Microsoft AI · Architecture and Azure billing · Checked 13 October 2026

From Copilot to a system you can operate.

Copilot gives people a place to work, Foundry runs agents and Fabric supplies governed data. The seats are bought in Microsoft 365, but agents, capacity and models are metered through an Azure subscription. Plan identity, permissions and spending stops together.

Find your starting point Follow the Azure bill

A focused map of Microsoft's AI stack rather than the full Azure catalogue. Start with one measurable workflow, and add a service and its meter only when the workflow needs them.

Architecture before shopping

One workflow. Several distinct responsibilities.

A service name does not tell you who can read a record, approve an action or pay for a failed loop. Make those boundaries explicit.

ExperienceCopilot + Studio
An employee asks for help. The agent gathers context and proposes work.
Billed as seats and Copilot Credits
Choose the user surface →
ExecutionFoundry + tools
A scoped specialist retrieves evidence, calls a model and invokes approved APIs.
Billed as tokens, provisioned throughput and Azure resources
Choose the runtime →
DataFabric + Search
Governed business records and authorized document passages ground the answer.
Billed as Fabric capacity, search units and storage
Choose the data route →

Across every layerEntra identity + Conditional Access + PIMPurview data controls + auditEvaluation + spending stops + accountable operators

An illustrative architecture, not a required bundle. A simple model call may need neither Studio nor Fabric. Permissions and spending limits must survive every handoff.

Choose the service by the job it owns, and know how it is billed
Your need	Starting point	Decision boundary	How it is billed
Employees working in Microsoft 365	Microsoft Copilot and Microsoft 365 Copilot	The app is the surface; the assigned license determines work features.	Per-user seat; Cowork, Code, Autopilot and frontier models also draw Copilot Credits
A business-owned workflow agent	Copilot Studio	Use connectors and managed authoring; establish environments, ownership and billing.	Copilot Credits from prepaid packs or Azure pay-as-you-go
Custom model or agent application	Microsoft Foundry	Own deployment, evaluations, tool contracts, identity and production operations.	Azure meters for tokens, provisioned throughput and supporting resources
Analytics and business data	Fabric + OneLake	Prepare trustworthy tables and semantic meaning before exposing a data agent.	Fabric capacity and OneLake storage on an Azure subscription
Grounding in documents	Azure AI Search / Foundry IQ	Check freshness, source permissions and query-time authorization.	Search service units, plus model calls when query planning runs
Speech, documents or image analysis	Foundry Tools	A specialist API can be simpler and easier to test than a general agent.	Per-request, per-page or per-hour service meters

Microsoft's current documentation: Foundry platform, Fabric overview and Foundry Tools. Billing routes are detailed in the Azure billing section.

The work surface

One Copilot app does not mean one entitlement.

Separate the client, the productivity subscription, preview access and the controls for agents.

The Copilot superapp

The unified Microsoft Copilot app brings personal and work experiences into one client across web, desktop and mobile. Microsoft says data does not flow between those account experiences. Work capabilities depend on the organization's licenses.

The 25 September announcement organizes work around Home, Code and Autopilot: Home brings Chat and Cowork together, Code builds apps and tools, and Autopilot handles persistent delegated work. Home and Code are rolling out through Frontier and Autopilot is in private preview, so the app arriving does not make every mode available. The license covers everyday Chat; the agentic modes also spend Copilot Credits, as the breakdown below shows.

App overview; 25 September announcement and availability.

Agent 365: know what is acting

Microsoft Agent 365 brings agent inventory, observability and governance together with Entra, Purview and Defender. Give each agent an owner, an identity, scoped tools and a lifecycle. Registration is a control point; you still need to test the runtime's actual enforcement.

Commercial availability began 1 May 2026. Standalone US list pricing is $15 per user per month on top of qualifying Microsoft 365, Defender and Purview licensing; E7 includes it. Microsoft lists no consumption charge for Agent 365 itself yet. Model consumption, agent compute and Windows 365 for Agents Cloud PCs are billed separately.

Agent 365 capabilities; licensing, prerequisites and excluded runtime costs.

What the seat covers, and what spends credits

Microsoft splits the Copilot app into everyday AI paid by the license and longer-running agentic work paid in Copilot Credits. The agentic modes need both: a Copilot license for access and a billing policy that an administrator sets up before anyone can use them.

Chat

Included in the license

Questions, drafts and summaries in the Copilot app and in Word, Excel, PowerPoint, Outlook and Teams. Auto picks the model and reasoning level for each request.

GPT 5.6 and Claude Sonnet are included with fair use and Opus with limits. Microsoft plans to let users continue on credits, or switch back to Auto, when they reach a limit.

Cowork

Copilot CreditsGenerally available

Delegated, multi-step work across Outlook, Teams, files and the Edge browser, including scheduled and event-driven tasks. Sensitive actions, such as sending email, ask for approval by default.

Model responses, tool and skill calls, image generation and browser tasks consume credits. Access comes only from a spending policy that selects Cowork; users can type /cost to see a task's approximate usage.

Code

Copilot CreditsFrontier rollout

Builds apps, trackers, dashboards and automations from a description, using the same technology as GitHub Copilot, with optional GitHub or Azure DevOps versioning.

Creating, refining and publishing consume credits. Cloud apps then run on Copilot Managed Runtime, where each user needs Copilot Credits or a Power Apps Premium license.

Autopilot

Copilot CreditsPrivate preview

Formerly Scout: a persistent agent with its own identity, memory, computer and workspace that follows up and runs recurring work when nobody is prompting it.

