Microsoft's New Copilot Agents: What Home, Code and Autopilot Mean for Your Business

Microsoft has just given Copilot its biggest shake-up since the product launched. On 25 September the company announced that Copilot is being restructured around three parts: Home, Code and Autopilot. The familiar chat assistant is still in there, but it is now only one piece of a much larger platform built around Copilot agents that work on their own.

The interesting part is not the new names. It is what sits behind them. Autopilot agents get their own cloud computer, their own identity inside your organisation, even their own email address. They keep working after you close your laptop. And Microsoft is changing how it charges for all of this, moving the agentic parts of Copilot from flat licenses to usage-based billing.

For business owners and managers, this announcement touches two sensitive areas at once: how your team works and how your software budget behaves. Both deserve a careful look, so let's sink in.

What Microsoft actually announced

The redesigned Copilot is organised into three areas. Home combines the existing Chat and Cowork features into one interface, and it now includes what Microsoft calls Office in Copilot. You can ask for a budget or a presentation and get back a real, editable Excel workbook or PowerPoint file, not just a wall of text in a chat window. A "Today" dashboard is planned for October, pulling your email, calendar, Teams messages and tasks into a single view.

Code is the second area, and it lets people who are not developers build applications, dashboards and automations by describing them in plain language. It is built on GitHub Copilot technology and runs inside something called the Copilot Managed Runtime, which we will come back to.

The third and most ambitious area is Autopilot, a persistent autonomous agent that Microsoft previously ran under the codename Scout. Home and Code are rolling out through Microsoft's Frontier early-access programme over the coming weeks, while Autopilot moves into private preview at the end of September. The Managed Runtime entered public preview on the day of the announcement.

Autopilot is an employee that never logs off

The clearest way to understand Autopilot is to stop thinking about chatbots. Each Autopilot instance receives its own dedicated cloud resources: a computer, a workspace, storage, and an identity inside your Microsoft tenant, complete with an email address. You trigger it with an @mention in Teams, Outlook or a document, and it carries on working whether you are online or not.

Microsoft's own example is telling. A single Autopilot agent can run a complete supplier review, from scheduling the meetings through chasing stakeholders for their input, without anyone prompting it along the way. It can watch a Teams channel and follow up on threads that would otherwise go quiet. It uses what Microsoft calls Microsoft IQ to learn how your organisation actually operates, and the company says permissions, auditing and governance are built in.

That last sentence is doing a lot of work, and we would treat it as a claim to verify rather than a fact. But the direction is unmistakable. Microsoft is no longer selling an assistant that answers questions. It is selling a digital worker that holds responsibilities.

Code puts app-building in everyone's hands, with guardrails

Copilot Code is Microsoft's answer to a question many companies have been asking: if AI can write software, why do we still have a backlog of small internal tools nobody has time to build? With Code, a manager can describe a dashboard or an approval workflow and get a working application, no developer required.

What makes this more serious than earlier no-code experiments is the Copilot Managed Runtime. Apps built with Code run in a sandboxed environment inside your own tenant, with Microsoft Entra identity integration, organisational policies enforced on connectors and data sources, and automatic versioning through Git. Administrators get an inventory of every app with visibility into access, usage and health. Code is expected to reach Microsoft 365 Premium and Pro subscribers later in 2026.

For businesses, this is genuinely useful for the long tail of internal tooling. It is not, in our experience, a replacement for properly engineered software. Anything customer-facing, anything that needs to scale, integrate deeply or carry your brand still belongs in the hands of a development team. The honest way to look at Code is as a filter: it clears the small stuff so your development budget goes where it matters.

What the new Copilot agents will cost

Buried under the product news is a billing change that will affect budgets more than any feature. The standard Copilot user subscription now covers everyday chat and Office integration. Everything agentic, meaning Cowork, Code, Autopilot and access to frontier models, moves to usage-based billing. An automatic router decides which model handles each request based on accuracy, speed and cost, and new FinOps tools are supposed to give IT and finance teams visibility into what is being spent.

Two details matter here. First, Microsoft has not disclosed per-task pricing, so nobody can model their costs precisely yet. Second, from 2 November new licenses sold through the CSP channel will have usage-based billing enabled by default. If your Microsoft licenses come through a reseller, that is a date to put in the calendar, because spending that used to be capped by a per-seat price will now grow with how much your agents actually do.

The logic of the change is easy to see from Microsoft's side. When an agent works around the clock, a flat monthly fee per human user stops making sense; the human is no longer the thing that limits usage. But it shifts budget responsibility onto customers, and companies that let agents multiply without spending policies will get some uncomfortable invoices.

The questions IT teams are already asking

The early commentary from IT administrators has been notably sober. The sandbox isolation and audit controls are vendor claims to test in a pilot, not something already proven in production. Several governance details of Autopilot remain undefined during the private preview, including approval mechanisms for agent actions, controls on outbound messages, and exactly how agent activity appears in audit logs.

There is also a compliance angle. The Frontier previews come without HIPAA coverage and without standard service level agreements, which means organisations in regulated industries should keep any pilot narrow and away from sensitive data for now.

And because every Autopilot instance is a separate identity in your tenant, your permission model has to treat each agent like a new hire: what can it read, what can it send, and who reviews what it did. Companies that never tidied up their internal permissions will find that agents inherit the mess.

Practical takeaways for businesses building digital products

If you run a business on Microsoft 365, the sensible move is a small, measured pilot. Pick one recurring process with low risk, give an agent clear boundaries, set spending caps from day one, and gather real cost data before making any procurement decision. Microsoft's Ignite conference in mid-November will bring more detail, and companies with actual pilot numbers will negotiate from a much stronger position.

Data readiness matters more than the technology itself. Agents act on what they can reach in your tenant, so messy SharePoint permissions, stale documents and undocumented processes translate directly into agent mistakes. Cleaning that up is unglamorous work, but it is the real preparation.

For companies that build digital products, there is a second-order effect worth noticing. Agents consume software through APIs and connectors, not through pretty interfaces. If your product or your internal systems have no clean programmatic access, they are invisible to this whole generation of tools. Building proper APIs, webhooks and structured data has quietly become a competitive requirement, and it is exactly the kind of work worth handing to an experienced development partner while your own team focuses on the core business.

Finally, keep some perspective. OpenAI shipped its Agents API earlier this month, Meta has its Muse agent out shopping for consumers, and now Microsoft has rebuilt its flagship productivity product around the same idea. The agent platform race is on, and locking your processes into one vendor before the dust settles carries its own risk.

Final notes

Microsoft's Copilot overhaul is the clearest signal yet that autonomous agents are moving from demos into ordinary work infrastructure. Home and Office in Copilot will get the headlines, but the durable changes are Autopilot agents with their own identities and a billing model that charges for what agents do rather than for the humans who employ them.

Treat the announcement the way you would treat a new class of employee: promising, unproven, and in need of a job description, a budget and supervision. Start with one small pilot, watch the costs, ask your IT team the hard governance questions, and make sure your own systems are ready to be worked on by software rather than only by people. The companies that do this calmly over the next few months will be the ones that benefit when the technology settles.