The New Copilot: Home, Code, and Autopilot
On September 25, Microsoft announced the biggest redesign of Microsoft 365 Copilot since launch. The product is now organized around three destinations. Home merges quick Chat with the longer-running assignments handled by Copilot Cowork, and Microsoft plans automatic routing so users can describe an outcome and let Copilot pick the right capability. Code lets employees describe an app, dashboard, automation, or desktop widget and have Copilot build it. Autopilot is a persistent agent that keeps working on an assignment after its human colleagues sign off.
Home also brings Office in Copilot: Word and PowerPoint editing and collaboration join Excel inside the Copilot surface. Users can create or revise the same files their colleagues are editing in the desktop apps, with tracked changes and syncing preserved. A planned feature called Today gathers signals from email, calendar, chats, and assignments, then drafts a follow-up or proposes a calendar change for you to approve. Today enters private preview in October and is planned for Outlook and Teams later.
The strategic shift is that Copilot is no longer a chat box bolted onto Office. It is now the place where people start work, delegate it, and build their own tools. Rollout is staggered: Home and Office integration reach Frontier participants over the coming weeks, Code reaches Frontier at the end of September, and a Code preview for Microsoft 365 Premium and Pro subscribers is planned for later in 2026. Shared @Copilot interactions in Teams also enter private preview at month-end.
What to do: Confirm whether your tenant is enrolled in Frontier and decide deliberately who gets early access. Pick a small pilot group, give them a short brief on the three destinations, and collect feedback on which one actually changes their day before you expand access.
Scout Is Now Autopilot, and It Is an Identity in Your Tenant
Microsoft Scout, unveiled at Build in June, has been officially renamed Autopilot. You give an Autopilot a name, a role, and an objective, and it runs with its own Entra identity, memory, compute environment, and workspace inside your Microsoft 365 boundary. It can follow Teams channels and threads, handle recurring responsibilities, pick up a project after several days without new instructions, and be @mentioned in Teams, Outlook, and documents like a coworker. In Microsoft's demo, an Autopilot named "Dot" monitored Black Friday prep across email, Teams, Dynamics 365 inventory, and spreadsheets, spotted a shipment problem affecting 18 stores, and pulled the right people together to fix it.
Autopilot expands to private preview at the end of September. It is not generally available, and Microsoft has not committed to a GA date. Earlier reporting targeted October, but the September 25 materials leave the date open. Satya Nadella said on the Sources podcast that trust is the central challenge for autonomous Copilot: "Every agent has to have an identity. Everything it does needs to be observed." Agent 365 provides the permissions, audit records, and admin controls. Microsoft says credentials are scoped per task and redacted from logs, destinations are allow-listed, sensitive actions can require human sign-off, and Purview labels and DLP apply before anything is sent or written.
Early reporting surfaced a few provisioning details admins need to know. An Autopilot's service identity does not automatically include a mailbox, calendar, or OneDrive. Those require a separate, optional agent user account plus licenses, currently limited to Frontier tenants. Each identity blueprint needs a human sponsor, permissions still require admin consent, and Conditional Access is the relevant control for agent identities rather than MFA. Also note the naming collision: Copilot's Autopilot has nothing to do with Windows Autopilot device provisioning.
What to do: Before joining the preview, write down how you would govern one Autopilot the way you would a privileged service account: its sponsor, what it can reach, which actions need human approval, which Purview policies apply, and how you would shut it off. Also update runbooks and ticket templates so "Autopilot" is never ambiguous.
Code and Copilot Managed Runtime: Citizen-Built Apps Become Governed Software
Copilot Code builds on the App Builder experience introduced in late 2025, but removes its biggest limitations. Code shares its foundation with GitHub Copilot and draws on organizational context and connected services. Generated apps can now call external services, and their code is editable and version-controlled in Git rather than locked away. Professional developers keep using GitHub Copilot, while business users build through the Copilot interface.
Under it all sits Copilot Managed Runtime, now in public preview. This Microsoft-operated hosting platform runs apps inside your Microsoft 365 boundary, whether they were built in Copilot Studio, Cowork, or Code. An SDK and CLI open the same deployment path to compatible third-party tools and pro developers. Entra handles identity, admins control permitted data connections and endpoints, and the admin center gets a consolidated app inventory showing access, activity, health, and policy. Microsoft's pitch is that hosting and sharing an app should be "as simple as saving and sharing a Word document."
That convenience is exactly why governance matters. When any employee can produce a working, hosted, data-connected app in minutes, app sprawl stops being theoretical. Reported defaults: CLI app creation is disabled, the eligible Copilot Studio preview path is enabled, and Cowork creation depends on Frontier onboarding. Build-time and runtime credit policies can be managed separately.
What to do: Review Managed Runtime defaults in the admin center now, while the service is still in preview. Decide which data connectors citizen-built apps can use, who can publish, and who owns an app once it has more than a few users. An app nobody owns is a future incident.
