Quick Reference: Technology by Need

Important: Use This as a Validation Tool, Not a Shortcut

These tables help you validate technology choices after you’ve completed the Decision Framework assessment. Selecting technologies without understanding your Business requirements, user eXperience needs, and Technology constraints often leads to suboptimal outcomes: either overpowered solutions that exceed budget, or underpowered solutions requiring costly rewrites.

First time here? Start with Decision Framework to build a systematic approach to technology selection.

Table of contents

  1. CAF Agent Adoption Quick Cues
  2. Lifecycle Status Quick Reference
  3. Availability updates
  4. Publish Target Quick Reference
  5. Technology by User Experience
  6. Agentic Retrieval Quick Facts
  7. Agent Development Approach Comparison
  8. Custom Engine Agent Tool Comparison
  9. Data Grounding Pattern by Source
  10. Developer Loop Quick Reference
  11. Builder Loop Quick Reference
  12. Data Plane Quick Reference
  13. Memory & Analytics by Technology
  14. Orchestration Complexity Decision Matrix
  15. Budget & Timeline Quick Guide
  16. Governance Decision Quick Reference

CAF Agent Adoption Quick Cues

  • Phases in one line: Plan for agents (business + tech + org + data), Govern & secure agents (responsible AI, controls, environment prep), Build agents (single vs multi, orchestrate, secure process), Manage agents (integrate, monitor, retire). CAF AI agent adoption.
  • When not to use an agent: Highly deterministic workflows or static Q&A/content generation → use code or plain RAG. Business plan.
  • SaaS vs build: Use SaaS agents when they meet requirements; if not, prototype on Foundry or Copilot Studio before committing to custom builds. Technology plan.
  • Single vs multi: Start single; go multi-agent only for hard security/compliance boundaries, distinct owning teams, or planned modular growth. Single vs multiple.
  • Prioritize use cases fast: Score business impact, feasibility, desirability; pick high-impact, low-friction pilots first.
  • Org roles: Platform owns guardrails, workload teams own use cases and data, AI CoE advises and standardizes. Organizational readiness.
  • Build safely: Use workflows for deterministic control, treat instructions as versioned config, gate tool calls, isolate memory per role/tenant, and run evaluations before production. Build process.
  • Operate with evidence: Keep an agent inventory/identity, centralize logging, track cost and quotas, red team regularly, retire unused agents. Manage agents.

CAF agent decision tree CAF decision tree: SaaS first, then build; validate single vs multi during planning.

How to prioritize agent use cases Use impact × feasibility × desirability to rank pilots.

Ground agents in governed data. Anchor retrieval in OneLake, Fabric, and Foundry rather than in copies, and give each workload a clear landing zone. See Grounded Q&A over enterprise content for the retrieval decision, and CAF: AI strategy for the data-strategy stage.

Sources (CAF):

Lifecycle Status Quick Reference

Active migration deadlines and forced transitions. If your project timeline crosses a deadline below, migration planning is mandatory.

Technology / API Current Status Deadline Successor Action
azure-ai-inference SDK Deprecated May 30, 2026 openai SDK (standard OpenAI() client) Migrate now
Assistants API Deprecated Aug 26, 2026 (date unconfirmed against current Microsoft Learn, verify before planning) Foundry Agents Service (Responses API) Plan migration. migration tool available
Classic agents (v1) Deprecated Mar 31, 2027 (date unconfirmed against current Microsoft Learn, verify before planning) Foundry Agents v2 (create_version()) Plan migration. migration guide
Foundry Workflows Retiring from Preview without a GA path on December 1, 2026 Dec 1, 2026 Role-dependent: Agent Framework, Logic Apps, A2A, or exported YAML on Hosted Agents (mixed maturity) Do not start new solutions. Map code-first orchestration, visual process, direct delegation, and hosted YAML separately
Classic Foundry portal Legacy Ongoing Foundry (new) portal Transition. Classic and new lack feature parity
Bot Framework Retired Dec 31, 2025 (passed) M365 Agents SDK + Toolkit Complete
azure-ai-agents SDK Deprecated March 2026 AIProjectClient in azure-ai-projects 2.x Remove standalone azure-ai-agents pin; use project_client.get_openai_client() for agent responses
azure-ai-projects 1.x Legacy Aligns with classic portal azure-ai-projects 2.x (GA: Python 2.0.1, Java 2.0.0; Beta: .NET 2.0.0-beta.1) Upgrade. 2.x targets the new Foundry portal and is the investment direction
Semantic Kernel (standalone) Maintenance Security patches only Microsoft Agent Framework Migrate for new projects

Sources:

Availability updates

Rollout timing is not a migration deadline. Use these labels to decide what you can evaluate and what you can put on a production plan.