Runs on usage-based billing. Give it an owner, scoped permissions, approval rules and a spending stop, as you would any persistent agent.

Frontier models such as GPT-6 Astra and Claude Fable 5.1 also draw credits wherever they are chosen, and higher reasoning effort in Cowork uses more of them. Advanced SharePoint work, such as large-scale synthesis or scheduled automations, asks users to confirm before it spends credits. Credit use per task depends on the models, context, tools and runtime involved, so estimate from a pilot rather than a per-prompt price.

License versus usage-based billing; Cowork access, billing and automated tasks; the /cost command; Copilot Credits Guide, October 2026; Managed Runtime eligibility; Copilot in SharePoint.

Claude, GPT and where the prompt is processed

Copilot routes work across Microsoft-hosted GPT models, GPT models operated by OpenAI and Anthropic's Claude models. In Cowork, users can pick Opus, Sonnet or Fable alongside the GPT models, and a badge shows which model answered. OpenAI and Anthropic both act as Microsoft subprocessors under the Product Terms and Data Protection Addendum, but where the processing happens differs.

Copilot settings that decide which models users in the EU, EFTA and UK get and where inference runs, checked 13 October 2026
Setting	Default in the EU, EFTA and UK	Effect
Anthropic as a Microsoft subprocessor	Off, set to no users	Turns on Claude in Copilot, Cowork and Researcher, and in Copilot Studio after a separate Power Platform setting. Processing is outside the EU Data Boundary and in-country commitments. Access can be limited to users or groups.
Copilot in Microsoft 365 apps with Anthropic models	On for tenants created after 25 March 2026	Lets Claude be the default model in Word, Excel and PowerPoint, processed outside the EU Data Boundary.
Flex routing during peak load	On for eligible EU and EFTA tenants created after 25 March 2026	At peak times, inference may run in the United States, Canada or Australia. Stored data stays in the EU Data Boundary apart from pseudonymized identifiers.
OpenAI as a Microsoft subprocessor	On since 24 July 2026 for eligible commercial tenants	OpenAI-operated GPT models stay inside the EU Data Boundary but outside in-country processing commitments.

An AI Administrator or Global Administrator changes these under Copilot > Settings. With Anthropic off, Cowork keeps working with the GPT models you allow and Auto chooses among them. Frontier never enables a provider for you, and some newer Anthropic models, such as those with data retention, need a separate opt-in under Anthropic's own terms. Treat each setting as a data-processing decision and review it with the GDPR guide.

Anthropic as a subprocessor and EU defaults; Anthropic models in Microsoft 365 apps; flex routing; OpenAI as a subprocessor; Cowork models and reasoning effort.

Three different uses of Frontier

Preview access

The Frontier program

Opt-in early access to preview features, subject to licensing and admin controls.

Bundle · $99

Microsoft 365 E7, The Frontier Suite

E5, Microsoft 365 Copilot, Agent 365 and Entra Suite. Launched 1 May 2026; the US Teams-inclusive list is $99 per user per month, paid yearly with an annual commitment.

Microsoft Frontier Tuning

A separate model customization offer, unrelated to preview access or the E7 bundle.

A preview program, a subscription bundle and model tuning solve different problems. Confirm country, Teams variant, feature status and contract before buying.

Frontier administration; E7 plans and pricing; Frontier Tuning.

Follow the money

The seat is in Microsoft 365. The meter is in Azure.

Agents for unlicensed users, Copilot Studio, Cowork, Fabric capacity, Foundry models and several security and compliance features bill usage to an Azure subscription. An organization that has never used Azure can acquire its first subscription this way.

How a Microsoft AI charge reaches an invoice

Three ways to pay, two admin centers and one Azure subscription.

Licensed per user
Seats
Bought in the Microsoft 365 admin center or from a CSP. The same price each month, whether or not the user is active.
- Microsoft 365 Copilot$30
- Copilot Business, up to 300 users$21
- Agent 365$15
- Microsoft 365 E7$99
Metered in Copilot Credits
Agent and Copilot usage
Prepaid capacity packs need no Azure subscription. Pay-as-you-go, P3 plans and overage bill one through a billing or spending policy, after any prepaid credits.
- Pay-as-you-go$0.01 a credit
- Capacity pack, 25,000 credits a month$200
- Copilot Credit P3, one year5-20% off
Azure resources
Capacity, models and compute
Created in the Azure portal, or by a partner, as resources in a subscription and priced by region and meter.
- Fabric capacityper second
- Foundry modelsper token or PTU
- Windows 365 for Agents, US$0.40 an hour

The subscription's billing account decides who invoices you. A CSP partner bills an Azure plan each calendar month, Microsoft bills a Microsoft Customer Agreement directly, and an Enterprise Agreement bills through its enrollment. In CSP, customers see Azure costs in Cost Management only after the partner enables cost visibility, which is off by default and shows retail rather than invoiced prices.

US list prices checked 13 October 2026. Seats are per user per month, paid yearly with an annual commitment; Copilot Business has an $18 first-year promotion for eligible existing Business customers. CSP and negotiated prices differ.

