
    Q1 -->|Azure meters| AzureQ{Workload shape?}
    Q1 -->|Prepaid commitment| CommitQ{What must it<br/>cover?}

    SeatQ -->|Frequent users| S_Full[M365 Copilot USL<br/>USD 30 per user per month<br/>Runway: days]
    SeatQ -->|Occasional users| S_Chat[Copilot Chat, included<br/>Shared tenant data is metered<br/>Runway: days]
    S_Full -->|Advanced work adds a meter| MeterQ

    MeterQ -->|Studio, Microsoft Copilot harness| M_Harness[Credits from first build<br/>Tokens, tools and harness<br/>USD 0.01 per credit]
    MeterQ -->|Studio, standard harness| M_Std[Billed after publish<br/>PAYG, packs or prepurchase<br/>Unused pack credits expire monthly]
    MeterQ -->|Cowork, Code, Autopilot, Work IQ API| M_UBB[Usage-based billing<br/>Copilot Credits<br/>Spending policy per owner]
    MeterQ -->|Apps on Managed Runtime| M_App[Two bills: build and run<br/>0.1 credit per API call<br/>or Power Apps Premium <i>Preview</i>]
    M_UBB --> M_Macc[MACC holder? Pick the billing<br/>subscription under the commitment]

    AzureQ -->|Variable or modest volume| A_Token[Foundry pay-per-token<br/>AI Search Dedicated<br/>Hosting you run]
    AzureQ -->|Managed agent runtime| A_Agent[Foundry Agent Service<br/>Model and tool meters<br/>Hosted adds compute]
    AzureQ -->|Steady high throughput| A_PTU[Foundry PTU<br/>Hourly or reserved<br/>Reservations burn first]
    AzureQ -->|Prototype retrieval only| A_Srch[AI Search Serverless<br/>Developer tier <i>Preview</i><br/>Not for production]

    CommitQ -->|PTU or Fabric capacity| P_Res[Reservations<br/>Applied first]
    CommitQ -->|Copilot Credits only| P_CCCU[Copilot Credit Pre-Purchase Plan<br/>CCCU, applied next]
    CommitQ -->|Credits plus Foundry, Fabric, GitHub| P_ACU[Microsoft Agent Prepurchase Plan<br/>ACU, applied last]
    P_Res --> P_Final([One-year term, auto-renew on,<br/>all purchases final])
    P_CCCU --> P_Final
    P_ACU --> P_Final

    style S_Full fill:#004578,color:#fff
    style S_Chat fill:#004578,color:#fff
    style M_Harness fill:#004578,color:#fff
    style M_Std fill:#004578,color:#fff
    style M_UBB fill:#004578,color:#fff
    style M_App fill:#8c5e00,color:#fff
    style M_Macc fill:#8c5e00,color:#fff
    style A_Token fill:#004578,color:#fff
    style A_Agent fill:#004578,color:#fff
    style A_PTU fill:#a52617,color:#fff
    style A_Srch fill:#8c5e00,color:#fff
    style P_Res fill:#8c5e00,color:#fff
    style P_CCCU fill:#8c5e00,color:#fff
    style P_ACU fill:#8c5e00,color:#fff
    style P_Final fill:#a52617,color:#fff

Validation Summary - Budget & Timeline Tradeoffs

Last Validated: October 14, 2026

The tree asks how the work is paid for, not how big the budget is. Runway is decoupled from budget band (see Diagram 4 and the Runway table below).

Seat Funded (USL)

Solution Cost Notes Documentation
M365 Copilot USD 30 per user per month, paid yearly; qualifying M365 plan required Everyday AI. Licensed users incur no extra charge for extensibility (connectors, agents, plugins). Model selection is part of the USL (some models with fair use, some with limits). Frontier models such as Astra and Fable are UBB Extensibility cost (2026-09-30); USL and UBB (2026-10-06)
Copilot Chat Included with an eligible M365 subscription Instruction-only and public-site agents add no charge. Shared tenant data (SharePoint, Copilot connectors) is metered in Copilot Credits via Copilot Studio. Optional PAYG work-grounded chat Extensibility cost; Copilot PAYG overview (2026-09-03)

Credit Metered (Copilot Credits)

Surface What bills Billing starts Managed in Status and limits
Studio, Microsoft Copilot harness LLM tokens, tools (including knowledge and MCP), and the harness From first build: creating, previewing, testing, evaluating Power Platform admin center Harness GA; upgrade tool Preview
Studio, standard harness Copilot Credits: PAYG, capacity packs, prepurchase (USD 0.01 per credit; USD 200 per 25,000-credit pack) After publish Power Platform admin center Unused pack credits do not carry over. Break-even vs PAYG is 20,000 credits per month
Cowork, Code, Autopilot, Work IQ API, advanced SharePoint/OneDrive, Teams Phone Agent Copilot Credits (UBB) Per use Microsoft 365 admin center, Copilot, Cost management; a billing policy must exist first Cowork, Code, and advanced SharePoint/OneDrive require a Copilot USL; Work IQ API, Managed Runtime, and Teams Phone Agent do not
Managed Runtime apps (Preview) Build: Cowork usage or the Studio harness. Run: per launch plus 0.1 credit per API call, or Power Apps Premium Build and run billed separately M365 admin center (separate runtime policy) Developers running locally need the entitlement too

Three admin surfaces issue Copilot Credits: the Power Platform admin center (Copilot Studio), the Microsoft 365 admin center UBB spending policies (Cowork, Code, Work IQ API, Managed Runtime), and Microsoft 365 PAYG billing policies (Copilot Chat, SharePoint agents). Reconcile all three.