Learn records the rename: “Microsoft 365 Copilot is now named Microsoft Copilot, and Microsoft 365 Copilot Chat is now named Microsoft Copilot Chat.” The work product now shares its name with the consumer Microsoft Copilot app. Confirm which experience you mean (requirements, updated 2026-09-24).

Technology Status Rollout timing Related path Action
Autopilot, formerly Microsoft Scout Private preview, per Jared Spataro’s Official Microsoft Blog post “expanding to private preview at the end of the month” (September 2026) Autopilot Confirm Autopilot’s deployment requirements. Scout’s documented desktop prerequisites do not establish them
Copilot Managed Runtime Preview on Learn; “public preview” in the announcement; “Paid Public Preview” in the Credits Guide Announced 2026-09-25 Runtime reference Pilot only. Confirm effective policies and keep code-review evidence; Git pushes are not audited
Cowork App skill Preview, Frontier only No rollout completion date stated as of 2026-09-25 Builder Loop Pilot with eligible Frontier users; test sharing and source access separately
Copilot Studio apps Preview, rolling out gradually Learn says the feature might not be available in your environment yet Builder Loop Check your environment and admin settings; code view is read-only
Copilot Code Frontier rollout announced; Learn documents UBB, but no Learn feature overview was found as of 2026-09-25 The same Official Microsoft Blog post gives “at the end of the month” and “in the coming weeks”; the runtime post calls entry points “already live” Full rollout statements Confirm access and spending policy for your cohort; do not assume broad availability
Power Apps vibe Preview; its overview recommends Cowork and Studio app building “moving forward” No retirement date stated in the overview as of 2026-09-25 Copilot Cowork or Copilot Studio app building Weigh the recommendation before new work. Conversion creates a separate code app
Foundry routines GA per the Foundry blog. Learn states no status as of 2026-09-25 and points .NET users to preview packages GA announcement: 2026-09-24 Foundry Agent Service Use for timer, recurring, and event triggers; the reminder tool remains preview
Foundry hosted-agent network egress controls Preview on Learn. The blog says “not GA, have no preview SLA, and are not intended for production use” Announced 2026-09-24 Hosted-agent guardrails Test only; network modes are Audit and Enforce (API value Enforced), not Blocking

Sources: New Copilot announcement; runtime details and sources; Foundry routines announcement and Learn guidance; Network egress controls (preview) and announcement. App-building sources are below.

Publish Target Quick Reference

Where can your agent appear? Match build platform to distribution surface.

Agent Built In Can Publish To
Foundry Agent Service M365 Copilot (one-click), Teams, custom app (Responses API), Agent 365 digital worker (Frontier)
Copilot Studio M365 Copilot, Teams, web, WhatsApp, custom channels, messaging channels via Azure Communications
M365 Agent Builder M365 Copilot (copy to Copilot Studio for deeper customization)
M365 Agents SDK M365 Copilot, Teams, web, email, SMS, custom channels (10+ via channel adapters)
Agent Framework Custom apps, web APIs, AG-UI protocol (Preview); surface via M365 Agents SDK for Copilot/Teams

Sources:

