Last week we mapped Microsoft’s complete first-party AI stack from silicon to cloud. This week, the story shifts to what happens when you open that stack to competitors’ models — and when you push AI into places users don’t want it. Microsoft added Anthropic’s Claude Opus 5 to the Copilot model selector on July 24, giving users their first taste of genuinely multi-vendor AI inside Microsoft 365. Two days earlier, Microsoft made a rare reversal: canceling Copilot Search for Outlook after user outrage. And today, July 28, the Model Context Protocol — the plumbing that connects AI agents to enterprise systems — finalizes its largest specification update, going fully stateless.

For IT leaders, this is the week where multi-model Copilot stops being a roadmap promise and becomes a configuration decision. Understanding which models are available, how they’re governed, and what your users actually want matters more than tracking benchmark scores.

1. Claude Opus 5 Joins Microsoft 365 Copilot

On July 24, Microsoft announced that Anthropic’s Claude Opus 5 is rolling out to the model selector in Microsoft 365 Copilot across Word, Excel, PowerPoint, Copilot Chat, Copilot Cowork, and Copilot Studio. This follows Claude Sonnet 5’s arrival in Copilot on July 2 and brings the total number of selectable models in Copilot Chat to at least five: Auto, Claude Sonnet 5, Claude Opus 4.8, GPT-5.5, and Claude Fable 5, with Opus 5 now replacing Opus 4.8 as the premium Opus option.

The positioning is deliberate. Microsoft frames Claude Sonnet 5 as the workhorse for everyday agentic tasks — moderate complexity, good speed, cost-efficient. Claude Opus 5 is the scalpel: long-document reasoning, multi-step analysis, complex structured content generation. GPT-5.6 remains the default for most interactions. In Copilot Cowork, a paired mode called “Sonnet + Opus Advisor” runs Sonnet for execution and Opus for review, balancing cost and quality.

Admin controls matter here. Anthropic models are off by default for UK, EU, and EFTA tenants because processing occurs outside Microsoft’s EU Data Boundary. Admins must explicitly enable Anthropic as a Microsoft subprocessor in the Microsoft 365 admin center under Copilot → Settings → AI providers. For commercial tenants outside those regions, Anthropic models are on by default. As of July 15, non-federal GCC customers can also opt in. New models appear automatically in the selector once enabled — no additional configuration per model is required.

What this means for IT leaders: Multi-model Copilot is now production reality, not a preview. Your users can choose Claude Opus 5 for a specific task and switch back to GPT-5.6 for the next one. This is a feature, not a problem — but it requires governance. Review your subprocessor settings, understand which user groups have access to which models, and ensure your data residency requirements accommodate Anthropic’s processing locations. The era of “which model is Copilot using?” being a hidden routing decision is ending; users will increasingly make that choice themselves.

2. Copilot Search for Outlook: Canceled

In a rare AI reversal, Microsoft confirmed on July 23 that it has canceled the Copilot Search integration for Outlook. The feature, which would have replaced Outlook’s classic search with a Copilot-powered interface, was walked back after significant user pushback. A Message Center update stated Microsoft had “decided not to move forward with this change” and apologized for the inconvenience.

The cancellation applies to Outlook for Windows, Android, iOS, and the web. Outlook Classic was never affected by the original plan.

This is notable not because Copilot Search was inherently bad, but because it represents Microsoft’s first major Copilot retreat. The pattern of replacing established, functional UI with AI-first interfaces is hitting a usability wall. Users rely on Outlook search for rapid, precise retrieval — keyword filters, sender sorting, date ranges — and a conversational Copilot interface, however sophisticated, adds friction to tasks that were already solved. The backlash sent a clear signal: AI should augment productivity workflows, not replace working tools for the sake of AI adoption metrics.

Simultaneously, Microsoft is removing Meeting Insights in Outlook starting mid-August 2026, replacing it with Microsoft 365 Copilot meeting preparation. Copilot in the Outlook editor is also rolling out for users with M365 Copilot licenses, initially on the web, new Windows client, and mobile — Mac and Outlook Classic will follow later in 2026.

What this means for IT leaders: The Copilot Search cancellation is a useful data point for adoption strategy. Pushing AI into every surface regardless of user need creates resistance. Communicate to your teams that Copilot is additive — it enhances search, drafting, and analysis — but Microsoft is learning (publicly) that not every feature needs an AI rewrite. Monitor user feedback on the Outlook editor integration and Meeting Insights replacement closely.