Where Microsoft AI usage is metered, where billing is connected and what actually stops spending
Service	Meter	Billing connected in	What can stop spending
Copilot Studio agents	Copilot Credits, $0.01 each on pay-as-you-go	Power Platform admin center: a billing plan links environments to a subscription	A monthly agent limit set to stop usage, or unlinking the environment. Prepaid-only environments disable custom agents at 125% of capacity.
Copilot Chat and SharePoint agents for unlicensed users	Copilot Credits; a successful SharePoint agent interaction uses 12	Microsoft 365 admin center: a pay-as-you-go billing policy for all users or one group	Disconnecting the policy, which can take up to two hours. The policy budget only sends email.
Cowork, Code, Autopilot, frontier models, Work IQ API, Copilot Managed Runtime and Teams Phone Agent	Copilot Credits, on top of the Copilot license the modes require	Microsoft 365 admin center: a spending policy in Copilot Cost management	Monthly policy and per-user spending limits. The policy is also the access gate, and new services join it automatically unless you turn that setting off.
Fabric capacity	Capacity units per second with a one-minute minimum; overage at three times pay-as-you-go	Azure portal or a CSP: an F SKU resource in a subscription	Pause or resize. Overage is on by default and its threshold is not a hard cap.
Foundry models and agents	Tokens or provisioned throughput, plus tools, Search and hosting	Azure portal: Foundry resources in a subscription. Partner models bill through Azure Marketplace.	Quotas and rate limits cap throughput, not spend. Build cost limits into the application.
Windows 365 for Agents	$0.40 per Cloud PC hour in the US, rounded up; $5 a month per always-available PC	Microsoft 365 admin center billing policy, or a Copilot Studio Cloud PC pool	Disabling it for the billing policy makes new Cloud PC check-outs fail.
Security Copilot	Security Compute Units, provisioned hourly plus overage; E5 and E7 include 400 a month per 1,000 licenses, up to 10,000	Azure subscription, or the Security Copilot portal	The provisioned count and an overage maximum. Unlimited overage is an option, not a ceiling.
Purview pay-as-you-go	Requests, processing units or GB, such as prompts and responses in non-Microsoft AI apps	An Azure subscription associated with the Microsoft 365 tenant	Turning off or narrowing the pay-as-you-go features and policies that generate the meter.

On the Azure invoice, Cowork, Copilot Studio and Work IQ usage appear together as one Copilot Credits service labelled Microsoft Copilot Studio; separate them by subscription or resource group when you need to know which one spent. The same billing policies also carry non-AI Microsoft 365 services such as Backup, Archive, extra SharePoint storage and document processing. GitHub usage can bill through an Azure subscription linked to a GitHub organization.

Microsoft documentation: Power Platform pay-as-you-go; Copilot Studio capacity and agent limits; Copilot Chat pay-as-you-go; usage-based billing for Copilot Credits; SharePoint agent messages; Fabric capacity overage; Foundry model billing; Windows 365 for Agents pricing; Security Compute Units; Purview billing models; CSP cost visibility. Prices: Copilot Credits and P3 tiers; Copilot Business; Microsoft 365 Copilot.

From 1 December 2026

Copilot Business through CSP turns usage billing on by default

Microsoft says new Microsoft 365 Copilot Business licenses bought through CSP will include the Azure subscription setup for usage-based billing, with pay-as-you-go as the default configuration for Copilot Cowork, Work IQ APIs and the GitHub Copilot harness. The default limit is 4,000 Copilot Credits per user per month: up to $40 per user at the $0.01 list rate, above the seat's $21 list price. The rollout initially excludes Australia, Belgium, Brazil, France, Germany, India, Italy, Korea, the Netherlands, Poland and Spain.

Choose the per-user limit, the users and the services before the first license is assigned. Ask the partner which subscription and billing account the license sets up, and who receives the alerts.

Partner Center announcement, 1 October 2026.

One agent turn can trigger several charges

In Copilot Studio's Standard harness, a generative answer is 2 credits, an agent action 5 and tenant Graph grounding 10. A reasoning model adds 10 credits per 1,000 reasoning tokens. Prompt tools, agent flows, voice minutes and computer use have their own rates, so a turn is not a fixed-price unit.

Employee-facing use by a person with a Microsoft 365 Copilot license is included, within fair-use limits, when the agent acts with that person's identity. Computer-using agents, agent flows started by other triggers, external channels and unlicensed users consume credits.

The GitHub Copilot harness in Studio bills credits for model, tool and harness consumption during authoring, testing, evaluation and runtime. Microsoft 365 Copilot licenses do not include it, and a separate GitHub Copilot coding subscription does not pay for it. In the Copilot app itself, Cowork, Code, Autopilot, frontier models and apps on Copilot Managed Runtime spend credits on top of the seat; see what the seat covers.

Standard rates and licensed-user inclusion; GitHub Copilot harness billing; Copilot usage billing announcement.

One illustrative turnStandard harness, pay-as-you-go

Generative answer: 2
Agent action: 5
Tenant Graph grounding: 10
Total: 17 credits = $0.17

At 2,000 such turns on each of 21 working days, that is 714,000 credits, or $7,140 a month, before retries, failed loops and evaluations.

Brakes before dashboards

An alert is not a brake.

Most budgets in this stack send email and let usage continue. Configure the controls that stop usage first, then add alerts for everything they do not cover.

Notifies only

Azure Cost Management budgets. Cost data usually arrives within 8-24 hours and budgets are evaluated every 24 hours. Resources keep running; an action group can call automation you build and test.
Microsoft 365 pay-as-you-go policy budgets. Email at thresholds; usage continues past the budget.
Fabric overage threshold. A rolling 24-hour limit evaluated periodically; charges can exceed it.
Anomaly and spending-policy alerts. Useful detection, without enforcement.

