Background: Why Wave 3 matters
Microsoft has been layering generative AI into Office apps for more than a year, but until now most features operated as one-off assistants that answer prompts, propose text, or generate slides. Wave 3 is the pivot from assistant-as-adviser to assistant-as-executor: agents that plan, act over time, reason across documents and apps, and return editable, app-native artifacts rather than blobs of exported text.
This is not just a UI update. It’s a shift in architecture and product strategy that embraces three principles:
- Embedded, app-native outputs — AI edits and creates inside Word, Excel, PowerPoint, and Outlook rather than producing external files that must be re-integrated.
- Multi-model intelligence — Microsoft is no longer single-vendor in Copilot’s brain; it will route tasks to the model best suited for the job.
- Operational control — the company expects agents to scale and has built Agent 365 to give IT and security teams governance tools similar to how they manage users today.
Taken together, Wave 3 aims to turn Copilot from an augmentation tool into a platform for automating complex, multi-step office work at enterprise scale.
What’s new in Wave 3: feature overview
Copilot Cowork — Anthropic’s Cowork inside Microsoft 365
Copilot Cowork is the marquee capability of Wave 3: a long‑running, multi‑step agent designed to execute complex workflows end-to-end. Microsoft says Cowork can orchestrate tasks such as preparing a customer meeting by assembling slides, pulling financials, emailing stakeholders, and scheduling prep time — all from a single request while keeping the user informed and able to intervene.
Key characteristics:
- Built with Anthropic’s agentic technology and the same “agentic harness” concept used by Claude Cowork, but integrated into Microsoft 365.
- Runs within a customer’s Microsoft 365 tenant and cloud environment so enterprise controls, audit logging, and governance apply.
- Uses Work IQ (Microsoft’s context layer) to reason over emails, files, meetings, and relationships when executing tasks.
- Initially available to a limited set of customers as a research preview and through Microsoft’s Frontier program in March 2026.
Why this matters: Cowork promises to move work off the user’s to-do list and into an automated process you can oversee. For enterprises, that means potential time savings at scale — but also greater need for governance and monitoring.
Agents in Word, Excel, PowerPoint, and Outlook
Wave 3 makes agentic functionality first-class inside Office apps. Microsoft describes these as no longer a separate “mode” but as embedded capabilities that produce app-native artifacts — e.g., a Word document with real styles and tracked changes, an Excel workbook with formulas and pivot tables, or a PowerPoint deck that respects your design system.
Highlights by app:
- Word: major rewrites, structural edits, document polishing and alignment to corporate style guides using contextual signals (meetings, files, emails).
- Excel: multi-step scenario modeling, automated formula repair and generation, pivot and dashboard creation that recalculates as source data changes.
- PowerPoint: generate decks that match brand kits, layouts and object styles, with the capacity to convert chat intent into finished slides.
- Outlook: chat-driven drafting, sending, and scheduling — including the ability to have Copilot send emails from within chat contexts.
Rollout: Excel and Word agent experiences are generally available; PowerPoint and Outlook are rolling out over the coming months.
Copilot Chat enhancements and model choice
Copilot Chat becomes the central entry point for chat-first creation and execution. Wave 3 expands chat’s role so you can start in conversation and then have the system create final artifacts inside the relevant app.
Crucially, Microsoft expands model choice: Claude (Anthropic) is now available in mainline Copilot Chat via the Frontier program alongside Microsoft’s OpenAI-powered models. The platform will pick the “right model for the job” by default, while also offering users and admins the option to select a specific model in many scenarios.
Agent 365 — governance and control plane
As agents proliferate, Microsoft introduces Agent 365, a control plane to observe, secure, and govern agents across an organization. It extends familiar management approaches — using Microsoft Admin Center, Entra (identity), Defender, and Purview — to handle agents the way IT handles employees.
Planned availability and pricing:
- Agent 365: generally available May 1, 2026, priced at $15 per user per month.
- Microsoft 365 E7 (Frontier Suite): available May 1, 2026, priced at $99 per user per month (bundles E5 + Copilot + Agent 365 + security features).
How Copilot Cowork actually works (and where it differs from Claude Cowork)
At a technical level, Copilot Cowork imports the agentic design pattern Anthropic demonstrated with Claude Cowork — i.e., models that can plan, invoke tools, operate asynchronously, and produce persistent outputs. Microsoft’s implementation layers three differentiators on top:
- Tenant-native execution — Cowork tasks execute within a customer’s Microsoft 365 tenant and cloud environment, enabling enterprise data protection, compliance, and identity integration.