Azure Consumption & Provisioned Throughput

Solution Planning band (framework estimate) Documentation
M365 SDK + Azure $600-2.3K/mo SDK free; Azure hosting (App Service ~$100-300/mo) + Azure OpenAI PAYG (~$500-2K/mo tokens) (docs)
Microsoft Foundry (Azure) Starter $1-5K/mo estimate PAYG tokens + AI Search Basic (~$73/mo per SU) (OpenAI pricing) | (AI Search pricing)
Foundry pay-per-token at scale $5-15K/mo PAYG tokens at scale + AI Search Standard S1 (~$248/mo per SU) (docs)
Foundry + Agent Service $10-30K/mo Managed orchestration PaaS (plus Hosted Agent active-session compute) + AI Search S2 (~$978/mo per SU) (docs)
Foundry PTU + Premium $30K+/mo PTU reservations + AI Search S2/S3 (docs)
AI Search Serverless Developer tier Consumption (Preview) Public Preview; billing began 2026-09-13 (Compute Units/hr + GB/mo). No SLA, not recommended for production, and no migration to or from other tiers (docs)

Runway Estimates

Runway Delivery shape What must already be true
Days Configure an existing governed surface (M365 Copilot, built-in agents, Declarative Agent) Data ready, access approved, no custom runtime
Weeks Compose managed capabilities (Copilot Studio, Logic Apps) Connectors, test environments, approval paths, support ownership exist. On the Microsoft Copilot harness, credits accrue from first build
Months Build and operate a custom runtime (Agent Framework or M365 Agents SDK on Foundry, ACA, or AKS) Landing zone, identity, networking, CI/CD, evaluation, and on-call are funded

Cross-Tier: The Two Prepurchase Plans

The Trade-off: One invoice vs. two different currencies. There is no single “unified” prepurchase. Microsoft ships two distinct Azure Reservations, and buying the wrong one is an expensive way to learn the difference.

One plan covers Copilot Credit usage only; the other spans Foundry, Copilot Studio, Fabric, and GitHub. Both are prepaid, both are non-refundable, and the narrow one always burns first. See Microsoft AI Stack for why that ordering decides what a broad commitment is actually worth.

  Copilot Credit Pre-Purchase Plan Microsoft Agent Prepurchase Plan
Unit CCCU (Copilot Credit CUs) ACU (Agent CUs)
Covers “eligible Copilot Credit usage” “select services across Microsoft Foundry, Microsoft Copilot Studio, Microsoft Fabric, and GitHub costs”, where the Copilot Studio entry also spans Dynamics 365 first-party agents and Copilot
Value 1 CCCU pays down US$1 of qualifying retail cost, purchased at a tiered discount (~5 to 6% in Microsoft’s worked examples) 1 ACU pays down US$1 of qualifying retail cost, purchased at a tiered discount (~5 to 6% in Microsoft’s worked examples)
Docs Copilot Credit Pre-Purchase Plan Microsoft Agent Prepurchase Plan

Both are 1-year terms with auto-renew on by default, and all purchases are final: no cancel, exchange, split, or merge.

Precedence: narrow benefits burn before broad benefits. Microsoft states it plainly: “Reservations always apply before prepurchase plans.” The order is:

  1. Microsoft Foundry PTU Reservations
  2. Microsoft Fabric Capacity Reservations
  3. Copilot Credit Prepurchase Plan
  4. Microsoft Agent Prepurchase Plan: “Applied last to remaining AI usage across all platforms.”

Coverage is now stated for Cowork, not for everything. Learn’s setup page says customers “can use Agent P3 and Copilot P3 credits for AI experiences, such as Cowork,” that P3 is “layered on top of pay-as-you-go,” and that when the subscription includes Agent P3 the system “uses the Agent P3 credits first and bills any overage charges only after those credits are exhausted.” It also says the admin center “displays only the credits available from Copilot Credit P3,” so the dashboard can understate prepaid cover. Code and Copilot Managed Runtime are not named on the prepurchase reservation pages as of 2026-10-14: confirm with your account team. MACC: eligible Copilot usage counts toward a MACC only if the billing subscription chosen in the Microsoft 365 admin center sits under the billing account that holds the commitment (guidance).

See Evaluation Criteria: Budget.

Cost Calculation Notes

  • Seat and meter are separate lines with separate owners; the USL is fixed per user and UBB follows the work
  • Dollar bands are framework estimates; derive from unit rates before quoting
  • Search is not the whole bill; model calls and semantic ranker bill separately
  • PTU (Provisioned Throughput Units) require Azure Reservations for cost optimization

Sources

Governance & Compliance Path

Interactive version available. Pan, zoom, and click nodes below, or open the full explorer.

