Microsoft Drops Copilot+ but Keeps Every Feature

In September 2026, Microsoft confirmed that its new Surface PCs would drop the Copilot+ name even though they still met the category’s 40-TOPS threshold and retained every Copilot+ feature. The company kept the machines’ AI capabilities and discarded the badge created to sell them.

Microsoft is shifting its strategic center from the PC badge to a work control plane above it. Through this assistant operating layer, Copilot translates a person’s intent into authorized actions across models, files, applications, and devices. To do that, Copilot must know who the user is, what the user may access, which actions require approval, and how an administrator can inspect or revoke them.

The disappearing badge separates capability from the purchase decision

Microsoft introduced Copilot+ PCs in May 2024 with a simple purchase rule: a Windows machine needed at least 40 TOPS of neural-processing performance. Microsoft emphasized local AI execution, making the NPU threshold both a technical specification and the reason buyers needed a new computer.

Microsoft then expanded features that had been exclusive to Snapdragon systems onto PCs using AMD and Intel chips. The expansion made Copilot+ less dependent on one processor family, but it weakened the category’s simplest consumer story. A badge works best when it marks a legible boundary; Microsoft kept broadening that boundary as it distributed Copilot features across hardware configurations.

The September 2026 Surface launch made the split explicit. Microsoft introduced a $1,149-and-up Surface Pro 12-inch and a $1,199-and-up Surface Laptop 13-inch, both using Snapdragon X2 Plus. The products preserve Surface as a reference platform for Windows hardware while removing the burden of explaining AI value through an NPU-qualified device class.

Reporters increasingly encountered Microsoft through systems that builders use and extend. From 2024 to 2026, the share of Microsoft coverage framed around consumers fell from 24.5% to 17.8%. Developer framing rose from 9.5% to 17.1%, while research framing rose from 10.6% to 20.9%.

For buyers, the 40-TOPS threshold still describes useful hardware. It no longer explains why Copilot can find relevant material, choose a tool, and finish work across applications without exceeding the user’s authority.

Microsoft is making intent the entry point to its products

Microsoft launched a unified Copilot app in September 2026 that combines AI chat, coding, and Autopilot agents. Word, Excel, GitHub, and the underlying agent tools remain distinct products; the app gives users one place to state an objective before Copilot routes the work into them.

Microsoft had already begun merging its consumer and commercial Copilot apps across mobile, web, and desktop. A May report described customers frustrated by multiple Copilot assistants and asking for one destination. Microsoft answered the navigation problem by consolidating the doorway rather than making users decide which assistant matched each task.

In June 2021, GitHub Copilot was a code-recommendation tool. By September 2026, Microsoft had placed coding beside chat and long-running agents inside one Copilot interface. Coding remains a distinct production system with its own economics, as the wider shift toward end-to-end agentic software work shows. Microsoft is nevertheless placing the initiation of that work beside other forms of knowledge work.

Microsoft also made macOS an entrance to its AI ecosystem with a native Copilot app released in February 2025. The move extends Copilot onto Apple’s desktop platform and loosens the link between Microsoft’s assistant and Windows hardware.

Google approaches the same control point from Android, cloud services, and its productivity tools. The struggle over the Android AI control plane already turns on permissions, entitlements, and resource limits rather than a single assistant icon. Shipping Copilot everywhere gives Microsoft reach; winning work requires context and authority behind the interface.

Delegation fails when permission is less precise than the prompt

Drafting text may require only a prompt and a destination. Sending messages, editing files, invoking software, or running for an extended period requires a principal, a defined scope, an observable record, and a way to stop. Microsoft can scale delegation only by making those relationships explicit.

Copilot Cowork uses Work IQ to ground actions in work data inside Microsoft 365. A generic model does not inherently know which document is authoritative, which meeting concerns the task, or which organizational boundary applies. Microsoft 365 can supply that context through the user’s existing work environment.

Microsoft also introduced the open-source Agent Control Specification to give developers granular and consistent controls over what agents may do. The specification acknowledges the harder engineering problem beneath interface consolidation: Microsoft must turn organizational policy into enforceable limits that survive across agents and applications.

For each task, Microsoft must identify the person or service delegating the work and give the agent only the context it needs. Each consequential action also requires an approval rule, an audit record, and a way for an administrator or user to revoke authority without dismantling the workflow.

The agent incidents involving Hugging Face and Mythos 5 raised questions about how much autonomy agents should receive and when they should seek human input. The closer Copilot gets to consequential work, the less Microsoft can rely on users to infer those limits from interface cues.

A unified Copilot can concentrate confusion as easily as it reduces fragmentation. Users may see one interface while Copilot switches among models, data sources, local processes, and remote actions behind it. Microsoft needs deployment accountability that tells users what acted, under whose authority, against which data, and with what route to reversal. One interface can conceal several risk boundaries, and software will enforce each boundary literally.

Every enterprise vendor is building from the context it controls

Across enterprise AI adoption, vendors are pairing models with the data, applications, and administrative relationships already inside their businesses.

Microsoft makes that structure unusually visible. Copilot Cowork incorporates Anthropic’s Claude Cowork technology, while Microsoft’s Frontier testing includes tools using both Anthropic and OpenAI models. Microsoft can change or combine intelligence suppliers while keeping Microsoft 365 context, Copilot distribution, and administrative control in its own layer.

These vendors are approaching the same action layer from different systems of record. Their agents may overlap in capability, but each begins with an advantage in the data and permissions its vendor already administers.

Salesforce’s specialization supplies important counter-evidence to the idea that models simply become interchangeable. A model tuned for CRM reasoning can improve the agent built on top of it. Microsoft’s multi-model testing makes a narrower point: even when model quality matters, an enterprise vendor can treat model selection as one architectural decision inside a larger system of data access, workflow routing, and governance.

Workers still cross applications no vendor fully controls. Microsoft owns a large collection of complements, but Copilot must cross organizational and software boundaries. Google, SAP, and Salesforce can contest the same action layer from the contexts they already administer. Customers keep that market open when they can move identities, models, and permissions among systems; one vendor gains power when moving them becomes costly or unsafe.

The endpoint survives by becoming a trust boundary

Local computation remains important after the Copilot+ badge disappears. Microsoft designed the original category to move more AI processing onto the device, and small models such as Phi-3 can run locally on phones and PCs without major functional compromises. Local execution can improve responsiveness and keep some uses more private.

The PC gives Microsoft a bounded place to combine hardware capability, device management, user identity, and application access. Cloud agents can handle work that requires remote data or long-running execution, while the endpoint can perform suitable inference locally and mediate approval for consequential actions. Microsoft’s continued Surface investment preserves a reference device for that hybrid arrangement.

Microsoft must make the division legible. A user needs to know whether Copilot is processing a task locally or remotely, which work data it can inspect, which application it is about to invoke, and whether the next step requires approval. Hardware performance cannot answer those questions, and a unified interface can hide them unless Microsoft exposes the boundaries.

In September 2026, Microsoft kept every Copilot+ capability and dropped the badge because the valuable boundary had moved. Local execution, managed identity, and revocable permission now meet at the endpoint, where the system must prove the chain of authority from person to agent to application. The PC earns its place by closing that chain.