Microsoft's 25 September rebuild turns Copilot from a chat box into a platform for long-running agents — and splits the commercial model in two. Chat stays on a fixed seat licence; Cowork, Code, Autopilot and frontier models move to usage-based billing. Here's what that actually breaks for teams running AI in production.
Last Updated: September 27, 2026
Microsoft rebuilt Copilot around three surfaces — Home, Code and Autopilot — moving it from a chat box to a host for long-running agents. Home and Code roll out through the Frontier program over the coming weeks; Autopilot hits private preview at the end of September. The bigger story for anyone integrating AI is commercial: everyday chat stays on the fixed seat licence, while Cowork, Code, Autopilot and frontier models shift to usage-based billing. That turns AI from a headcount line item into a variable cost that needs governance, and it makes an opaque model router a budget risk as well as a quality risk.
What Microsoft actually shipped
On 25 September Microsoft announced what it calls the new Copilot, structured around three capabilities. Per the official announcement, Home is the new starting point where Chat and Cowork come together, with Word, Excel and PowerPoint built into the experience through Office in Copilot. Code lets people build their own solutions with tools to run them safely, powered by the same underlying technology as GitHub Copilot. And Autopilot is described as a persistent, proactive and personal agent that keeps working when you are not.
The rollout is staged. Microsoft's EMEA newsroom confirms Home and Code arrive gradually through the Microsoft Frontier program over the coming weeks, with Autopilot expanding to private preview at the end of the month, alongside previews of Today in Home and @Copilot in Teams.
Underneath the three surfaces sits the part that integration teams should read twice. Coverage of the launch notes a Copilot Managed Runtime that provides infrastructure for running code inside an organisation's Microsoft 365 environment — supporting apps built with Cowork, Code and Copilot Studio as well as third-party and pro-code development — plus Fabric IQ pulling Power BI semantic models into Copilot, Dynamics 365 and Power Platform integrations, and a plugin registry where Microsoft, partner and custom plugins are managed from one catalogue.
Read that as a stack, not a feature list: an execution runtime, a data layer, a plugin registry and an agent surface, all inside the tenant that already holds your identity, your mail and your files. That is the same shape as a bespoke agent platform, shipped through the largest enterprise software distribution channel in existence.
The meter is the story
Microsoft split Copilot's commercial model along the line between answering and doing. The user subscription licence covers everyday AI — quick answers, first drafts, summaries and analysis — across Word, Excel, PowerPoint, Outlook and Teams at a fixed cost, with a router called Auto at its heart. Agentic work goes elsewhere: Microsoft states that Cowork, Code, Autopilot, new long-running agentic capabilities and frontier models including Astra and Fable all run on usage-based billing, with new FinOps for AI capabilities to manage spend.
Auto weighs accuracy, speed and cost on each request to route to the model best suited for the job.
— Microsoft, Introducing the new Copilot
That sentence is doing enormous work. A router optimising accuracy, speed and cost inside a fixed-price licence has a structural incentive to be thrifty on the margin. Ethan Mollick has made the point repeatedly in public that routers can quietly hand a genuinely hard task to a weaker model, and the user never sees the handoff — they just see a slightly worse answer they have no way to attribute. When the fixed-cost tier routes automatically and the expensive tier is the one you pay per unit for, quality variance and billing variance become the same problem.
| Dimension | Fixed seat licence (USL) | Usage-based billing (UBB) |
|---|---|---|
| What it covers | Chat plus Copilot in Word, Excel, PowerPoint, Outlook and Teams | Cowork, Code, Autopilot, long-running agents, frontier models |
| Model choice | Auto routing, with model selection available | You choose the model and match cost to value |
| Cost behaviour | Predictable per head | Variable, scales with agent activity |
| What finance needs | A seat count | Budgets, alerts, attribution — FinOps for AI |
| Main failure mode | Silent quality drift from routing | Silent spend drift from autonomous work |
What this changes for teams integrating AI
Three practical shifts, in order of how soon they will land on your desk.
How the buying conversation changes
A persistent agent that already has the user's identity, mail and calendar beats a bolt-on integration for a large class of routine delegation work.
Generic summarise-and-draft orchestration stops being billable. What remains valuable is domain logic, systems your tenant cannot reach, evaluation and guardrails.
Once agents bill by usage, an architecture decision is a cost decision. Whoever signs the invoice now has an opinion on your loop design.
The uncomfortable one is the second. If your automation roadmap is largely "agent reads the thread, drafts the reply, files the doc," Microsoft just shipped that natively to everyone in the Frontier program. The work that survives is the work that touches systems Copilot does not see, enforces rules Copilot cannot be trusted to enforce, or proves — with evidence — that the output was correct.