View as static Mermaid diagram
%%{init: {'theme':'dark'}}%%
flowchart TD
    Start([What must be true<br/>before this agent ships?])
    Q_Risk{Assist or execute?<br/>What breaks if wrong?}
    T1[Tier 1 Low risk<br/>Owner and basic monitoring]
    T2[Tier 2 Medium risk<br/>Owner plus domain validator]
    T3[Tier 3 High risk<br/>Security and RAI review first]
    Q_Bound{Who operates the<br/>trust boundary?}
    B_Tenant[Tenant-managed<br/>Microsoft 365 Copilot]
    B_Conn[Connector-mediated<br/>Copilot Studio, Logic Apps]
    B_Cloud[Customer-operated cloud<br/>Foundry, Container Apps, AKS]
    Q_Resid{Residency or DPA duty<br/>on the model path?}
    W_Anthropic[⚠ Anthropic path<br/>Outside EU Data Boundary]
    Id_Name[Every agent gets an<br/>Entra Agent ID]
    Q_Auth{Whose permissions<br/>does it carry?}
    A_User[On behalf of the caller<br/>Cannot exceed the user]
    A_Fixed[⚠ Fixed human credentials<br/>One login serves every caller]
    A_Own[Agent identity roles<br/>Grant per tool, least privilege]
    Q_Ctl{Which control layer<br/>is missing?}
    C_Fleet[Inventory: Agent 365 and<br/>Foundry Control Plane]
    C_Data[Data protection<br/>Purview DLP, partial for Cowork]
    C_Runtime[Runtime interception<br/>Agent Governance Toolkit<br/><i>Preview</i>]
    C_Shadow[Shadow AI, Local agents<br/>Windows blocking only<br/><i>Preview</i>]
    Q_Net{Network isolation<br/>needed?}
    N_Entra[Not Azure-hosted<br/>Entra network controls]
    N_Public[Public endpoint<br/>Entra ID and RBAC]
    N_PE[Private endpoint in<br/>Egress stays public]
    N_BYO[BYO VNet<br/>Delegated subnet, your firewall]
    N_Managed[Managed VNet<br/>Microsoft runs the network<br/><i>GA not stated</i>]
    N_Tools[⚠ Check tool support<br/>behind the VNet]

    Start --> Q_Risk
    Q_Risk -->|Assists one person<br/>or a small team| T1
    Q_Risk -->|Answers questions or<br/>runs internal services| T2
    Q_Risk -->|Executes changes or<br/>faces customers| T3
    T1 --> Q_Bound
    T2 --> Q_Bound
    T3 --> Q_Bound
    Q_Bound -->|Stays inside the<br/>Microsoft 365 tenant| B_Tenant
    Q_Bound -->|Reaches other systems<br/>through connectors| B_Conn
    Q_Bound -->|Runs in our own<br/>Azure subscription| B_Cloud
    B_Tenant --> Q_Resid
    B_Conn --> Q_Resid
    B_Cloud --> Q_Resid
    Q_Resid -->|Yes| W_Anthropic
    Q_Resid -->|No| Id_Name
    W_Anthropic --> Id_Name
    Id_Name --> Q_Auth
    Q_Auth -->|On behalf of the<br/>invoking user| A_User
    Q_Auth -->|One fixed person,<br/>set at authoring| A_Fixed
    Q_Auth -->|The agent itself,<br/>app-only| A_Own
    A_User --> Q_Ctl
    A_Fixed --> Q_Ctl
    A_Own --> Q_Ctl
    Q_Ctl -->|Cannot list the<br/>agents that exist| C_Fleet
    Q_Ctl -->|Agents read or emit<br/>labeled data| C_Data
    Q_Ctl -->|Must stop a tool call<br/>before it runs| C_Runtime
    Q_Ctl -->|Agents nobody<br/>registered| C_Shadow
    C_Fleet --> Q_Net
    C_Data --> Q_Net
    C_Runtime --> Q_Net
    C_Shadow --> Q_Net
    Q_Net -->|Not Azure-hosted| N_Entra
    Q_Net -->|No| N_Public
    Q_Net -->|Inbound only| N_PE
    Q_Net -->|Full isolation,<br/>we own the network| N_BYO
    Q_Net -->|Full isolation,<br/>Microsoft runs it| N_Managed
    N_BYO --> N_Tools
    N_Managed --> N_Tools

    style T1 fill:#0b6a0b,color:#fff
    style T2 fill:#8c5e00,color:#fff
    style T3 fill:#a52617,color:#fff
    style B_Tenant fill:#004578,color:#fff
    style B_Conn fill:#004578,color:#fff
    style B_Cloud fill:#004578,color:#fff
    style W_Anthropic fill:#a52617,color:#fff
    style Id_Name fill:#004578,color:#fff
    style A_User fill:#0b6a0b,color:#fff
    style A_Fixed fill:#a52617,color:#fff
    style A_Own fill:#004578,color:#fff
    style C_Fleet fill:#004578,color:#fff
    style C_Data fill:#004578,color:#fff
    style C_Runtime fill:#4b2070,color:#fff
    style C_Shadow fill:#8c5e00,color:#fff
    style N_Entra fill:#004578,color:#fff
    style N_Public fill:#004578,color:#fff
    style N_PE fill:#004578,color:#fff
    style N_BYO fill:#004578,color:#fff
    style N_Managed fill:#004578,color:#fff
    style N_Tools fill:#8c5e00,color:#fff

Gates are ANDed. Answer all five in sequence.

Validation Summary: Governance & Compliance Path

Last Validated: October 14, 2026

Five gates, ANDed rather than alternatives. Microsoft’s own order is risk tiers first, then identity, access, and data protection.