Can stop usage

Copilot Studio agent limits. A monthly limit for each agent, set to stop usage, in the Power Platform admin center.
Copilot Credit spending limits. Policy-level and per-user monthly limits in Microsoft 365 Copilot Cost management. A policy also grants access: even a one-credit limit lets someone start Cowork work, so leave people out of the policy to keep them out.
Prepaid-only capacity. Custom agents are disabled at 125% of prepaid capacity. Adding pay-as-you-go overage removes this stop.
Unlink or disconnect. Unlinking an environment ends its pay-as-you-go billing; disconnecting a Microsoft 365 policy removes access to metered agents within about two hours.
Pause Fabric; cap Security Copilot. Pausing ends capacity billing after charging usage already smoothed forward. An overage maximum bounds Security Copilot.

Set up the Azure side before the first metered agent

Choose who invoices. Decide whether AI meters belong to a CSP Azure plan, a Microsoft Customer Agreement or an Enterprise Agreement. A Global administrator can create a subscription from the Microsoft 365 admin center on its linked MCA billing account, which is a different account from an Azure plan your partner manages.
Separate the subscription. Keep SaaS meters in a dedicated subscription, or at least one resource group per billing policy, with a named business owner and finance owner.
Treat billing rights as privileges. Attaching pay-as-you-go typically needs an admin-center role plus Owner or Contributor on the subscription. Keep both groups small, use PIM and review them.
Scope before you enable. Link only the pilot's environments and groups. Turn off automatic coverage of new services wherever each new meter needs approval.
Set stops, then alerts. Configure agent and per-user limits first. Add Azure budgets with action groups at subscription and resource-group scope for what the stops do not cover, and ask a CSP to enable cost visibility.
Reconcile monthly. Microsoft 365 dashboards count credits while the invoice shows currency, and Copilot Studio usage is attributed to environments and agents rather than people. Report cost per accepted task.

Azure budget timing and behavior; Microsoft 365 billing policies, budgets and subscription creation; spending limits and automatic service coverage; agent stop limits and attribution; pausing Fabric. Use the 90-180 day review to judge whether the spend produced value.

Lower the bill · checked 13 October 2026

Same token price. A different effective cost.

Matched GPT-6 list prices are identical on Azure and OpenAI, and data-zone residency costs more. Savings come from commitments you fully use and from agreements you already have.

USD per 1 million tokens, Standard short-context tier up to 272K input tokens. Azure Global Standard compared with direct OpenAI Standard.
Model	Azure input / cached / output	OpenAI input / cached / output
GPT-6 Astra	$10 / $1 / $50	$10 / $1 / $50
GPT-6 Sol	$2 / $0.20 / $10	$2 / $0.20 / $10
GPT-6 Luna	$0.10 / $0.01 / $0.50	$0.10 / $0.01 / $0.50

Azure's US Data Zone rates are 10% above Global Standard and its EU Data Zone rates 20% above: GPT-6 Astra costs $12 input and $60 output per million tokens in the EU Data Zone. Cache writes, long-context uplifts, tools and infrastructure are excluded. Both providers discount asynchronous Batch processing; compare Batch with Batch, not with an interactive Standard request.

Microsoft's published deployment-price table; OpenAI model pricing for Astra, Sol and Luna.

Which commitment pays first

When benefits overlap, Microsoft applies them in this order. Usage left over moves to the next step and finally to pay-as-you-go.

Foundry PTU reservationApplied hourly to provisioned throughput. Unused hours expire, and capacity is billed while idle.
Fabric capacity reservationApplied hourly to capacity units. Unused reserved units don't carry over.
Copilot Credit P3One-year commit units for Copilot Credit usage, in nine tiers from 3,000 units for $2,850 (5% off) to 3 million for $2.4 million (20% off).
Microsoft Agent P3One-year commit units for eligible Copilot Credit, Foundry, Fabric and GitHub usage: 20,000 for $19,000, 100,000 for $90,000 or 500,000 for $425,000 (5%, 10% or 15% off).
Pay-as-you-goEverything else, at retail or your agreement's rate.

Each commit unit offsets $1 of eligible retail consumption. Sources: benefit order and Agent P3 rules, Copilot Credit P3 and published tiers.

Rules that change the arithmetic

Purchases are final. P3 plans can't be cancelled, exchanged, split or merged. They renew automatically unless you change the setting, and unused units expire at the end of the year.
Discounts don't combine. With an existing agreement discount, the effective P3 saving can be lower than the tier percentage.
EA prepayment isn't drawn. A P3 purchase is its own invoice line rather than a draw on an Enterprise Agreement's Azure Prepayment.
Check each meter. The plans pay eligible Microsoft-metered usage. Partner models sold through Azure Marketplace follow their own terms.

Reservations, savings plans and MACC

Month or year reservations can lower the cost of deployed Foundry PTUs, but a financial reservation does not guarantee available model capacity. The advertised up to 70% example compares GPT-4o hourly PTU with a one-year reservation using May 2025 prices; it is not 70% off Standard tokens or direct OpenAI.

Azure savings plans cover eligible compute and database infrastructure, not serverless model tokens. Eligible spend can fulfill a Microsoft Azure Consumption Commitment (MACC), but MACC is not a public token discount. Ask for an agreement-specific quote.

PTU reservation mechanics; headline and footnote; savings-plan coverage; MACC accounting.

A discount is useful only if you use the commitment

At the 15% Agent P3 tier, $425,000 prepays $500,000 of eligible retail consumption. Use only $400,000 before expiry and plain pay-as-you-go would have cost less. Break-even utilization is 85%; at the 5% tier it is 95%, before migration and operating costs. Size a plan from at least a few months of measured meters, not from a pilot forecast.

Break-even figures are arithmetic illustrations, not quotations.

When an Azure quote is worth requesting: if you already spend materially on the OpenAI API, ask for a quote that names the eligible meters, the P3 tier, any reserved throughput you would fully use and your existing commitments. Compare total cost per successful task, including quality, geography, latency, Search, storage, networking, support and operations.