Technology by User Experience

Where Users Interact Recommended Technologies Use When
Microsoft 365 Apps Free Microsoft 365 Copilot Chat (included) + Copilot connectors (Graph connectors) for baseline pilots; Microsoft 365 Copilot add-on + declarative agents for work-grounded copilots; Frontier Word/Excel/PowerPoint creation agents (Preview) require admin Frontier opt-in1 and Anthropic data-sharing consent; mobile parity for custom engine/message-extension agents (iOS/Android)2; Copy to Copilot Studio copies data sources/actions but GPTs/custom actions must be reattached3 Need managed copilots embedded in Word, Excel, Outlook, or Teams with tenant-level governance. Start with the free chat surface and graduate to the add-on when Graph grounding or in-app assistants are required; use Frontier creation agents only for controlled pilots
Always-On Personal Agent (private preview) Autopilot, formerly Microsoft Scout: a cloud-hosted personal agent in Jared Spataro’s 2026-09-25 Official Microsoft Blog post. Scout’s desktop setup requirements apply to Scout; their applicability to Autopilot is not established as of 2026-09-254 Controlled exploration of persistent coordination across work surfaces, after confirming deployment and identity requirements
Microsoft Teams Only Teams SDK, Copilot Studio, M365 Agents SDK Teams-centric chat, channels, meetings, or calling scenarios where admins may enforce “only during the call” retention
Custom Web/Mobile App Microsoft Foundry, Microsoft Foundry Agent Service (Standard setup) Building standalone applications while keeping files, search, and thread storage in customer-owned Azure resources
Governance / Registry Microsoft Agent 365 (GA 2026-05-01; per-user, E5 recommended prerequisite, included in M365 E7; Agent 365 SDK and CLI remain Preview)56; Foundry Control Plane; M365 Agent Registry7; Agent Governance Toolkit v4.1.0 (Public Preview OSS middleware; zero GA features) Compose fleet governance with application-layer enforcement; AGT is not a managed registry or service
Custom Web/Mobile UI with streaming Microsoft Agent Framework + AG-UI protocol (Preview) Need Server-Sent Events streaming, backend tool rendering, shared state, and human approvals in bespoke front-ends
Multiple Channels M365 Agents SDK Deliver one agent across Microsoft 365 Copilot, Teams, web, email, SMS, and other channels
Power Platform Copilot Studio, AI Builder Integrated with low-code Power Apps/Power Automate workloads
Enterprise Workflows Azure Logic Apps agentic workflows with an agent loop (Consumption explicitly in preview; Standard carries no preview banner on the agent loop, but Microsoft never states Standard is GA and some Standard capabilities are marked preview; check the exact capability), MCP Server Workflow automation that needs autonomous/conversational agent patterns with Easy Auth guardrails
Data Grounding / RAG Microsoft IQ: Foundry IQ (mixed GA/Preview), Work IQ (APIs GA; Work IQ MCP is Preview), Fabric IQ (Preview workload), Web IQ (Limited Access); Copilot connectors (M365); Microsoft 365 Copilot Search API (Preview) for OneDrive hybrid semantic+lexical search8 Match the grounding layer to the data domain: enterprise documents (Foundry IQ), work context (Work IQ), business semantics (Fabric IQ), public web (Web IQ), tenant-scoped content (Connectors), or hybrid OneDrive search (Search API)
Developer Tools GitHub Copilot in the IDE; Copilot cloud agent (async, renamed from “coding agent”); MCP servers IDE and development workflow integration. Note: GitHub App-based Copilot Extensions were deprecated 2025-11-10 with MCP servers as the migration target; client-side VS Code Copilot Extensions remain supported

Sources:

Confidence Level: High for placement guidance; status varies by capability and should be verified in the linked Microsoft Learn pages before production planning.

Agentic Retrieval Quick Facts

  • Knowledge agents are now knowledge bases. Agentic retrieval is GA via the 2026-04-01 REST API; portal experiences remain preview-only, as do answer synthesis, multi-turn retrieval, and newer source types.
  • Knowledge sources (Preview): indexed SharePoint, remote SharePoint (Copilot Retrieval API, ACL-trimmed), indexed OneLake, web/Bing, search index, Azure Blob; ingestionParameters wraps embeddings/chat models/Content Understanding; portal creates 2025-08-01-preview objects. Migrate to 2025-11-01-preview.
  • Semantic ranker is available on free tier (quota limits); enable on the service before using KBs.
  • Hybrid/vector preview (2024-09-01-preview): MRL truncationDimension, filterOverride for vector-only filters, debug subscores for RRF, token-based Text Split parameters.
  • Content Understanding skill (Preview) replaces Text Split for richer chunking; billed to Foundry resource when used via contentExtractionMode.

Sources:

Agent Development Approach Comparison

Approach Declarative Agents Custom Engine Agents
Definition Microsoft-managed orchestration where you supply instructions, knowledge, actions Bring your own orchestration, models, and tooling for bespoke agents
Best For Rapid delivery of guided experiences in M365 apps Advanced workflows, multi-agent patterns, or non-M365 channels
Development Model Low-code (Copilot Studio) or pro-code via Agents Toolkit scaffolding Pro-code using Agents SDK, Teams SDK, or custom frameworks
Orchestration Microsoft 365 Copilot orchestrator handles planning and grounding You decide orchestration (Semantic Kernel, LangChain, Teams AI action planner, etc.)
M365 Integration Native to Microsoft 365 Copilot UI, Teams, and SharePoint Requires explicit integration via M365 Agents SDK or Teams SDK
Typical Timeline Days to a few weeks Weeks to months
Skill Level Makers or full-stack developers Professional developers
Availability GA GA

Sources:

Confidence Level: High (official Microsoft guidance)

Custom Engine Agent Tool Comparison

Teams SDK and M365 Agents SDK coexist with differentiated fit: Teams SDK for Teams-native collaborative experiences, M365 Agents SDK for broader multi-channel distribution and custom orchestration portability.