3. MCP 2026-07-28: The Protocol Goes Stateless

Today, July 28, the Model Context Protocol finalizes its largest specification update since launch. MCP — the open standard that connects AI agents to enterprise tools, data sources, and APIs — is going stateless, and the changes are significant for any organization building agent infrastructure.

The headline: the initialize handshake is gone. No more initialize/initialized round trip between client and server. Protocol version, client info, and capabilities now travel in _meta on every request. A new mandatory server/discover RPC lets clients fetch server capabilities on demand. The Mcp-Session-Id header is eliminated entirely — servers no longer retain client memory between requests, meaning any request can land on any server instance. This enables horizontal scaling behind ordinary load balancers, which was previously impossible with MCP’s stateful sessions.

Tasks are redesigned. The Tasks feature moves from core spec to an official extension (io.modelcontextprotocol/tasks). Blocking tasks/result is replaced by polling via tasks/get, and tasks/list is removed outright. This is a breaking change for any agent pipeline that relied on synchronous task completion.

Enterprise-Managed Authorization (EMA) is stable. Reached stable status on July 6, EMA lets central IT decide which MCP connections are allowed — enterprise-grade access control for AI agent integrations. Anthropic, Microsoft, and Okta have all implemented it. This is the governance layer that makes MCP viable for regulated environments.

Deprecations: Roots, Sampling, Logging, and HTTP+SSE transport all move to deprecated status with a minimum 12-month migration window through mid-2027. Full JSON Schema 2020-12 support is now mandatory. W3C Trace Context is required for distributed tracing.

Beta SDKs are available now: Python mcp v2.0.0b1, TypeScript v2, Go v1.7.0-pre.1, and C# v2.0.0-preview.1. Old clients interoperate with new servers through a fallback path, so production stays live past July 28 — but migration should happen within the deprecation window.

What this means for IT leaders: If your organization is building or deploying MCP-based agent integrations, this spec change is the most significant architectural shift since MCP launched. The stateless model is better for scale and resilience, but it requires SDK updates and testing. More importantly, EMA reaching stable means you can now enforce enterprise access control on MCP connections — treat this as a security baseline, not an optional feature. Review your agent integrations against the new spec and plan migration within the 12-month window.

4. Microsoft Agent 365: First Major Enterprise Deployment

Microsoft Agent 365 — the control plane for governing, monitoring, and securing AI agents at enterprise scale — reached general availability on May 1, 2026. This week, it landed its first publicly named major enterprise deployment: Manulife announced an expanded partnership with Microsoft on July 22 that includes deploying Agent 365 alongside expanding Microsoft 365 Copilot adoption to over 30,000 employees.

Agent 365 is not a product that builds agents — it’s the governance layer above them. It connects to Microsoft’s existing identity, compliance, and security stack: Entra Agent ID provides each agent with a managed identity, Microsoft Purview handles data governance, and Microsoft Defender provides threat protection. The value proposition is visibility — a single place to see every agent in your environment, what it can access, who owns it, and what it’s doing.

This matters because agent sprawl is already a real problem. Organizations have agents built in Copilot Studio, Azure AI Foundry, and third-party platforms, often without central tracking. Agent 365 turns these from invisible software into managed entities with sponsors, owners, and defined scopes. For Entra administrators, this also means new governance responsibilities: the July 24 Entra ID update explicitly frames AI agent security as an identity problem, recommending agent inventories, blueprint-based governance, and Conditional Access policies scoped to agent identities using Custom Security Attributes.

What this means for IT leaders: If you have more than a handful of agents in your environment — and if you’ve deployed Copilot Studio or Azure AI Foundry, you almost certainly do — Agent 365 is the governance tool you need. The Manulife deployment is a reference point for large-scale adoption. Start with agent inventory: Entra ID > Agents > Agents overview gives you visibility. Then define ownership: every agent needs a sponsor (business accountable) and an owner (technical responsible). Treat agents like any other nonhuman identity — because that’s exactly what they are.

5. Copilot Admin and Agent Updates

Several administrative and agent-focused updates landed this week:

Microsoft 365 Admin Agent is generally available as of July 22. Entra built-in administrators can now use natural language to manage Microsoft 365 admin center tasks through Copilot Chat — a meaningful productivity gain for helpdesk and operations teams.

Copilot agents now support multiple owners, announced July 24 and rolling out worldwide in August 2026. Teams can co-manage and edit agents collaboratively with equal permissions, improved sharing controls, and support for organization-wide publishing. This removes a significant bottleneck for organizations where a single agent owner created a bus-factor risk.

Admins can now block up to 1,000 specific websites from being used as web sources by Copilot. This web content control feature is available from mid-July 2026 and is off by default until configured. It provides granular control over what Copilot can reference when generating answers — useful for compliance, competitive sensitivity, or quality control.