October 2026 · Azure CSP promotions

Announced 12 October. Available from 2 November 2026.

Five CSP discounts on the services around AI.

Time-bound discounts on eligible pay-as-you-go usage for eligible customers, redeemed through a CSP partner. Each figure is a maximum, and none of them discounts Foundry model tokens.

Up to30%

Azure Database Modernization

Eligible Azure database usage for customers migrating on-premises SQL Server workloads.

Up to20%

Fabric for Microsoft 365 customers

Fabric pay-as-you-go capacity for eligible Microsoft 365 customers new to Fabric.

Up to20%

Fabric for Azure database customers

Fabric pay-as-you-go capacity for Azure database customers with no prior Fabric spend.

Up to15%

App PaaS for Foundry customers

Eligible Azure App PaaS usage for Microsoft Foundry customers.

Up to15%

Data services for Foundry customers

Eligible Azure data services usage for Microsoft Foundry customers.

Before you count on it

Partner Center or its eligibility API decides eligibility and redemption. Microsoft has not published country lists, eligible meters, promotion durations or end dates for these offers, so ask your CSP for the eligibility result, redemption window and final customer price.

Microsoft applies the discount to the partner's Azure billing and leaves its partner-earned credit intact; your price depends on what the partner passes through. Do not add these offers to a P3 or reservation saving without written stacking terms.

October announcement and 2 November availability; promotion eligibility and redemption lifecycle.

Prove the workflow

Start in Studio when the business can own the agent.

A policy assistant, service-desk triage or approval workflow is a useful starting point when scope, evidence and actions are clear.

Use managed authoring well

Copilot Studio gives makers agent instructions, knowledge, connectors, tools and workflow orchestration. Choose an authoring harness deliberately: capabilities, supported channels and billing differ. Keep development, test and production in controlled Power Platform environments, and decide which of them may bill pay-as-you-go.

Start with a read-only answer or a draft. Add a write action only after its API checks the caller, record scope, allowed operation and any required human approval. A connector being available is not permission to use it on every record.

Copilot Studio overview; harness-specific usage billing.

Keep what the pilot proved

Graduation is an architecture decision.

Move specialist execution into Foundry when custom orchestration, model choice, deployment control or a reusable API justifies the engineering.

Prove it in Studio
Validate the workflow with business users. Save instructions, tool schemas, representative inputs, permissions and acceptance cases in version control.
Delegate a specialist
Retain Studio's useful experience and call a Foundry agent or scoped backend. The documented connected Foundry agent path is public preview, uses the new Foundry portal and requires an enabled Activity protocol endpoint.
Promote through engineering gates
Reimplement the pieces that need code, add evaluation and rollback gates, and deploy separately. Foundry agents can also publish directly into Copilot and Teams; organization-wide distribution needs Microsoft 365 admin approval.

There is no verified universal one-click export. Treat connected agents, custom tools and a new implementation as distinct options. Studio's bring-your-own-model prompt connector is another route with its own limits: current documentation requires Chat Completions and excludes GPT-5 and later models. Studio's credit rates exclude those Foundry model calls, which bill separately in Azure.

Connect a Foundry agent; publish to Copilot and Teams; BYOM prompt compatibility.

The developer platform

Use Foundry for the parts you need to engineer.

Microsoft Foundry is the current name for Azure AI Foundry. Models, agent hosting, tools, evaluation and governance sit within an Azure operating boundary.

Models and agents

Start with a direct model call if that solves the task. Add a prompt agent for managed instructions and tools, or a hosted agent when you need your own code and orchestration. Check each feature's availability, deployment type, model region and data handling separately.

Use tracing to connect a request to retrieval, model calls, tool decisions and outcome. Evaluate quality, authorization failures, latency and cost before deployment. A multi-agent demonstration adds coordination overhead as well as possible specialization.

Foundry capabilities and runtime choices.

Search and the former Cognitive Services

Azure AI Search supplies document indexing, keyword and vector retrieval and semantic ranking. Foundry IQ builds a managed knowledge layer on Search's agentic retrieval. Query planning may add model calls and latency; compare it with a simpler retrieval pipeline.

Foundry Tools, previously Azure AI services and Cognitive Services, covers Speech, Translator, Language, Document Intelligence, Content Understanding, Vision and Content Safety. Use document extraction or speech recognition when that is the actual job. Check service-specific meters and retirement notices.

Search and agentic retrieval; current specialist service catalogue.

A retrieval index is a second copy of your security problem

Propagate source access rules and deletion changes into the retrieval system. Enforce permissions before passages reach the model. Search documents several approaches, including application security filters and native ACL or sensitivity-label paths that remain preview features. Foundry IQ does not automatically honor every source's permissions.

Document-level controls and status; Foundry IQ permission limits. Read the RAG guide for retrieval evaluation.

Match the model to the task

Microsoft builds models, too.

Foundry's value includes Microsoft's own models alongside other providers. A small specialist can be the better choice for one step of a larger workflow.

Current Microsoft-developed families, grouped by job and checked 13 October 2026. MAI hosted services, MIT open weights, models adapted from Qwen and domain research have different deployment and licensing boundaries. A model in the catalogue is not necessarily a managed, generally available API; check the linked card before deployment.

Reasoning, code and agents

MAI-Thinking-1

Foundry previewHosted

General text reasoning, mathematics, coding and function calling. A candidate for a reasoning or agent step.

Global Standard, with no current PTU support. A hosted service, not verified open weights. Model and deployment details.