Tool / Mode Copilot Studio (standard harness) Copilot Studio (GitHub Copilot harness) Teams SDK M365 Agents SDK
Primary Use Case Structured conversational flows with deterministic branch control Reasoning-heavy agents and workflows that complete multi-step business processes Collaborative agents inside Teams Pro-code agents running across M365 and third-party channels
Orchestration Topics plus generative orchestration Enhanced orchestration runtime reasoning over instructions, knowledge, tools, memory, and skills Built-in Teams AI action planner Bring your own orchestration (Semantic Kernel, LangChain, custom)
Supported Channels Microsoft 365 Copilot, Teams, partner apps, mobile apps, custom websites Microsoft 365 Copilot, Teams, partner apps, mobile apps, custom websites Microsoft 365 Copilot, Teams Microsoft 365 Copilot, Teams, web, email, SMS, Office add-ins, custom sites
Development Experience Low-code UI with Power Platform controls Low-code UI with component authoring, native file creation, and evaluation loops Visual Studio/VS Code libraries for C#, TS/JS, Python Agents Toolkit scaffolding for .NET/JS with multi-channel deployment
Ideal Team Makers or fusion dev teams preferring explicit flow control Makers or fusion teams needing adaptive reasoning with managed platform controls Teams-focused pro dev squads Professional developers delivering enterprise-scale agents
Classic Topics Support Yes No (use skills, instructions, and tools instead) N/A N/A
Billing starts After publish When you start building Per your hosting Per your hosting
Status GA GA (2026-08-03) GA GA

The harness is chosen at creation and an agent cannot move between harnesses. Microsoft states that agents created on the GitHub Copilot harness “can’t be transferred to the standard harness, and vice versa.” Reuse works best within a harness: on the GitHub Copilot harness, skills export as Markdown packages and import into other agents there. Skills are specific to that harness, so what carries across all three is more basic, namely instructions, knowledge sources, and connectors. See Technologies for the full comparison.

Feature maturity can change quickly for Studio orchestration components and channel capabilities. Reconfirm per-feature GA/Preview in current Microsoft Learn release notes before rollout.

Sources:

Confidence Level: High (all GA, official Microsoft documentation)

Data Grounding Pattern by Source

Data Source Type Recommended Approach Technologies
SharePoint/OneDrive Copilot connectors (Graph connectors) Microsoft Graph Connectors SDK, Copilot Studio connectors
External Structured Data Copilot connectors (inside M365) or Azure AI Search (Microsoft Foundry Agent Service standard setup) Copilot connectors, Azure AI Search
Unstructured Documents Vector search with chunking Azure AI Search, Azure OpenAI Embeddings
Real-Time Transactional Data API-based grounding API plugins, Functions, Logic Apps agent workflows
Multimodal Content Azure AI Content Understanding (Preview) Process documents, images, audio, video with reasoning
Database Vectors AI-capable databases with managed embeddings Azure Cosmos DB for NoSQL (flat, quantized flat, DiskANN; standard setup: 3 × 1000 RU containers; no GA date stated in docs), Azure Database for PostgreSQL (pgvector GA, plus azure_ai for in-database embeddings), Azure SQL Database / SQL MI / SQL Server 2025 / SQL database in Fabric (VECTOR_SEARCH is Preview on all three platforms)
Microsoft Fabric Platform Direct data access Lakehouse (Delta tables), Warehouse (T-SQL), OneLake (ADLS Gen2 APIs), SQL analytics endpoint
Microsoft Fabric via Agent Conversational data layer Fabric data agent (GA, formerly “AI skill”; requires F2+ or P1+ capacity) with Copilot Studio or Microsoft Foundry Agent Service

Sources:

Confidence Level: High for GA technologies (Fabric data agent, pgvector); Medium for Preview capabilities (VECTOR_SEARCH, Content Understanding)

Developer Loop Quick Reference

One agent, four handoffs. This is the relay race a developer’s work actually runs: from the keyboard, to the cloud, to a governed runtime. Each leg has a different baton.