Treating usage-based agents like seats
Seat licences make AI cost linear in headcount, so nobody instruments it. Long-running agents make cost linear in activity — and a retry loop, a chatty tool call or an agent politely re-reading a 200-page attachment every morning are all invisible until the invoice.
Trusting the router without measuring it
Automatic routing is genuinely useful for the long tail of cheap requests. It is dangerous for the small set of requests where a wrong answer is expensive — contract terms, financial reconciliation, anything regulated — because the downgrade is invisible at the point of use.
What the announcement is carefully not saying
Four gaps worth naming. First, pricing specifics: Microsoft says agentic work moves to usage-based billing and that pricing evolves alongside capability, but the announcement is a model, not a rate card. Until you see unit economics for a real workload, any TCO comparison against a bespoke build is guesswork.
Second, availability. Home and Code are gated behind the Frontier program and Autopilot is in private preview — this is not something you can plan a Q4 rollout around for a general tenant.
Third, autonomy and accountability. An agent that keeps working when the user is away will, at some point, act on stale context or a misread instruction while nobody is watching. The announcement describes capability; it does not describe what your audit trail looks like when an autonomous action needs to be explained to a regulator or a customer.
Fourth, the citizen-developer surface. Code lowering the barrier so non-developers can build apps and automations inside the tenant is a real productivity unlock and a real governance problem — sandboxed execution constrains blast radius, it does not review business logic. Expect a wave of unreviewed internal tools that nobody can maintain, which is the SharePoint-and-Access story with a better interface.
What to do in the next 30 days
Inventory the agent work you have already built or scoped. Mark each item as "Copilot will do this natively" or "Copilot structurally cannot do this." The second list is your actual roadmap.
Stand up cost governance before enablement, not after. Budgets, alerts, owners, and a policy on which teams may run long-running agents at all.
Build the evaluation set. Twenty to fifty representative high-stakes prompts with known-good answers, scored monthly against whatever the router hands you.
Write the autonomy policy: what an agent may do unattended, what requires approval, how actions are logged, and who is accountable when it is wrong.
The strategic read is straightforward. Microsoft has decided that agents, not chat, are the product, and that agents should be metered. If you are integrating AI into a business, your differentiation just moved from "we can orchestrate an LLM" to "we can make autonomous systems correct, governed and affordable." That is a harder and more durable place to stand. If you want help drawing that line for your own stack, start here or look at how we scope this in engagements.
Frequently Asked Questions
What are Home, Code and Autopilot in the new Copilot?
Microsoft describes Home as the new starting point where Chat and Cowork come together, with Word, Excel and PowerPoint built in through Office in Copilot. Code lets people build their own solutions with tools to run them safely, powered by the same underlying technology as GitHub Copilot. Autopilot is a persistent, proactive and personal agent that keeps working when you are not at your machine.
When can my organisation use these features?
Not immediately. Home and Code begin rolling out gradually through the Microsoft Frontier program over the coming weeks, and Autopilot expands to private preview at the end of September 2026. General availability dates were not given. Treat this as a signal for planning and architecture decisions rather than something you can schedule a tenant-wide deployment around this quarter.
What exactly changes about Copilot pricing?
The model splits in two. Everyday chat, plus Copilot across Word, Excel, PowerPoint, Outlook and Teams, stays on the fixed-cost user subscription licence with automatic model routing. Cowork, Code, Autopilot, new long-running agentic capabilities and frontier models such as Astra and Fable move to usage-based billing, where you choose models and match cost to value. Microsoft is shipping FinOps for AI capabilities to manage that spend.
Why is automatic model routing a risk?
Because the trade-off is invisible to the person using it. Microsoft says Auto weighs accuracy, speed and cost on each request. That is sensible for the long tail, but on genuinely hard tasks a router can select a weaker model and the user simply receives a slightly worse answer with no indication why. For high-stakes work, pin the model explicitly and maintain a fixed evaluation set so you can detect quality drift rather than guess at it.
Does this make custom agent development pointless?
No, but it changes what is worth building. Generic summarise-draft-file orchestration inside Microsoft 365 is becoming a platform feature, so building it yourself is hard to justify. Value moves to work Copilot structurally cannot do: systems outside the tenant, domain-specific logic, deterministic guardrails, evaluation harnesses, and cost governance across a mixed estate of vendor and in-house agents.
Research digest
AI Research Briefing
Honest insights on AI agents, Small Language Models, and local RAG. No hype. Only when we have something worth sending.
- No hype, just measurable outcomes
- Read by 2,400+ engineers
- Unsubscribe anytime