MAI-Code-1.1-Flash

In GitHub CopilotLocal download

Agentic coding: changes, debugging and tests, with image input. Evaluate it within the coding harness and repository.

Local download announced 7 October; local Copilot integration is still forthcoming. A general Foundry endpoint and open-weight license are not established. Release and access; model repository and image input.

Microsoft-Decision-1

FoundryOpenRouter

Classification, routing, prioritization and verification over predefined options, with probability outputs. Useful before an expensive escalation.

Microsoft post-training of Qwen3.5-9B. Probability is not authorization. Model behavior and deployment.

MagenticBrain-14B

MIT weightsFoundry preview

Text-based orchestration, declared tools and coordinating specialist agents.

Qwen3-derived. Runtime code owns execution, retries and authority. Official model card.

Fara1.5

MIT weightsFoundry catalogue

Screenshot-based browser computer use: turn an instruction and visual state into browser actions.

4B, 9B and 27B Microsoft adaptations of Qwen3.5. The harness must control the actual browser and permissions. Research release.

Voice, images and vision

MAI-Transcribe-2

Catalogue preview

Recorded audio to text, with diarization, timestamps, keyword biasing and clean or verbatim styles.

Foundry and the Azure Speech Fast API. A transcription model, not a conversational reasoning model. Microsoft model card.

MAI-Transcribe-2-Streaming

Preview

Live transcription, dictation and captions, returning partial and final text.

Foundry and Azure Speech, using WebSockets or the Speech SDK. File and streaming transcription support different features. Streaming API.

MAI-Voice-2.1 + Flash

Launched 1 October

Expressive multilingual text-to-speech, consistent voices and emotion controls. Creating a voice needs appropriate consent.

Foundry and Azure Speech. Text-to-speech is one component of a voice agent. Voice models and requirements.

MAI-Image-2.6 + Flash

Foundry previewHosted

Image generation and editing, reference images, layout, rendered text and visual consistency. Flash offers a different cost and latency trade-off.

Hosted commercial models. Earlier 2.5 variants remain listed; evaluate the current versions first. Variants and deployment.

Florence-2

MIT weights

Vision extraction: captions, object detection, phrase grounding, OCR and segmentation.

For local or custom deployment. A vision specialist with no direct managed Azure endpoint established here. Model card and license.

Small models and specialists

Phi-4 + mini-instruct

MIT weightsFoundry

Compact text generation, extraction, instruction following and bounded local tasks.

Selected Foundry deployments. Status and hardware needs vary by version. Phi-4 model card; current variants.

Phi-4 reasoning + mini-reasoning

MIT weightsFoundry

Smaller-model reasoning over mathematics, code and structured problems.

Test task quality and reasoning-token overhead rather than assuming frontier-model equivalence. Reasoning model card.

Phi multimodal + visual reasoning

MIT weights

Phi-4-multimodal-instruct takes text, images and audio to text. Phi-4-reasoning-vision-15B reasons over images and documents. Phi-Ground-Any maps screenshot instructions to GUI coordinates.

Multimodal and Reasoning Vision have Foundry catalogue routes; Ground-Any is specialist research weights. Multimodal; Reasoning Vision; Ground-Any.

MAI-Cyber-1-Flash

Restricted access

Vulnerability discovery and remediation inside the MDASH multi-agent security system.

For verified defenders through a register-interest route; not a general Foundry chat model. Access conditions.

Domain research models

Individual licenses

Aurora for weather; Muse for gameplay and world modeling; MatterGen, RetroChimera and EvoDiff for materials, chemistry and proteins; GigaTIME, Prov-GigaPath, CxrReportGen and MedImageInsight for medical imaging.

Each has its own license and lifecycle state. Keep them outside the ordinary enterprise agent shortlist; catalogue access does not establish clinical authorization. Microsoft publisher inventory.

Microsoft also describes MAI models specialized for Excel. Product-specific results do not establish a separately offered public Excel-model API. Microsoft's production account. The MAI model hub and publisher catalogue are the routes to recheck new releases.

Compare capabilities with the right alternative

ChatGPT, Claude and Grok can name user products as well as model families. Compare the API model, version and execution harness you would actually deploy. Kimi K3 and GLM are current alternatives; downloadable weights and a managed third-party endpoint are different hosting choices.

A workload comparison, not an assertion of equal intelligence or benchmark rank
Workload	Microsoft shortlist	Alternatives and deciding test
General reasoning and tools	MAI-Thinking; Phi reasoning for bounded work.	OpenAI GPT-6, Claude Fable/Opus/Sonnet, Grok 4.7, Kimi K3, GLM-5.3. Test task completion, tool correctness, latency and full cost.
Coding	MAI-Code in Copilot; MAI-Thinking; Phi reasoning.	OpenAI/Codex, Claude, Grok, Kimi and GLM coding workflows. Run the same repository tasks, tests and tool permissions; measure accepted changes.
Images and visual tasks	MAI-Image for creation; Phi vision or Florence for understanding.	OpenAI image generation and other actual image generators for creation; Claude/OpenAI/Grok/Kimi/GLM vision models for understanding. Image input support does not imply image output.
Speech products	MAI-Transcribe + a reasoning model + MAI-Voice.	OpenAI or xAI realtime voice routes and open speech pipelines. Compare end-to-end latency, interruption, language accuracy and consent controls.
Routing and classification	Microsoft-Decision-1; Phi.	Small OpenAI/Claude models or suitable Kimi/GLM/open specialists. Test calibrated error rates, abstention, escalation and cost.
Computer use	Fara; Phi visual reasoning and grounding.	Claude/OpenAI computer-use systems or open multimodal agents. Test the complete browser harness, approval gates and recovery from a wrong click.