Leg Surface Status What it is for
1. Inner loop GitHub Copilot in the IDE GA Synchronous, developer-in-the-seat editing and chat
2. Async loop Copilot cloud agent (renamed from “coding agent”) GA on paid plans including Student; not available on Copilot Free Hand off an issue or prompt; the agent works on a branch and opens a pull request
3. Behavior as config Custom agents (.md with YAML front matter) + AGENTS.md agent instructions Custom agents GA for cloud agent, VS Code, and Visual Studio; Public Preview for JetBrains/Eclipse/Xcode Version your agent’s personality and rules alongside the code. GitHub’s own caveat: agent instructions are “currently not supported by all Copilot features”
4. Embed it in your product GitHub Copilot SDK GA: Python, TypeScript, Go, .NET, Java, Rust Wraps the Copilot CLI engine over JSON-RPC. BYOK to OpenAI, Microsoft Foundry, or Anthropic; Entra/managed identity works via bearerTokenProvider composition with the Azure Identity SDK
5. Run it under governance Microsoft Foundry Hosted Agents See the Hosted Agent constraint card Managed endpoint, scaling, identity, and observability for your own agent container

The seam that bites: Foundry Hosted agents support Python and C# only.

Microsoft Foundry does list the GitHub Copilot SDK among the frameworks you can bring to a Hosted agent. But the hosted-agent runtime states plainly: “Hosted agents support Python and C#.” So a Go, Rust, or Java Copilot SDK agent is not directly hostable on Foundry Hosted Agents. You host it yourself (Azure Container Apps, AKS, or your own runtime) and integrate via the Responses API. Choose your SDK language with the hosting endgame in mind, not the other way round.

Sources:

Builder Loop Quick Reference

Name the owner before choosing the builder. For apps on Copilot Managed Runtime, the host supplies platform controls; your team still owns the behavior. The governance, connection, and sharing details below describe that host. Power Apps vibe uses Power Platform controls and hands off by conversion to a new code app.

Leg Surface Status What it is for
1. Set ownership and governance Apps in the Microsoft 365 admin center; environment groups and connector policies Preview Name the business owner and technical reviewer. Confirm effective access, sharing, and spending policies. The Advanced connector policies only setting can bypass classic data policies; inspect it before connecting data
2. Describe Copilot Code; Cowork App skill; Copilot Studio apps; Power Apps vibe Code: announced Frontier rollout with differing timing statements; see Availability updates. App skill: Preview, Frontier only. Studio apps: Preview, rolling out gradually. Power Apps vibe: Preview, restricted regions and English only Turn the requirement into a draft, then test it. Studio’s code view is read-only. Power Apps vibe’s overview recommends Cowork and Studio app building moving forward
3. Connect Approved connectors and MCP servers Preview runtime. By default, 18 Microsoft first-party connectors and 12 Microsoft MCP servers are available, including Work IQ MCP servers. Admins can widen access toward the 1,500+ connectors the SDK can call. Work IQ MCP is Preview Put approved data behind the app. Check each recipient’s source permissions and connections; catalog size does not tell you what the tenant allows
4. Share managedapps.cloud.microsoft; share links or direct sharing with people and groups Preview Publish only after validating results, permissions, failure handling, and support ownership. Users find the apps they can access in one portal
5. Hand over A developer using the repository, coding-agent plugin, or ms CLI; Power Apps vibe conversion creates a new code app Managed Runtime Preview; Power Apps vibe Preview Hand over the repository with the requirements and tests. The blog’s Cowork example continues “without forking or standing up a separate runtime, platform resources, and governance stack”. Power Apps vibe forks instead. Neither path removes the need for review

Sharing is a release decision. Learn warns: “Preview features aren’t meant for production use and might have restricted functionality.”

Cowork says “Anyone with the link can open and use the app, including all its data”. The runtime FAQ says sharing “doesn’t grant access to the app’s underlying data”. The Studio FAQ says “Sharing an app might not automatically share its underlying data.” These descriptions remain unreconciled as of 2026-09-25. Test app data and connected data with an account that lacks source access before sharing real data.

Purview marks “Git push / code change” as “Not audited”. The Data & tools inventory is not a complete view of destinations, actions, or permissions. Keep code-review evidence separately. Known-answer cases are one test layer; business owners validate rules and results, while technical reviewers check risks those owners cannot assess.

Sources:

Data Plane Quick Reference

Grounding is not one product. It is a shelf of stores with different shapes, different guarantees, and, critically, different maturity labels.