Control Status in Microsoft’s words License or cost Watch
Risk tiers (CoE) Guidance None; effort rises per tier Tier 3 is business-critical and external-facing. Move an agent up a tier as its scope or autonomy grows.
Anthropic in Copilot, Researcher, Studio, Power Platform, Copilot in apps Subprocessor; excluded from EU Data Boundary and in-country commitments Included in the product meter On by default outside EU, EFTA, and UK. Cowork switch is tenant-wide. Data-retention models sit outside the DPA.
Claude in Foundry Non-Microsoft Product; Anthropic is seller, operator, and processor Marketplace CCU meter, MACC-eligible Azure-hosted lists Global and US Data Zone only. Anthropic-hosted may process outside the region. GA Claude models run a 12-month lifecycle.
Entra Agent ID Generally available to all Entra customers Platform included; protections need Agent 365 Studio creates one for each new agent since July 2026.
Agent 365 GA 2026-05-01, Commercial segment Per user; in E7; add-on for E5, A5, Business Premium. $15 is blog-stated. Unit is described three ways (per user, “at least one user” to enable, per agent instance in Frontier). Confirm before budgeting. The Oct 1, 2026 FedRAMP High statement is a control assessment, authorization pending.
Registry, basic governance, lifecycle rules, Registry Sync Documented Included with base Microsoft 365 plans Sync covers Bedrock, Vertex AI, Agentforce, Genie, Claude, Oracle, Snowflake. Manual only, read-only, key-based credentials, deletions not reconciled, observability “Coming soon” for four.
Policy templates, observability, tool control, Entra/Purview/Defender/Intune layers Documented Needs E7, G7, or Agent 365 Several are not in GCC.
Foundry Control Plane GA (Nov 2025) AI gateway needed for advanced features Verify per capability.
Purview DLP for Cowork Label exclusion and web-search blocking; prompt blocking is Preview Agent-wide Purview controls need Agent 365 External-email blocking lists Copilot and Copilot Chat. One rule cannot mix SIT and label conditions.
Agent Governance Toolkit Public Preview, no GA features MIT OSS; you operate it Latest formal release is v4.1.0 (2026-06-09). A v5.0.0 tag sits on a temporary publish commit.
Shadow AI, Local agents Preview (Frontier) E5, Intune, Defender for Endpoint; GSA metadata needs Entra Internet Access, Entra Suite, or E7 Blocking: OpenClaw, Gemini CLI, and six VS Code extensions, on managed Windows only. Blocking one Node.js agent blocks the run type.
Network option Egress Inbound Status Cost and operations
Public Public Public or IP-restricted Default None
Private endpoint only Public Private endpoint GA Secures only inbound
BYO VNet Your subnet and firewall Private endpoint GA /27 minimum, /24 recommended for hosted agents. Storage, AI Search, and Cosmos DB platform-managed or yours. Same region as the VNet. Fixed after the first hosted agent.
Managed VNet Microsoft-managed Private endpoint No page-level banner; CLI group Preview; GA not stated 18 regions. FQDN rules create a managed Azure Firewall (cost, not shareable, not BYO). No outbound logging. Mode changes are one-way.

Tool support behind a VNet applies to Responses API agents, not classic-portal agents. Supported: MCP, AI Search, Code Interpreter, Function Calling, File Search, OpenAPI, Azure Functions, A2A (Fabric IQ is partial). Not supported: Fabric Data Agent, Logic Apps, Browser Automation, Computer Use, Image Generation. Bing, web search, and SharePoint grounding use public endpoints.

Sources: Governance & Compliance Path

Multi-Agent Orchestration

Interactive version available. Pan, zoom, and click nodes below, or open the full explorer.

View as static Mermaid diagram
%%{init: {'theme':'dark'}}%%
flowchart TD
    Start([Need more than one agent?])
    Why{What forces a<br/>second agent?}
    Solo[Single agent with tools<br/>and skills. Not multi-agent]
    SoloLink([Event triggers, schedules<br/>See Question 9])
    C_Platform{Where does each<br/>agent run?}
    C_Studio[Copilot Studio<br/>connected agents<br/><i>Preview</i>]
    C_Foundry[Foundry Agent Service<br/>A2A v1.0]
    W_Style{How will you<br/>build it?}
    W_Studio[Copilot Studio<br/>child agents<br/><i>GA not stated</i>]
    W_MAF[Microsoft Agent Framework<br/>Python, .NET GA. Go Preview]
    W_SDK[Need Teams or Copilot reach?<br/>M365 Agents SDK hosts it]
    W_SK[Semantic Kernel or AutoGen<br/>Successor: Agent Framework]
    W_Third[LangGraph and others<br/>Third-party, host anywhere]
    W_Retire([Foundry Workflows today?<br/>See the Migration Calendar])

    Start --> Why
    Why -->|More steps or roles,<br/>or a growth forecast| Solo
    Why -->|Security boundary or<br/>separate teams own it| C_Platform
    Why -->|Measured failure,<br/>tool overload| W_Style
    Solo --> SoloLink
    C_Platform -->|Low-code,<br/>Microsoft 365 reach| C_Studio
    C_Platform -->|Azure,<br/>pro-code| C_Foundry
    W_Style -->|Low-code makers| W_Studio
    W_Style -->|New code-first build| W_MAF
    W_Style -->|Existing SK or<br/>AutoGen code| W_SK
    W_Style -->|Existing LangGraph<br/>or other framework| W_Third
    W_Style -->|Foundry Workflows<br/>today| W_Retire
    W_MAF --> W_SDK

    style Solo fill:#0b6a0b,color:#fff
    style C_Studio fill:#004578,color:#fff
    style C_Foundry fill:#004578,color:#fff
    style W_Studio fill:#004578,color:#fff
    style W_MAF fill:#4b2070,color:#fff
    style W_SDK fill:#4b2070,color:#fff
    style W_SK fill:#4b2070,color:#fff
    style W_Third fill:#8c5e00,color:#fff