Current alternative families: Anthropic model overview; OpenAI model catalogue; Grok in Foundry; Moonshot Kimi; Foundry's Fireworks route for Kimi K3 and GLM. These links establish offerings; they do not prove equivalent performance.

One catalogue, different data and billing boundaries

For example, the Fireworks route requires admin enablement and shares data with that provider. Current documentation excludes those deployments from Microsoft's EU Data Boundary and from PCI cardholder-data use. Models sold by Azure bill as Azure meters with Microsoft support; partner and community models bill through Azure Marketplace under their own terms. A charge on an Azure invoice does not make every model share the same processor, geography or compliance terms.

Provider and data-processing conditions; billing for models sold by Azure and by partners.

The data foundation

The agent needs business meaning, not just files.

Fabric brings ingestion, engineering, analytics and business consumption together. OneLake is its shared data lake; a lakehouse is one way to organize and query that data.

OneLake, lakehouse and ADLS

OneLake comes with a Fabric tenant and is built on Azure Data Lake Storage. Shortcuts can expose supported existing storage without copying it; mirroring is a separate replication mechanism. A Fabric lakehouse combines files and Delta tables, Spark processing and a read-only SQL analytics endpoint.

You can retain an existing ADLS Gen2 lake or other supported source and connect it where appropriate. Fabric is useful when several teams need shared data engineering, analytics and semantics. A small document assistant does not need a new enterprise lakehouse just to start.

OneLake architecture and shortcuts; lakehouse overview.

From governed data to agent context

Define owners, trusted tables, metrics, relationships and freshness targets. Fabric data agents can expose analytical questions to Copilot Studio, while the data agent still runs in Fabric. Both agent access and underlying source permissions matter. Choose authentication deliberately: User authentication uses the caller's access; Maker authentication can expose data through the maker's permissions. Responses passed to Studio can leave Fabric's geographic and compliance boundary.

Fabric IQ adds business context and relationships to the data story. Neither an ontology nor a semantic model grants permission to take an operational action. Keep the write API and its approval rules separately controlled.

Fabric data agents in Studio, permissions and cross-service data boundaries. See the Fabric IQ walkthrough and the operational intelligence guide.

Fabric bills like Azure infrastructure

Budget capacity, OneLake storage and connected services separately from model tokens, and test concurrency and throttling with representative workloads before choosing a size.

Capacity is an Azure resource. F SKUs are bought in the Azure portal or through a CSP, priced by region and billed per second, with a one-minute minimum, until paused.
F64 changes the license count. On F2-F32 every Power BI viewer needs Pro or Premium Per User; on F64 and larger, viewers with a Free license and the Viewer role can read shared content.
Overage is on by default. New capacities pay off excess use at three times the pay-as-you-go rate, up to a rolling threshold that is not a hard cap. Pausing ends compute billing after charging usage already smoothed forward.
Reservations are hourly. Unused reserved units don't carry over. Microsoft's planning guide suggests pay-as-you-go scale-up for occasional peaks and reservations for capacity needed more than four days a week.

Microsoft 365 customers new to Fabric may qualify for an announced CSP promotion. Check this month's offers

Buying capacity; per-second billing and viewer licensing; capacity overage; pause and resume; capacity planning.

Baseline before autonomy

Control the identity. Protect the data. Authorize the action.

These are different enforcement points. Buying a security suite does not configure them for you.

Entra + Conditional Access + PIM

Use Entra for human and workload identity. For people, configure Conditional Access with appropriate MFA, device and risk requirements; test policies in report-only mode and maintain emergency access. Use Privileged Identity Management for time-limited, approval-based administrator elevation instead of standing privilege. Include the roles that can attach billing: Owner or Contributor on a subscription that carries AI meters can add spending.

Give workloads managed identities or federated credentials and the smallest resource roles they need. Human MFA policies do not protect every service-to-service call. Workload Conditional Access has separate licensing and applies to supported service principals; managed identities are excluded from that policy path. Agent identity controls have their own feature status and scope.

Conditional Access; PIM; workload policy limits.

Purview is configured data protection

Use Purview to discover and classify sensitive information, apply labels and supported DLP policies, investigate AI data exposure, and retain audit evidence. Repair overshared source data before allowing agents to search it.

Coverage differs across Microsoft 365, Studio, Foundry and third-party apps. Foundry protection requires explicit integration and configuration; some policy enforcement needs user context, while other scenarios provide audit and classification. Protection for AI apps outside Microsoft 365 uses Purview's pay-as-you-go billing through Azure. Test both permitted and denied requests, copied content, attachments and downstream tools.

Content Safety addresses harmful content. Purview addresses data security and compliance. The application's authorization code still decides whether a caller may read a record or execute a write.

AI app coverage; Foundry integration and prerequisites; pay-as-you-go billing model.

The safety net needs a contract

Buy a route to help. Assign the work of operating.

Managed AI removes some infrastructure work. Governance, availability, security and financial responsibility still need owners.

CSP: the commercial and support contact

A Cloud Solution Provider can supply subscriptions, help with licensing and billing, diagnose issues and escalate service problems to Microsoft. CSP support responsibilities follow the partner relationship; agree contact methods, coverage hours and escalation expectations.

In CSP, the partner invoices Azure consumption for each calendar month and decides whether you can see your costs in Azure Cost Management. Partner-earned credit and CSP promotions change the partner's cost; ask how they reach your price. A reseller relationship does not automatically include architecture reviews, 24/7 operations or incident recovery.