Store Retrieval capability Status Pick it when
Azure AI Search (name unchanged) Agentic retrieval, hybrid, vector, semantic ranking Agentic retrieval GA via the 2026-04-01 REST API; portal experiences remain preview-only You need governed enterprise retrieval with ACL/label enforcement; it is the engine underneath Foundry IQ
Azure Database for PostgreSQL pgvector similarity search; azure_ai extension for in-database embeddings and LLM calls pgvector GA Your operational data already lives in Postgres and you want retrieval next to it. Note: there is no “PostgreSQL agent” product
Azure Cosmos DB Vector search: flat/kNN, quantized flat, DiskANN; must be enabled as a feature No GA date stated in the docs; ultra-high-throughput vector search is Private Preview Globally distributed transactional data that also needs vector lookup
Azure SQL / SQL MI / SQL Server 2025 / SQL database in Fabric VECTOR_SEARCH() over the native vector type VECTOR_SEARCH is Preview on all three platforms (SQL Server 2025 also needs the PREVIEW_FEATURES database-scoped configuration; TOP_N is deprecated) Your system of record is SQL and you want retrieval without a second datastore, with eyes open about preview status
Microsoft Fabric / OneLake Fabric data agent over lakehouse, warehouse, KQL, semantic models Fabric data agent GA (formerly “AI skill”), F2+/P1+ capacity; Fabric IQ is a Preview workload The question is analytical (“what happened, and why”) rather than document lookup
Azure Managed Redis Vector and cache-side retrieval GA, Entra-native, built on Redis Enterprise You need low-latency semantic caching or session state beside the agent

Infrastructure deadline worth putting on the calendar: Azure Cache for Redis Enterprise and Enterprise Flash retire 2027-03-31 (disabled 2027-04-01); Basic, Standard, and Premium retire 2028-09-30 (disabled 2028-10-01). Azure Managed Redis is the forward path.

Sources:

Memory & Analytics by Technology

Common Customer Confusion: Grounding (RAG) ≠ Memory ≠ Analytics

Technology 📋 Grounding (RAG) 💾 Memory / Thread Storage 📊 Analytics / Transcripts Admin Access Retention Control
M365 Copilot ✅ M365 content per request ⚠️ User-level memory9; user history lives in mailbox/Graph ⚠️ Copilot activity history, Teams call summaries if policy allows ⚠️ Purview/eDiscovery only ✅ Purview retention policies; Teams calling policy can force “only during the call”
Copilot Studio ✅ M365 data, Dataverse, connectors ⚠️ Dataverse variables per conversation ✅ Session reports, transcript downloads, ROI analytics ✅ Transcript Viewer role required ✅ Toggle Dataverse save, default 30-day retention, bulk delete
Microsoft Foundry Agent Service ✅ Azure AI Search, Cosmos DB, Fabric, tools ✅ BYO thread storage in Cosmos DB (standard setup) ⚠️ Custom telemetry (App Insights, OpenTelemetry) ✅ Customer RBAC on Azure resources ✅ Customer deletes threads/files in own storage
M365 Agents SDK ✅ Custom (developer implements) ⚠️ Custom (developer implements thread storage) ⚠️ Custom (Application Insights, custom logging) ⚠️ Custom (developer implements) ⚠️ Custom (developer implements)

Key Compliance Questions:

  • Where is conversation history stored? → M365 Copilot (user mailbox/Purview, plus user-level memory), Copilot Studio (Dataverse), Microsoft Foundry Agent Service (customer Cosmos DB), SDK (customer datastore)
  • How long is it retained? → M365 Copilot (Purview retention or user deletion), Copilot Studio (policy configurable, 30-day default), Microsoft Foundry Agent Service (customer-defined lifecycle), SDK (custom)
  • Who can query chat logs? → M365 Copilot (admins via eDiscovery, users see activity history), Copilot Studio (admins with Transcript Viewer role), Microsoft Foundry Agent Service (Azure RBAC), SDK (custom controls)
  • Can we scrub PII? → Microsoft Foundry Agent Service (customer deletes in Cosmos/Storage), Studio (bulk delete transcripts), M365 Copilot (user clears activity history, Purview retention)
  • Can we run without saved transcripts? → Teams calling “Only during the call” mode keeps speech-to-text transient; Copilot Studio toggle stops Dataverse saving.