Copilot mobile app gains a Record feature for voice notes and AI-generated transcripts, rolling out from late August 2026. Recordings are stored in OneDrive and governed by existing M365 security and compliance policies. This extends Copilot’s reach into mobile-first scenarios where typing is impractical.

A policy link setting for blocked Copilot Chat users is now available — admins can configure a custom URL that appears when users are blocked, directing them to company policy documentation. Small but important for user experience.

6. Mistral Integration and Model Catalog Expansion

Microsoft deepened its partnership with Mistral AI, integrating Mistral Medium 3.5 and OCR4 into Microsoft Foundry, with Medium 3.5 also available in Copilot Studio. Microsoft will also leverage Mistral’s European AI infrastructure built on NVIDIA Vera Rubin generation GPUs.

This is strategically significant for two reasons. First, it expands the model catalog available in Copilot Studio beyond Microsoft’s own MAI models and OpenAI/Anthropic options — giving enterprises more choice for building custom agents, especially for European workloads where data residency matters. Second, Mistral’s European infrastructure gives Microsoft a credible EU-based inference path, which could eventually help address the EU Data Boundary gap that currently blocks Anthropic models for European tenants by default.

7. Microsoft Q4 FY2026 Earnings: AI in the Spotlight

Microsoft reports Q4 FY2026 earnings on July 29 — the day after this post publishes. Wall Street consensus expects revenue of approximately $87.6 billion (up ~15% year-over-year) and EPS of ~$4.24 (up ~16%). Microsoft’s own guidance: $86.7B to $87.8B.

The AI metrics to watch:

  • AI annual revenue run rate: Crossed $37 billion in Q3 FY2026 (ending March 31), up 123% year-over-year. Q4 should show whether that acceleration is continuing or normalizing.
  • Microsoft 365 Copilot paid seats: Surpassed 20 million in Q3, growing 250% year-over-year. With Copilot now standard in Microsoft 365 Business plans (as of July 1), the seat count should jump significantly.
  • Azure AI / cloud growth: Q3 Azure grew 40% (39% constant currency). Guidance for Q4 implies 39-40% constant currency growth. Azure AI services are the primary driver.
  • AI infrastructure spending: Microsoft guided above $40 billion in AI capex for Q4 alone. Watch for forward guidance on FY2027 capex — this is the number that will determine whether Microsoft’s AI investment cycle is accelerating or plateauing.

What this means for IT leaders: The earnings report matters less for the headline numbers and more for the forward guidance. Microsoft’s AI capex trajectory signals how aggressively they’ll push model development, datacenter expansion, and Copilot feature velocity in the next fiscal year. If Copilot seat counts show a significant jump from the Business plan bundling, expect Microsoft to accelerate feature delivery — and pricing pressure on standalone Copilot licenses.

Strategic Next Steps

  1. Configure Claude Opus 5 access — review your subprocessor settings in the Microsoft 365 admin center. Decide whether Opus 5 should be available to all users or restricted to specific groups handling complex analysis tasks
  2. Inventory your agents — use Entra ID’s agent overview to identify every agent in your environment. Assign sponsors and owners to each one. If you don’t have Agent 365 licensing, budget for it in FY2027 planning
  3. Plan MCP migration — if you have MCP-based integrations, schedule SDK upgrades to the 2026-07-28 spec within the 12-month deprecation window. Test stateless behavior behind your existing load balancers
  4. Enable web content controls — configure blocked site lists in Copilot settings to prevent sensitive or unreliable sources from informing AI-generated content
  5. Prepare for multi-owner agents — document your agent ownership model before the August rollout. Identify which agents should be org-wide published versus team-scoped
  6. Watch the earnings call — July 29, after market close. Track AI revenue run rate, Copilot seat growth, and FY2027 capex guidance. These numbers will shape Microsoft’s product roadmap velocity for the next 12 months

The Copilot story this week is about boundaries — where AI belongs (the model selector, agent governance, admin tools) and where it doesn’t (replacing Outlook search). Microsoft is learning, publicly and sometimes painfully, that AI adoption isn’t about coverage metrics. It’s about giving users the right tool for each task, whether that tool runs on GPT-5.6, Claude Opus 5, or — as Microsoft confirmed this week — classic keyword search. The organizations that succeed with Copilot will be the ones that learn the same lesson: AI is a capability to be deployed selectively, not a feature to be applied universally.

Follow along at https://x.com/kkaminski for daily Copilot and Microsoft AI analysis.