Microsoft's CSP customer-support requirements; CSP Azure billing periods; Azure plan and partner-earned credit.

MSP: the operating help you contract

A Managed Service Provider can design the landing zone, harden identity, run monitoring, handle incidents and maintain cost controls. A company can serve as both CSP and MSP, but agree each responsibility and its price.

Keep customer ownership of the tenant, subscriptions, repositories, data and recovery access. Use narrow delegated privileges, reviews and a documented offboarding path.

Azure landing-zone design areas. The ownership and acceptance questions below are this site's operational guidance.

Make the operating agreement reviewable before production
Responsibility	Evidence to ask for	Acceptance check
Governance and policy	Resource hierarchy, allowed regions, policy assignments, exceptions and named owners.	An unauthorized deployment is denied; an exception expires.
Identity and security	Conditional Access, PIM, scoped identities, secret handling and tool allowlists.	A revoked user or agent cannot retrieve or act.
Availability and recovery	Service dependencies, quotas, retry limits, backups, recovery targets and on-call contacts.	A model or region outage reaches a tested fallback or manual queue.
Networking	Private endpoints where supported, DNS, outbound rules and external-provider data routes.	Approved clients connect; unexpected public or outbound paths fail.
Logging and audits	Correlated traces, action receipts, retention, redaction, alerts and reviewer access.	Reconstruct who approved and executed an action without exposing unnecessary sensitive data.
FinOps	Meter inventory across both admin centers and Azure, stop controls, cost visibility, commitment utilization and forecast.	A runaway agent stops at its limit; cost per accepted outcome is reported monthly.
Portability and exit	Exportable data, prompts, tool schemas, evaluation sets, infrastructure code and replacement-provider test.	Restore a workflow from customer-owned assets; measure the work of changing model or platform.

Borrow a starting point

Inspect the accelerator before you inherit it.

Use sample architectures to shorten discovery. Review their identity, data, deployment and maintenance assumptions before adapting them.

Microsoft's solution accelerators

The catalogue spans Foundry, Fabric, OneLake, Search, Purview and Studio. Start with Unified Data Foundation for the data layer, Data and Agent Governance and Security for controls, or Multi-Agent Custom Automation for orchestration. Its “actively maintained” labels are publisher status, not independent production assurance. Deploying one creates billable resources in your subscription.

Browse the current accelerator catalogue →Unified Data Foundation →Data and Agent Governance and Security →Multi-Agent Custom Automation →

FibreOps / BRK241-frontier: A fibre-outage workflow with analysis, coordination and dispatch agents, Event Hubs, Teams, evaluation and tracing. Dynamics Field Service is mocked; local deterministic mode helps inspect the flow before deploying Azure resources.
On-Call Copilot Multi-Agent: Concurrent triage, summary, communications and post-incident review using Foundry hosted agents and routing. Useful for structured drafts; test incident accuracy and approval boundaries before sending messages or making changes.
Agent Patterns Demo Pack: Visual orchestration patterns, traces and replay using Foundry Local and Microsoft Agent Framework, with a cloud-model option. Compare the coordination cost with a single agent.
Foundry CI/CD: GitHub Actions and Azure DevOps pipeline references for evaluation gates and controlled promotion. Supply the application configuration, evaluation data, tests and deployment scripts the templates expect.
Build 2026 DEM331: Archived conference demo, 2 September 2026. Shows Content Understanding extracting context from field documents for a fibre-cut response agent. Read it for the tool/data design, with no assumption of ongoing maintenance.
LocalStack for Azure: Public preview. Local emulation of selected Azure services for development and testing, with documented parity gaps. It does not establish Foundry inference equivalence or replace real-cloud identity, networking and quota tests.
Azure GPU Setup reference: Historical instructions. This guide targets Ubuntu 16.04 and NV6-era manual driver/CUDA/TensorFlow setup. Use current supported VM images and vendor documentation for a new deployment.

Selected demonstrations

Watch the layer you are deciding about.

The newest relevant Microsoft Mechanics walkthroughs found for these topics, checked 13 October 2026. Vendor demonstrations explain capabilities; they do not prove your deployment's security or ROI.

Microsoft Mechanics · 5 October 2026

Studio: build and evaluate

Authoring, specialist agents, MCP tools, workflow orchestration and human review. Use this to decide what belongs in the managed builder.

Microsoft Mechanics · 14 September 2026

Agent 365: controls across agents

Registry, tool governance, security templates and spending controls. Map the demonstration to the agents and licenses in your own tenant.

Microsoft Mechanics · 9 July 2026

Token economics: design affects cost

Context, input and output tokens, caching and routing. Pair the explanation with the current price and commitment tables above.

Microsoft Mechanics · 23 June 2026

AI landing zones: own the plumbing

Model and tool gateways, policy, telemetry, OneLake and Purview. A useful architectural walkthrough for the CSP/MSP conversation.

Microsoft Mechanics · 28 April 2026

Fabric IQ: add business context

An explanation of operational meaning and relationships over enterprise data. Evaluate permissions and data quality alongside the ontology.

A useful first step

Fund one outcome and the controls around it.

For an operations copilot, start with retrieving the right procedure, summarizing an incident and drafting the next action. Keep execution behind an authorized approval.

Choose the simplest workable user surface and runtime. Use existing governed data where possible. Before launch, link the pilot to its own billing policy and set a limit that stops it. Record correctness, escalation rate, total human time and total cost per accepted task. Expand autonomy only when the evaluation cases, permissions, spending stops, recovery route and operating ownership hold up under real use.

Choose the starting point Set the spending stops Set the 90-180 day review Design the retrieval layer