Sources:

Confidence Level: High (all technologies GA)

Orchestration Complexity Decision Matrix

Complexity Level Characteristics Recommended Technologies
Simple (Q&A) Static lookup, single-turn answers, basic RAG Search or classic RAG first; add an agent only for action, adaptation, or open-ended reasoning
Moderate (Task Execution) Multi-turn, 1-5 actions, simple branching Declarative Agents with API plugins; Foundry Prompt Agents when Foundry-managed configuration fits
Complex (Workflows) Sequential workflows, conditional logic Declarative Agents + Power Automate, Agent Framework workflows
Advanced (Multi-Agent) Agent-to-agent delegation, parallel execution Copilot Studio multi-agent (per-feature status varies: the what’s-new feed shows “Connect other agents” as Preview; Microsoft states no blanket GA); Foundry incoming A2A endpoint (Preview) for direct delegation; Agent Framework (GA core) for code-first orchestration
Expert (Custom Reasoning) Custom orchestration logic, model selection Custom engine agents with Microsoft 365 Agents SDK, Teams SDK, or Microsoft Foundry Agent Service

Component fit heuristic (Copilot Studio, GitHub Copilot harness):

  • If behavior is always true, use Instructions.
  • If behavior is scenario-specific and repeatable, use a Skill.
  • If behavior requires exact system actions, use Tools plus deterministic workflow and approvals.
  • If behavior requires a separate domain owner or trust boundary, consider Connected agents.

Sources:

Confidence Level: High (official patterns documented)

Budget & Timeline Quick Guide

Timeline Assumptions: Estimates assume a 2-4 person team with existing Azure or Microsoft 365 tenant, standard use case complexity, and no specialized compliance requirements.

Scenario Fastest Path Most Cost-Effective (Long-Term)
Extend M365 Copilot (Knowledge Only) Graph Connectors (days) Graph Connectors (included with Microsoft 365)
Extend M365 Copilot (Knowledge + Actions) Declarative Agent (Copilot Studio, 1-2 weeks) Declarative Agent (M365 Agents Toolkit, 2-4 weeks)
Custom Multi-Channel Agent M365 Agents SDK with templates (2-4 weeks) Microsoft Foundry (consumption-based, 4-8 weeks)
Multi-Agent Orchestration Copilot Studio connected agents (Preview) for low-code delegation Agent Framework (GA core) for code-first orchestration; Foundry A2A (Preview) only for direct delegation
Enterprise Workflow Automation Azure Logic Apps agentic workflows (1-2 weeks; Consumption explicitly in preview, Standard status not stated by Microsoft) Microsoft Foundry + Agent Framework (4-8 weeks)

Key Budget Considerations:

  • M365-Centric: Per-user licensing (Copilot Studio, M365 Copilot add-on)
  • Azure-Native: Consumption-based (Azure OpenAI tokens, AI Search queries, compute)
  • Hybrid: Mix of per-user and consumption models
  • Two prepurchase plans, not one: The Copilot Credit Pre-Purchase Plan buys CCCUs against eligible Copilot Credit usage; the Microsoft Agent Prepurchase Plan buys ACUs against select services across Foundry, Copilot Studio, Fabric, and GitHub. One ACU pays down US$1 of qualifying retail cost, purchased at a tiered discount. Learn now documents Copilot Credit P3 for Cowork and Work IQ API spending policies. The checked pages do not name Code or Copilot Managed Runtime for prepurchase coverage as of 2026-09-25. See Evaluation Criteria: Budget
  • Discount precedence (verbatim): “Reservations always apply before prepurchase plans.” The order is Microsoft Foundry PTU Reservations → Microsoft Fabric Capacity Reservations → Copilot Credit Prepurchase Plan → Microsoft Agent Prepurchase Plan (applied last, to remaining AI usage across all platforms). The governing principle is simple: narrow benefits burn before broad benefits.
  • Hidden Costs: Team training, governance setup, responsible AI evaluation, ongoing maintenance

Billing Currencies: The Five Wallets

Microsoft does not have one AI meter. It has five, and they do not convert into each other. Track all of them or your forecast is fiction.

Wallet Unit What it meters Notes
Copilot Credits Credit Microsoft 365 side: agent usage, Copilot Cowork activity, Work IQ APIs, Code and Autopilot (UBB), and Copilot Managed Runtime apps (Preview; each app launch plus 0.1 credit per API call) Pay-as-you-go at US$0.01 per credit; runtime entitlement can change which charges apply
GitHub AI credits Credit GitHub Copilot premium usage 1 credit = US$0.01. Each assigned Copilot Business license adds 1,900 credits per month; Enterprise adds 3,900. Credits form a shared monthly pool at the billing entity; unused credits do not carry over
Azure consumption Tokens, PTU, batch Microsoft Foundry models and platform services Pay-per-token, Provisioned Throughput Units, batch, and priority processing
Fabric Capacity Units CU Fabric workloads, including Fabric data agents and Foundry Tools in Fabric F2 to F8192; Fabric data agents require F2+ or P1+
Azure SRE Agent AAU AAU Azure SRE Agent runtime 4 AAU per agent-hour always-on, plus variable usage. (The term is AAU; “Azure Agent Units” is not a Microsoft term)

The USL funds everyday AI at a fixed price. UBB adds metered advanced AI and frontier models; these experiences consume Copilot Credits and require a USL for access (USL and UBB, updated 2026-09-25). Managed Runtime app users have a different entitlement rule: Copilot Credits or Power Apps Premium, not a Microsoft 365 Copilot license. Power Apps Premium avoids runtime credits except for separately billed services such as Work IQ APIs or usage above its API request limits (admin overview).