Validation Summary: Multi-Agent Orchestration

Last Validated: October 14, 2026

Connected Agents / Sub-Agent Pattern

Platform Direction Status in Microsoft’s words Meter Watch
Copilot Studio, Microsoft Copilot harness Calls Studio, Foundry, remote A2A, and M365 Agents SDK agents Harness announced GA 2026-08-03. The connected-agent page (updated 2026-10-14) states no status. What’s new (2026-10-06): “(Preview) Connect other agents…” Copilot Credits, including build and test Harness GA is not feature GA.
Copilot Studio, standard harness Calls Studio agents, A2A, Foundry, Fabric data agent, Agents SDK A2A GA (April 2026). Foundry, Fabric data agent, and Agents SDK “currently available in public preview” Copilot Credits Split a Studio agent past roughly 30 to 40 choices of action.
Copilot Studio as a server Exposed over MCP or A2A Preview; early-release environments only Copilot Credits OAuth authorization code flow with the caller’s identity.
Foundry Agent Service Incoming endpoint and outgoing A2A tool v1.0 GA; v0.3 Preview and served by default when no version is sent Tokens on both agents; Hosted Agent compute v1.0 is JSON-RPC only; Foundry agent targets are text only with no streaming. .NET SDK is beta. The azd toolbox flow creates the preview tool.
Fabric data agent Consumed by Studio, Copilot, or Foundry Core data agent GA; cross-platform consumption routes are Preview (Foundry “Microsoft Fabric” tool retires 2026-11-06) Fabric F2+ or P1+ Not supported behind Foundry VNet.

Agent Workflow Orchestration

Foundry Workflows is retiring. Microsoft’s wording, quoted exactly:

“Microsoft Foundry is retiring workflows on December 1, 2026. If you’re looking to build new workflows, use Microsoft Agent Framework. To migrate existing workflows, see the Migration guide section of this article for all supported paths.”

Three documented migration paths, each with a different center of gravity:

  • Microsoft Agent Framework: the recommended path; exported YAML is reusable with minimal changes.
  • Azure Logic Apps: keeps a visual designer, aimed at business-process automation.
  • Agent-to-agent (A2A): lightweight handoffs.

Exported YAML definitions remain executable when deployed as Hosted Agents. After 2026-12-01, the visual designer and in-portal execution are unsupported.

Technology Status Orchestration Patterns Documentation
Microsoft Agent Framework GA core (Python 1.21.0, C#/.NET 1.25.0); Go (0.1.1) is Public Preview; A2A packages beta Microsoft’s stated “direct successor” to Semantic Kernel and AutoGen: Sequential, Concurrent, Handoff, Group Chat, Magentic Agent Framework
Semantic Kernel & AutoGen Superseded by Agent Framework; no sunset date announced v1.x supported for the foreseeable future with critical and security fixes; migration guides exist for both SK and AutoGen Semantic Kernel Agents
M365 Agents SDK GA Channel and host SDK; bring your own orchestrator (integrate Agent Framework or third-party) M365 SDK
LangGraph Third-party State graph management for complex workflows Third-party framework

Single-Agent Triggers (Not Multi-Agent Orchestration)

Technology Status Event Handling Documentation
Foundry routines GA (Learn 2026-09-09; blog 2026-09-24) Timer, recurring, and GitHub/Teams event triggers on a single prompt or hosted agent Automate agents with routines
Logic Apps agentic workflows Consumption explicitly Preview; Standard has no preview banner on the agent loop, but Microsoft never states Standard is GA Event triggers + MCP Server (triggers a SINGLE agent, NOT multi-agent orchestration) Agentic workflows
Azure Functions + Agent Service GA Event-driven single agent invocation (event routing, NOT coordination) Agent Service
Event Grid + Foundry GA Event routing to trigger agents independently (NOT orchestration) Azure Event Grid

Sources: Multi-Agent Orchestration

Upgrade Paths

Interactive version available. Pan, zoom, and click nodes below, or open the full explorer.

View as static Mermaid diagram
%%{init: {'theme':'dark'}}%%
flowchart TD
    Start([Start: Low-Code Agent or App]) --> Q0{What are you<br/>upgrading?}

    Q0 -->|An agent| Q1{Need more<br/>capability?}
    Q0 -->|A business-built app| A1{How was the<br/>app built?}

    Q1 -->|No| Stay[Stay on the standard harness<br/>Learn points rule-based,<br/>well-defined agents here]
    Q1 -->|Yes| Q2{What kind of<br/>capability?}

    Q2 -->|Better reasoning, model choice,<br/>skills, memory| Path1[Enhance<br/>Upgrade to Microsoft Copilot harness<br/><i>Preview</i>]
    Q2 -->|Complex orchestration| Path2[Extend<br/>Connect a specialist agent<br/><i>Preview</i>]
    Q2 -->|Custom UI or protocol| Path3[Migrate<br/>M365 Agents SDK or Foundry]

    Path1 --> Warn1[Copy, not transfer, and it uses credits<br/>Not migrated: knowledge sources,<br/>Power Fx, child agents]
    Warn1 --> Result1([Microsoft Copilot harness agent<br/>Model picker built in<br/>Credits from first build])

    Path2 --> Ext1{Which harness runs<br/>the primary agent?}
    Ext1 -->|Standard| Result2a([Studio plus Foundry agent <i>Preview</i><br/>or A2A agent GA])
    Ext1 -->|GitHub Copilot| Result2b([Studio plus Foundry, A2A,<br/>M365 SDK or Studio agents<br/><i>Preview</i>])

    Path3 --> Result3([Full pro-code solution<br/>Azure hosting and engineering])

    A1 -->|Cowork or Copilot Studio app| B1[Copilot Managed Runtime<br/>Git-backed <i>Preview</i>]
    A1 -->|Power Apps vibe| V1[Power Apps vibe<br/><i>Preview</i>]