- Work IQ grounding — agents have access to a persistent context built from your files, meetings, and communications so their outputs are grounded in the organization’s current state.
- Observability and governance — actions are auditable, reviewable, and can be paused or steered by users and admins.
From a user perspective, Cowork looks like a single command that kicks off a visible, multi-step run. Behind the scenes, the agent decomposes the job into tasks, loops through app integrations, pulls in data, and returns app-native artifacts the user can edit.
Limitations to watch:
- Copilot Cowork in Microsoft is cloud-run, not local execution — a deliberate tradeoff for centralized governance but a difference from some Claude Cowork deployments that emphasized local processing.
- Availability starts as a research preview in Frontier; typical enterprise customers should expect controlled pilots before organization-wide rollout.
Multi-model intelligence: benefits and complexity
Microsoft’s explicit embrace of multi-model intelligence is strategic and pragmatic. Rather than betting exclusively on a single provider, Copilot will route problems to the model best suited for the task — OpenAI’s latest generative reasoning models for some tasks, and Anthropic’s Claude for others.
Benefits:
- Access to a broader set of reasoning strengths (different models excel at different tasks).
- Faster adoption of breakthrough innovations from various model vendors without rewriting product plumbing.
- Flexibility for enterprises that may prefer particular vendor properties for regulatory, accuracy, or compatibility reasons.
Operational complexity:
- Model selection introduces additional decision points for IT: Which model is default? Which models are allowed? How do you control data flows when multiple vendors are involved?
- Billing and telemetry become more complex when different models impose separate metering or cost structures behind the scenes.
- Testing and validation: model behavior can vary significantly by provider; organizations must expand QA and compliance reviews to account for the mix.
On balance, multi-model intelligence is a clear product advantage, but it raises non-trivial governance and cost-management needs that IT teams must plan for.
Security, privacy, and governance: the critical control plane
Microsoft positions Wave 3 as intelligence plus trust — and the trust piece is where enterprise IT will spend most of their energy. Key guardrails Microsoft highlights include:
- Enterprise Data Protection (EDP) and enforcement of tenant permissions and sensitivity labels so agents only process content when allowed.
- Saving outputs to OneDrive/SharePoint with tenant-level controls; minimizing version sprawl by keeping work inside managed storage.
- Agent 365 for visibility, governance, and lifecycle control of agents.
- Integration with existing security services — Defender, Entra, Intune, and Purview — to extend familiar controls to agents.
Practical risks and gaps to plan for:
- Agent sprawl: Microsoft reports rapid agent creation in previews; without rules and quotas, the number of active agents could balloon and create management headaches.
- Phishing and social engineering against agents: if agents send/receive emails or interact with external systems, they can be the target of malicious inputs — Microsoft explicitly warns agents can be phished just like humans.
- Data residency and third-party models: when an agent uses an external model vendor, organizations must validate where data is processed and whether that meets regulatory obligations.
- Explainability and auditing: automated, multi-step agent actions need clear, tamper-resistant audit trails and human-readable rationales for compliance.
Recommendation for IT: treat agents like privileged service accounts. Inventory them, require approval for production agents, apply least privileges, and bake agent controls into incident response playbooks.
Real-world benefits: what teams will actually gain
Wave 3’s most immediate wins are practical and often mundane — the kinds of time savings that compound across teams.
Productivity scenarios that improve immediately:
- Sales teams generate tailored decks and one-click send-to-prospect workflows, shaving hours from prep time.
- Finance teams run scenario modeling in Excel and get native formulas, explanations, and pivot tables rather than static reports.
- HR and legal teams draft and iterate policies or offer letters grounded in the organization’s current documents and precedent.
- Managers get meeting prep packs assembled automatically from recent files, emails, and transcripts.
What makes these wins repeatable is the move to app-native outputs: generated documents behave like any other corporate file (editable, versioned, auditable) instead of being throw-away artifacts.
Quantifying ROI: Microsoft claims substantial adoption growth in Copilot seats and agent deployments, but enterprise ROI will vary by process complexity, data maturity, and governance readiness. Pilot with high-value, repeatable workflows and measure time-to-completion, error reduction, and rework avoided.
For admins and IT leaders: rollout checklist
If your organization is evaluating Wave 3, start with these practical steps:
- Join Frontier or request early access to test Copilot Cowork and model choices in a controlled environment.
