Copilot daily usage accelerates
Microsoft chairman and CEO Satya Nadella placed Copilot adoption at the center of the company's latest earnings message, disclosing that daily users grew nearly 50% in the three months to September compared with the prior quarter and more than five times the level of a year earlier. The figure offers one of the clearest signals yet that Microsoft is turning its broad generative AI investment into repeated, habitual use across work software. It does not, however, resolve the question investors have pressed for since Copilot became a paid product: how many customers actually pay for it. Microsoft still did not disclose a paid subscriber count, leaving the market to infer conversion from a patchwork of usage, seat, and revenue metrics.
The context matters. Three months earlier, Copilot had roughly 30 million users, according to figures circulating around the company's AI business, and some observers described that number as paid seats. This quarter, Nadella said the number of customers paying for Copilot by usage grew more than 75%. That is a different measure from raw daily users. Daily activity can be lifted by bundled access, pilot programs, employee trials, and workplace accounts that are provisioned but not necessarily billed as standalone subscriptions. Paid-by-usage growth is closer to a revenue signal, though it still does not reveal average revenue per user, churn, or the mix between consumer and commercial accounts. The combination suggests strong momentum, but it leaves room for interpretation.
One app, more surfaces
Microsoft has spent the past several quarters trying to make Copilot feel less like a feature bolted onto individual products and more like a single layer across work. Nadella described the biggest update to Copilot so far as bringing Chat, Cowork, Autopilot, and Code into one unified experience. The Copilot app had already added coding and Autopilot in September, and the company is now positioning those capabilities as parts of one assistant rather than separate tools with separate interfaces. That consolidation is important for adoption. A fragmented assistant forces users to remember which product does what; a unified assistant can carry context from a conversation to a document, a codebase, a meeting, or a workflow.
The early results cited by Microsoft are encouraging on their face. User satisfaction rose nearly 20% in the quarter to another all-time high, Nadella said. The number of customers with more than 50,000 Copilot seats grew eight times in a year, a sign that large organizations are not merely testing the product in small pilots but expanding it across substantial employee populations. Microsoft also said its high-end Microsoft 365 E7 bundle now has millions of seats. Large-seat deployments are strategically valuable because they create internal usage norms, generate more feedback, and make Copilot part of daily workflows rather than an optional add-on. They also raise the stakes for reliability, security, and cost control.
Copilot's expansion fits a broader pattern in enterprise software. Productivity vendors have been racing to embed AI into documents, spreadsheets, email, meetings, and chat, hoping that distribution beats standalone AI tools. Microsoft's advantage is its installed base across Windows, Microsoft 365, GitHub, and Dynamics 365. Its challenge is proving that the assistant is useful enough to justify incremental spending and trustworthy enough to handle sensitive corporate data. The daily-user growth suggests that more people are opening Copilot regularly; the seat growth suggests that more organizations are willing to commit at scale. The missing piece is the paid conversion rate, which Microsoft has chosen not to make public.
Foundry and the multi-model bet
The earnings message was not only about Copilot. Microsoft said revenue at Foundry more than tripled in a year. Nadella described Foundry as part of the common platform underneath Microsoft's own AI products, a layer that lets customers build, deploy, and manage AI applications with different models. The number of customers building with several AI models grew five times, while revenue from open-source models grew ten times. Those figures point to a strategic bet: Microsoft wants to be the place where enterprises assemble AI systems from multiple providers, not only the place where they buy one company's model.
That flexibility is becoming a competitive necessity. Enterprises rarely want to depend on a single model for every task. They may use frontier models for complex reasoning, smaller or cheaper models for routine work, and open-source models for cost, control, or data-residency reasons. Nadella said Microsoft offers customers a range of AI options, including some that cost less than frontier models. He also said Microsoft is making Copilot more useful with each model release by consolidating frontier gains for broad deployment. In practice, that means Microsoft is trying to turn rapid model progress into product improvements without forcing customers to rebuild their workflows every time a new model appears.