    B1 --> B2{Outgrown<br/>its builder?}
    B1 --> Warn2[Decide at creation: repo type is fixed.<br/>Git pushes are not audited.<br/>Bind your own GitHub repo]
    B2 -->|No| B3([Stay: name an owner,<br/>budget the run bill])
    B2 -->|Yes| B4[Same repository<br/>GitHub Copilot or ms CLI<br/><i>Preview</i>]

    V1 -->|Convert| V2[Power Apps code app<br/>New app, disconnected from the plan]

    style Stay fill:#004578,color:#fff
    style Path1 fill:#8c5e00,color:#fff
    style Path2 fill:#a52617,color:#fff
    style Path3 fill:#4b2070,color:#fff
    style Warn1 fill:#8c5e00,color:#fff
    style Warn2 fill:#8c5e00,color:#fff
    style B1 fill:#004578,color:#fff
    style B4 fill:#4b2070,color:#fff
    style V1 fill:#8c5e00,color:#fff
    style V2 fill:#004578,color:#fff

Validation Summary: Upgrade Paths

Last Validated: October 14, 2026

Agent Upgrade Mechanisms

Mechanism Status What it does and costs Documentation
Harness upgrade (standard to GitHub Copilot) Preview (prerelease docs, 2026-10-05) Ask Copilot Studio to “upgrade my X agent”. It creates a copy, carries instructions, tools, and suggested prompts, and returns a report of what migrated, needs attention, or was skipped. Custom topics “can be rebuilt as skills or workflows.” Not migrated: knowledge sources, Power Fx expressions, child agents. The upgrade consumes credits, and billing then starts at first build. Agents “can’t be transferred to the standard harness, and vice versa” Upgrade agents | Agents overview
Model choice GA models by region; experimental models are not for production Dropdown on the Microsoft Copilot harness. Many entries are “GA (cross-geo)”, meaning processing outside the region; some are early access. Names are deliberately not in the diagram Model availability and controls
Connected agents (Microsoft Copilot harness) Preview (release feed 2026-10-06; how-to pages carry no banner) Studio, Foundry, A2A, and M365 Agents SDK agents; the primary agent must be on the harness. Foundry: Activity protocol must be enabled (the step sends Foundry-Features: AgentEndpoints=V1Preview), current-portal agents only, Entra user-login connection. A2A: remote agent reachable over the public internet (custom connector for on-premises or VNet). SDK agents: messaging endpoint, no agent card. The orchestrator rewrites the request; handoffs have a payload limit Connected agents overview and the Foundry, A2A, and SDK pages
Foundry agent handoff (standard harness) Preview The older route on the standard harness; outbound A2A is GA on this harness Connect to a Foundry agent (preview)
Shared knowledge Studio ↔ Foundry IQ integration GA per the 2026-10-07 blog (release feed still lists Preview as of 2026-10-06; the connect page carries no label). Azure AI Search GA. Foundry IQ itself is mixed One Foundry IQ connection per agent; one knowledge base can serve many agents; private networking per the blog Connect to Foundry IQ | Blog

App Promotion Mechanisms

Path Status Handoff Documentation
Copilot Managed Runtime Preview Git-backed; the repository is platform-managed or your own GitHub repository, fixed at creation. A developer continues with the GitHub Copilot plugin or ms CLI. Runtime entitlement applies to developers running locally. Purview lists Git pushes as “Not audited”. “Same repository, without forking” is the launch blog’s Cowork example; Learn describes no Studio-specific handoff Overview | SDK
Power Apps vibe Preview. Learn recommends Cowork and Studio app building “moving forward” “Converting a vibe app creates a new code app … disconnected from the original plan” Overview | Convert
Power Apps code apps GA per the official GitHub README; the Learn overview states no status Already code. No Learn page reviewed describes a migration between Managed Runtime and code apps, so do not draw one Overview | README

Sources: Upgrade Paths

Lifecycle Check

Interactive version available. Pan, zoom, and click nodes below, or open the full explorer.

View as static Mermaid diagram
%%{init: {'theme':'dark'}}%%
flowchart TD
    Start([Pre-flight: can we<br/>build on this?])
    Q_Date{Retirement vs go-live<br/>and operating life?}
    Q_Succ{Successor named?}
    Plan[⚠️ Budget the migration now,<br/>or start on the successor]
    Q_Mat{What status does<br/>Microsoft state?}
    Rerun[↺ Re-run this check<br/>on the successor]
    Block[🚫 Block: ask Microsoft,<br/>no new build]
    Q_Prod{Production or<br/>customer-facing?}
    P_Pilot[⚠️ Pilot only<br/>No commitments, set an expiry]
    Q_Exit{Owner signs exit plan<br/>and rewrite budget?}
    P_Go[⚠️ Proceed with caution<br/>Pin versions, track GA]
    P_Wait[🚫 Wait for GA or pick<br/>a GA alternative]
    F_Lab[Lab only<br/>No production data<br/><i>Experimental</i>]
    Q_Fin{Commitment outlives<br/>status or date?}
    Fin_Fix[⚠️ Shorten the term,<br/>exchange or cancel first]
    Go[✅ Proceed to the<br/>Decision Framework]