- Map high-value workflows (sales decks, budgeting, compliance reporting) suitable for agent automation.
- Define policies for model selection — which vendors are permitted and under what circumstances.
- Set up Agent 365 governance ahead of scale: agent approval workflows, quotas, monitoring dashboards, and alerting.
- Update data classification and sensitivity labels so agents process only what is permitted.
- Run red-team scenarios for agent-targeted phishing and malicious prompts to test resilience and response playbooks.
- Budget for cost-modeling: agent usage and multi-model routing may introduce new metering and compute spend.
These are not optional. Agents change the operational envelope of productivity software; the difference between safe pilot and corporate incident is often in early governance.
The competitive and strategic landscape
Wave 3 signals Microsoft’s choice to be a platform for an ecosystem of models rather than a single-model vendor. That has immediate market implications:
- It reduces vendor lock-in and gives Microsoft flexibility to integrate the “best” model features as they appear.
- It puts Microsoft in direct competitive conversation with Anthropic, OpenAI, and enterprise players (Salesforce, Google) who are building agent or task-focused products.
- It raises the bar for incumbents: traditional SaaS features are now threatened by agents that can replace or automate parts of the software value chain.
For Anthropic, the collaboration is both validation and accommodation: Microsoft integrates Claude’s tech into a massive enterprise footprint while running it under Microsoft governance. For customers, that can be attractive — Anthropic’s strengths delivered inside Microsoft’s security posture — but it also concentrates more responsibility on Microsoft to manage third-party model behavior and vendor risk.
Known limitations and remaining questions
Wave 3 is not a finished product; it arrives with caveats and open questions enterprises must respect:
- Availability varies: many features begin in Frontier previews and roll out regionally over months. Some markets may see delayed support.
- Model behavior consistency: different models produce different styles and may disagree on facts; testing and guardrails are necessary.
- Regulatory and compliance unknowns: for highly regulated industries, the third-party model processing and telemetry need careful legal review.
- Cost and licensing tradeoffs: Microsoft’s new bundles (Agent 365 at $15/user/month, M365 E7 at $99/user/month) create choices: per-user pricing remains prominent, but heavier agent usage could push customers toward different consumption models over time.
- Vendor dependencies: while multi-model is an advantage, it also means organizations must manage relationships and technical dependencies across multiple AI providers.
Flagging unverifiable claims: Microsoft’s internal adoption statistics and forward-looking agent counts come from company briefings; treat those as vendor-reported metrics until independent audits or third-party studies corroborate them.
Bottom line: who wins, who needs caution
Wave 3 is an important evolution for Microsoft 365 Copilot. For knowledge workers and teams burdened by repetitive, multi-step processes, Copilot Cowork and embedded agents have the potential to deliver measurable productivity gains. The combination of model choice and a governance-first approach is smart product strategy: it gives customers flexibility while promising enterprise controls.
That said, the shift to agents changes the operational dynamics of IT. The most successful adopters will be those that:
- pilot with high-value workflows,
- treat agents as first-class, governed digital workers,
- invest in monitoring, and
- update policies and incident response for agent-specific threats.
If you’re an IT leader, treat Wave 3 as the beginning of a new administrative lifecycle — not a typical feature rollout. If you’re a user, expect to offload more of the repetitive, low-creativity work to agents, but keep a close eye on outputs and retain final sign-off on critical artifacts.
Practical recommendations (fast checklist)
- For CIOs and CISOs: start Agent 365 planning now; set acceptance criteria for agents and establish an approval board.
- For IT admins: enable tenant-level sensitivity labels and OneDrive/SharePoint controls before enabling Cowork pilots.
- For compliance teams: document where models process data and update vendor contracts accordingly.
- For procurement: model forecasted agent usage and negotiate bundled E7 pricing if forecasts justify it.
- For team leads: identify two high-impact workflows to pilot Cowork (e.g., monthly reporting, customer-prep packs) and measure cycle-time improvements.
Wave 3 of Microsoft 365 Copilot is a consequential move: it converts Copilot from an assistant into a procedural workforce that lives inside the apps enterprises use every day. That opens exciting productivity possibilities — and it raises governance, security, and cost-management demands that must be met if those possibilities are to be realized safely. For organizations willing to invest in governance and phased pilots, the payoff could be significant; for those that rush without controls, the operational and compliance risks could be painful. The next several months will show whether agentic productivity becomes an incremental efficiency or a foundational change in how knowledge work is delivered.