Pricing and FinOps: Copilot Credits Become the Meter for Agentic Work
Microsoft clarified that everyday use (Chat and the Office apps) stays inside the $30 per user/month enterprise subscription, and it did not raise that price. Cowork, Code, and Autopilot, plus advanced SharePoint capabilities and premium models, bill on top of the subscription in Copilot Credits. Microsoft's modeled example shows how quickly that adds up: 20 everyday tasks plus five complex Cowork assignments came to roughly $69–$73 per user per month depending on the model. That is illustrative, not a price list, and detailed rates for Code and Autopilot still have not been published.
The FinOps tooling expanded to match. Agent 365 cost management now covers Autopilot, Code, and Managed Runtime, and Copilot Studio agents follow in October. New this week is group-specific model access, starting with Cowork. Admins can restrict which model families a group can use, which also constrains automatic model routing. Users can see their consumed credits, remaining allowance, and history. Spending policies and balances are manageable through Microsoft Graph APIs, and requests for more credits can flow into existing approval processes. Metered services stay off until an admin creates a spending policy, but note that new policies enable "Auto-apply new services" by default.
This fits the trend we covered last week: seats for everyday assistance, meters for autonomous work. The mental model for finance teams is closer to cloud compute than to software licensing. Budgets, alerts, and showback by department are about to become part of running Copilot.
What to do: Audit your existing spending policies for the "Auto-apply new services" setting before Code and Autopilot land in your tenant, and turn it off where you want an explicit decision. Then set group-level model access so premium reasoning models go to the teams that actually need them.
Cowork and Copilot Studio: Skills, Models, and Plugins
Several builder-facing updates shipped alongside the relaunch. Declarative agents can now carry packaged skills: a folder with a SKILL.md instruction file, added through Agent Builder (up to eight per agent, 50 MB per package) or the Agents Toolkit. Cowork added shareable skills, a rebuilt Notebooks experience, a reasoning effort slider that trades speed and cost for depth, a model picker built into the prompt box (split into GPT and Claude sections), and a renamed Automations area for scheduled prompts. Brand kits gained a strict brand-adherence toggle that limits Copilot to your template's layout slides.
Model choice keeps widening. GPT-6 Astra and Claude Fable 5.1 are now selectable in Cowork and Copilot Studio, with admins controlling availability in the Microsoft 365 admin center. Nadella confirmed that Copilot will stay multi-model even as Microsoft's own models power parts of Auto mode. On context and extensibility: Fabric IQ integration is generally available for Chat and Cowork, Work IQ gains Power Platform and Dynamics 365 connections in public preview, and a shared plugin registry gives admins one place to approve Microsoft, partner, and internal extensions. Nine new partner plugins also arrived, including Harvey, LSEG, Miro, monday.com, Moody's, and Morningstar.
The quiet headline is skills. A SKILL.md package is portable, reviewable institutional knowledge: how your team writes a proposal, runs a close, or triages a ticket, captured once and reused by every agent that needs it. Combined with Copilot Studio now creating an Entra Agent ID automatically for every new agent, the building blocks for well-governed, well-instructed agents are all in place. What most organizations lack is the practice of writing good ones.
What to do: Identify three repeatable processes your teams explain over and over, and write each as a skill: clear steps, examples of good output, and explicit boundaries. Store them in source control, review them like code, and attach them to your agents. Well-written skills are the cheapest way to make every agent, and every credit, more effective.
The Bigger Picture: Delegation Is Now the Product
Put this week together and Microsoft's direction is clear. Copilot has moved from answering questions (Chat), to completing tasks (Cowork), to building tools (Code), to owning ongoing responsibilities (Autopilot). Each step hands more responsibility to the agent, and each one is metered, identity-bound, and auditable by design. Anthropic's move earlier this month to merge Claude chat and Cowork points the same way across the industry: the assistant is becoming a delegate.
The limiting factor is no longer capability. It is whether an organization knows how to delegate well: how to scope an objective, set boundaries, write the instructions and skills an agent needs, choose the right model for the cost, and review the result. Those are human skills, and they can be taught. The organizations getting value from Autopilot next year will be the ones building that muscle during this preview window.
What to do: Treat the next 60 days as a readiness sprint. Train a core group on prompting, skill authoring, and agent governance. Stand up spending policies and an agent register. Run one supervised Autopilot or Cowork workflow end to end. When general availability arrives, you want a playbook, not a pilot.
Sources
- Introducing the new Copilot with Home, Code and Autopilot — The Official Microsoft Blog
- Microsoft revamps its Copilot AI with a persistent Autopilot agent and hosting for AI-generated apps — VentureBeat
- Microsoft unveils Copilot super app, targeting business users with AI agents — Fortune
- Microsoft 365 Copilot: Autopilot Replaces Scout in Private Preview — WindowsForum
- Satya Nadella on Microsoft's agent bet and AI's trust problem — Sources
- Evolution of the Copilot pricing model — Microsoft Tech Community
- Manage Copilot Credits with usage-based billing — Microsoft Learn
- NEW Copilot Updates Are Here — September 2026 — M365 Copilot Connection
- What's New in Microsoft 365 Copilot: September 2026 — A Guide to Cloud & AI
- What's new in Copilot Studio — Microsoft Learn
- Release Notes for Microsoft 365 Copilot — Microsoft Learn