    Start --> Q_Date
    Q_Date -->|Retired already,<br/>or before go-live| Q_Succ
    Q_Date -->|Inside the<br/>operating life| Plan
    Q_Date -->|None, or after the<br/>operating life| Q_Mat
    Plan --> Q_Mat
    Q_Succ -->|Yes| Rerun
    Q_Succ -->|No| Block
    Q_Mat -->|GA| Q_Fin
    Q_Mat -->|Public Preview,<br/>or not stated| Q_Prod
    Q_Mat -->|Frontier or<br/>experimental| F_Lab
    Q_Mat -->|Deprecated or<br/>superseded| Q_Succ
    Q_Prod -->|No, pilot or<br/>internal demo| P_Pilot
    Q_Prod -->|Yes| Q_Exit
    Q_Exit -->|Yes| P_Go
    Q_Exit -->|No| P_Wait
    P_Go --> Q_Fin
    Q_Fin -->|Yes| Fin_Fix
    Q_Fin -->|No| Go
    Fin_Fix --> Go

    style Rerun fill:#004578,color:#fff
    style Block fill:#a52617,color:#fff
    style P_Pilot fill:#8c5e00,color:#fff
    style P_Go fill:#8c5e00,color:#fff
    style P_Wait fill:#a52617,color:#fff
    style F_Lab fill:#8c5e00,color:#fff
    style Go fill:#0b6a0b,color:#fff

Check the complete schedule of passed and upcoming retirements in Quick Reference: Lifecycle Status.

Validation Summary: Lifecycle Check

Last Validated: October 14, 2026

Microsoft’s word What it means for this check Where you see it
GA Production-safe unless a retirement date falls inside your operating life Learn banners; announcements by the product owner
Public Preview “Without a service-level agreement”; not recommended for production without a signed exit plan Azure and Copilot Studio previews
Frontier Opt-in, default off, preview terms. Agent 365 FAQ: no SLA, removable at any time, not for product evaluations Copilot Frontier setting, Copilot Code, Cowork App skill, Shadow AI, Local agents
Experimental May change or be removed Studio experimental models, Foundry preview models
Not stated Silence. Treat as Preview Studio child agents, harness connected-agent how-to pages, Managed VNet
Deprecated Date announced; migrate Classic agents, standalone Python/.NET azure-ai-agents, Foundry Workflows (2026-12-01), Foundry Fabric tool (2026-11-06)
Retired Date passed; remove dependencies Assistants API (2026-08-26), azure-ai-inference (2026-08-26), GitHub Models (2026-07-30), Bot Framework (2025-12-31)

Models are a lifecycle surface of their own. GA Foundry models retire 18 months after launch (12 months for Anthropic, DeepSeek, Fireworks, and Mistral). Dates are not extendable, Provisioned deployments are not auto-upgraded, and inference returns 410 Gone after retirement. Read the date from the Models API, not from a diagram.

Sources: Lifecycle Check

IQ Layer Selection

Interactive version available. Pan, zoom, and click nodes below, or open the full explorer.

View as static Mermaid diagram
%%{init: {'theme':'dark'}}%%
flowchart TD
    Start([What knowledge<br/>does the agent need?]) --> Domain{What kind of<br/>answer does it need?}

    Domain -->|What does our<br/>policy say?| FIQ[Foundry IQ<br/>Curated documents<br/>Permissions implemented]
    Domain -->|Who decided what,<br/>and when?| WIQ[Work IQ<br/>Mail, meetings, chats<br/>Permissions inherited]
    Domain -->|What do the numbers<br/>and entities say?| FBIQ[Fabric IQ<br/>Numbers and entities<br/>Permissions inherited]
    Domain -->|What is true in the<br/>world right now?| WEBIQ[Web IQ<br/>Fresh public web<br/><i>Limited Access</i>]
    Domain -->|Look up one record| Rows[Not an IQ layer<br/>See Data Grounding Decision]
    Domain -->|More than one kind| Combine[Combine layers<br/>Each layer is its own tool<br/>The meters add up]

    FIQ --> FIQ_GA[GA<br/>Knowledge bases, minimal<br/>extractive retrieval]
    FIQ_GA --> FIQ_Preview[LLM query planning,<br/>remote sources, portals<br/><i>Preview</i>]
    FIQ_Preview --> License1[Azure meters<br/>Search units or CUs,<br/>plus model tokens]

    WIQ --> WIQ_GA[GA<br/>Work IQ APIs, MCP,<br/>Copilot Studio integration]
    WIQ_GA --> WIQ_Preview[retrieve tool, Foundry tool,<br/>Business Applications<br/><i>Preview</i>]
    WIQ_Preview --> License2[Copilot Credits<br/>API access independent<br/>of Copilot licensing]

    FBIQ --> FBIQ_GA[GA<br/>Data agents, semantic models,<br/>Chat, Cowork, Fabric IQ MCP]
    FBIQ_GA --> FBIQ_Preview[Ontology, graph, Ontology MCP,<br/>data agents consumed elsewhere<br/><i>Preview</i>]
    FBIQ_Preview --> FBIQ_Retiring[Foundry Fabric tool<br/>for data agents<br/><i>Retires 2026-11-06</i>]
    FBIQ_Preview --> License3[Fabric capacity F2+<br/>Chat, Cowork: Copilot licensing]

    WEBIQ --> WEBIQ_Alt[No access yet?<br/>Foundry web search tool or<br/>Foundry IQ Web source]
    WEBIQ_Alt --> License4[Limited Access terms<br/>Bing paths are metered]

    style FIQ fill:#264f78,color:#fff
    style WIQ fill:#4a3278,color:#fff
    style FBIQ fill:#783228,color:#fff
    style WEBIQ fill:#0e5a6f,color:#fff
    style Rows fill:#004578,color:#fff,stroke-dasharray: 5 5
    style Combine fill:#1a6b3c,color:#fff
    style FIQ_GA fill:#0b6a0b,color:#fff
    style WIQ_GA fill:#0b6a0b,color:#fff
    style FBIQ_GA fill:#0b6a0b,color:#fff
    style FIQ_Preview fill:#8c5e00,color:#fff
    style WIQ_Preview fill:#8c5e00,color:#fff
    style FBIQ_Preview fill:#8c5e00,color:#fff
    style FBIQ_Retiring fill:#a52617,color:#fff

Key distinction: The four layers are not interchangeable, and the choice follows the shape of the question rather than the brand: what the policy says (Foundry IQ), how the work happened (Work IQ), what the numbers and entities say (Fabric IQ), what the world says right now (Web IQ). Looking up one record is not an IQ layer. Data Grounding Decision covers that case and the engine underneath Foundry IQ.