Sources:

Sources:

Confidence Level: Medium (pricing models subject to change; verify current rates for production planning)

Governance Decision Quick Reference

Requirement M365 Trust Boundary Azure Workload Boundary
Data Residency M365 tenant region Azure region selection
Identity Management Entra ID (automatic) Entra ID + Azure RBAC
Approval Workflows M365 admin center (Integrated Apps) Azure custom workflows
Audit Logging Purview + M365 audit logs Azure Monitor + Log Analytics
Data Classification Purview DLP policies Azure Policy + custom tagging
AI Governance Microsoft Purview AI protections (DLP, compliance, audit) Microsoft Foundry control plane + Azure Policy
Deployment Control IT admin approval (tenant-wide settings) Azure DevOps, CI/CD pipelines
Real-Time Meeting/Call Transcripts Teams calling policy can enforce “Only during the call” to avoid post-call storage Build ephemeral transcript handling or disable logging in custom stack

Copilot Studio compliance snapshot: Covered under HIPAA BAA, HITRUST CSF, FedRAMP High, SOC, ISO 9001/20000-1/22301/27001/27017/27018/27701, PCI DSS, CSA STAR, UK G-Cloud, OSPAR, K-ISMS, Singapore MTCS Level 3, and Spain ENS High (audit artifacts via the Microsoft Service Trust Portal). (Updated: 2024-12-20)

Best For:

  • M365 Trust Boundary: Organizations with strong Microsoft 365 governance already in place
  • Azure Workload Boundary: Organizations requiring granular control, custom policies, or Azure-native compliance

Sources:

Confidence Level: High (official Microsoft compliance documentation)

Quick Navigation

Next: Resources - Official Microsoft references to keep research evidence-backed

  1. Copilot Frontier is the early access program for experimental and preview features in Copilot apps and agents; enable via Microsoft 365 admin center > Copilot > Settings > User access > Copilot Frontier. Source: Manage Microsoft 365 Copilot scenarios (Retrieved: 2026-03-16). ↩

  2. Mobile parity for custom engine agents and message-extension agents on iOS/Android. Source: Microsoft 365 Copilot release notes (November 24, 2025). ↩

  3. Copy to Copilot Studio copies agent data sources and actions; GPTs and custom actions must be reattached. Feature availability may vary by tenant. Source: Declarative Agents overview (Updated: 2025-12-01). ↩

  4. Foundry also has a documented autopilot definition: “An agent either has its own agent user account or it doesn’t”. The Copilot app’s product description says it has its own identity, but Microsoft has not stated whether it is an instance of that Foundry definition as of 2026-09-25. See Technologies for Scout’s desktop prerequisites and the separate Foundry preview path. Do not confuse these uses with Windows Autopilot device provisioning or the GitHub Copilot app’s Autopilot session mode. ↩

  5. Agent 365 SDK (Preview) extends agents with Entra-backed identity, notifications, OpenTelemetry observability, and governed MCP servers. Agent 365 platform is GA; SDK remains Preview. Source: Agent 365 SDK (Retrieved: 2026-01-09). ↩

  6. Agent 365 CLI (Preview) is a cross-platform CLI for deploying and managing Agent 365 apps on Azure; install via dotnet tool with --prerelease. Agent 365 platform is GA; CLI remains Preview. Source: Agent 365 CLI (Retrieved: 2026-01-13). ↩

  7. Agent Registry lifecycle actions in M365 admin center: publish, activate, deploy, pin, block, remove, delete, reassign owner, export inventory. Source: Agent Registry documentation (Updated: 2026-01-23). ↩

  8. Microsoft 365 Copilot Search API (Preview) for hybrid semantic + lexical search across OneDrive via Graph /beta. Source: Search API overview (Updated: 2025-10-20). ↩

  9. Microsoft 365 Copilot user-level memory allows personalized experiences based on user preferences and context. This is distinct from org-wide conversation logging; memory is user-controlled, while conversation history follows Purview retention policies. Source: Data, privacy, and security for Microsoft 365 Copilot (Updated: 2026-01-07). ↩

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.