Foundry's growth also supports Microsoft's cloud narrative. Microsoft Cloud revenue surpassed $63 billion, with growth accelerating to 30% year over year. AI workloads are not the only driver, but they are increasingly intertwined with cloud consumption. Training and inference require compute, storage, data services, and governance tools. The more customers build agents and AI applications, the more they consume underlying cloud services. Foundry sits at that intersection. Its tripled revenue and the fivefold growth in multi-model customers suggest that Microsoft is making progress in the platform layer, even if the market remains highly competitive.
Agents, authority, and the control plane
Nadella also used the update to frame the next phase of AI adoption around agents: software that can take actions, call tools, and complete tasks with less human prompting. More than 50 million AI agents now run in Agent 365, the system Microsoft uses to help customers secure and govern them, he said. That is up 30% in the quarter. Microsoft is linking Agent 365 with Windows so agents can work within secure boundaries on a device using MXC. The company says agents including OpenAI Codex, Anthropic's Claude Code, and Meta Muse are adopting the system. If that adoption holds, it would make Microsoft a control point for agents built on multiple model providers, not just its own.
The concept Nadella emphasized most sharply was separating intelligence from authority. In a post after the earnings call, he wrote that Perception enforces identity, permissions, monitoring, and containment outside the model itself. That is a subtle but important architectural claim. As models become more capable, the risk is that intelligence and permission become entangled: a model that can reason about a task may also be trusted to execute it without enough checks. Separating the two means the model can generate a plan or a recommendation while a separate system decides whether the action is allowed, which data can be accessed, and what happens if something goes wrong.
For enterprises, that separation is likely to matter as much as raw model quality. An agent that can read email, update records, write code, or move money is useful only if it can be governed. Identity, permissions, monitoring, and containment are the controls that let organizations grant authority selectively and revoke it when needed. Microsoft's pitch is that Agent 365 and Perception can provide that control layer across agents from different vendors. The strategy echoes the way companies already manage devices, applications, and user identities. If agents become a new class of enterprise actor, they will need a similar control plane.
Cost, quality, and the road to deployment
Microsoft also pointed to GitHub Copilot Fusion, announced that morning, which Nadella said delivers frontier-level quality at up to 77% lower cost. Cost is an underappreciated part of the AI adoption story. Early generative AI pilots were often judged on capability alone, but production deployments are judged on unit economics. A coding assistant that is slightly less capable but dramatically cheaper can be more attractive for routine tasks. A model routing system that sends simple requests to cheaper models and complex requests to frontier models can reduce spending without sacrificing quality on the tasks that matter most.
Nadella's post-call commentary tied these threads together. He described Copilot as the new operating system for work and said Microsoft is bringing the business logic and capabilities of Dynamics 365 directly into Copilot, helping organizations stand up what he called their own infinite SaaS factory. That is an ambitious framing, but it reflects a real shift: AI assistants are moving from answering questions to participating in business processes. If Copilot can connect to enterprise data, respect permissions, invoke workflows, and govern agents, it becomes more than a chat interface. It becomes an access layer for work itself.
The numbers Microsoft disclosed will be parsed closely because they mix usage, satisfaction, seats, revenue, and platform growth. Daily users up nearly 50% quarter over quarter and more than fivefold year over year show demand. Satisfaction up nearly 20% suggests that product changes are landing. Customers with more than 50,000 seats up eightfold in a year show enterprise scale. Foundry revenue tripling and multi-model customer growth show platform traction. More than 50 million agents in Agent 365 show that the agent economy is no longer theoretical. What remains undisclosed is the most direct measure of commercial success: how many people pay, how much they pay, and how durable that spending is.
That gap is likely to keep the Copilot debate alive. Microsoft can point to rapid daily-use growth and large seat deployments as evidence that AI is becoming part of everyday work. Skeptics can note that daily users are not the same as paid subscribers and that bundled access can obscure true demand. The company's answer is to keep expanding the product surface, lowering cost, and adding governance that makes agents safe enough for production. Nadella's phrase about separating intelligence from authority captures the bet: the winner in enterprise AI may not be the company with the most capable model alone, but the company that can give the most capable models a trusted place to work.
Source:TNW | Microsoft News