Validation Summary: IQ Layer Selection

Last Validated: October 14, 2026

Work IQ Studio status refreshed: November 3, 2026, against the Studio integration page, updated October 22. The integration is GA; Business apps workloads remain Preview.

Read the qualifier, not the brand. Every layer has a production path and Preview parts beside it. Plan on the first column, and price the second into the schedule.

Layer Production path today Preview parts to plan around Meter and permissions
Foundry IQ Knowledge bases and agentic retrieval over Azure AI Search through the 2026-04-01 REST API, with minimal extractive retrieval and the knowledge source types not marked preview: search index, Azure Blob, OneLake, Web 2026-08-01-preview: LLM query planning, answer synthesis, non-minimal reasoning effort, multi-turn. Remote SharePoint, Work IQ, Fabric data agent, Fabric ontology, and MCP server sources; indexed SharePoint, Azure SQL, and File sources. The Foundry and Azure portals are preview-only for all agentic retrieval features Azure AI Search plus Azure OpenAI. The free search tier and a free agentic retrieval token allocation cover a proof of concept. Permissions are implemented: ACL sync, Purview labels, Entra identity at query time; the remote SharePoint source trims at query time
Work IQ APIs (Chat, Context, Tools, Workspaces) GA 2026-06-16 per Microsoft’s announcement; the Learn API pages carry no status label. Work IQ MCP is GA per the Agent 365 tooling page (updated 2026-10-13); Work IQ in Copilot Studio is GA (updated 2026-10-22) The MCP retrieve tool; Work IQ as a Foundry tool; Business Applications in Work IQ, on by default for new environments from 2026-09-30 and for existing environments from November 2026 (EEA, EU, and government exceptions). Dataverse intelligence is folding into it (“will no longer be available soon”, no date). Do not start on the per-workload MCP servers (Mail, Calendar, Teams, SharePoint, OneDrive, User, Word, Copilot): they are legacy, kept for backward compatibility Copilot Credits, usage-based: a fixed component for Tools and a variable one for Chat and Context. API access is independent of Microsoft 365 Copilot licensing (Work IQ overview, 2026-09-25). Inside Microsoft 365 Copilot experiences the license covers licensed users; custom and third-party agents are usage-billed. The Foundry quickstart (2026-07-01) still lists a Copilot license per calling user, so the pages conflict: verify your surface. Entra delegated auth only
Fabric IQ Fabric data agent (needs F2+ or P1+), Power BI semantic models, Power BI answers in Microsoft 365 Copilot Chat, the Fabric IQ plugin in Copilot Cowork, and Fabric IQ MCP over Power BI reports and semantic models The IQ workload, ontology, Graph, and Ontology MCP. Every documented route that consumes a data agent from another platform is Preview: Copilot Studio, Microsoft Copilot, and Foundry through the Fabric IQ (OneLake Catalog) tool; responses can leave the Fabric compliance boundary or region. Retiring: the Foundry Microsoft Fabric tool for data agents stops being supported on 2026-11-06 Fabric capacity: F2 is 2 CUs, about $263 per month if left on (East US list). Chat and Cowork need Microsoft 365 Copilot licensing instead; Cowork adds usage-based billing. Permissions: Fabric roles, OneLake security, row-level security
Web IQ None. Limited Access (waitlist; access is subject to intended use, Azure account history, and other terms) Not applicable No published meter found. Without access: the Foundry web search tool (Grounding with Bing Search and Custom Search; the page states no GA or preview status; the Data Protection Addendum does not apply and data leaves compliance and geographic boundaries) or the Web knowledge source in Foundry IQ

Sources: IQ Layer Selection

Using This Guide

How to navigate: Start with Complete Decision Flow for end-to-end guidance, or jump to any of the nine specialist trees: Persona-Based Flow (by role and ownership), Data Grounding (trust boundary and store), Complexity Assessment (four behavior tiers), Budget & Timeline (funding lanes), Governance Path (five ANDed gates), Multi-Agent (Soloist vs. Orchestra), Upgrade Paths (agents and business-built apps), Lifecycle Check (pre-flight readiness), or IQ Layer Selection (Foundry IQ, Work IQ, Fabric IQ, Web IQ).

Diagram conventions: Diamonds = decisions, Rectangles = technologies, Circles = start/end. Colors: Blue = Microsoft core, Purple = developer tools, Green = low complexity/safe path, Orange = medium complexity/caution, Red = high complexity/enterprise/warning.

Last Updated: October 14, 2026 Next: Quick Reference - Fast lookup tables and the canonical Migration Calendar

Next Steps

Fast lookup tables: → Quick Reference

Detailed comparisons: → Feature Comparison

Real-world examples: → Scenarios

Evaluate readiness: → Evaluation Criteria

Back to top

Copyright © 2025. This documentation is based on official Microsoft sources and best practices.

This site uses Just the Docs, a documentation theme for Jekyll.