The next measure of AI momentum is work transformed

In yesterday’s earnings call, we shared that Microsoft 365 Copilot has surpassed 30 million paid seats, with net seat adds more than doubling quarter over quarter.¹ This is an important milestone, but it’s also a signal that points to a much broader shift: Across every industry, AI is moving from assistant to active participant, becoming an essential part of how work gets done, rather than just a tool people use. Organizations are redesigning workflows around it. And employees are directing agents to take on tasks, projects, and processes, expanding their agency and increasing their impact.

The first phase of enterprise AI was defined by adoption—deployments, licenses, and hours saved—real gains that will continue to compound. The next phase will be defined by transformation. At Microsoft, it is our mission to empower every person and every organization on the planet to achieve more, and we are excited by the growing opportunity to fulfill that mission as humans and agents collaborate to build what was previously impossible.

New data from Microsoft 365 Copilot telemetry signals, along with examples from our customers around the world and teams inside Microsoft, demonstrate this transformation in action. It shows up in three ways: in how deeply and broadly people now use AI, in what AI makes possible—from faster work to net-new capabilities—and in how organizations tailor it to specific roles. Let’s take a closer look, starting with the technology itself.

A step change in innovation

In the past few months, a new kind of AI product has emerged: a model wrapped in an agentic harness that can close its own feedback loop—planning, executing, testing, and correcting its work to deliver a finished result. This approach enabled Copilot Cowork, which became generally available in June. Cowork can take on multi-step work: you define the task, and it runs it end to end and returns a completed result. Its multi-model design lets you run the models a task needs, so capability scales with the work as more models become available. And testing shows that Cowork was 30% to 40% cheaper² compared to a single model option. We also introduced our first autopilot agent, Microsoft Scout, which stays active in the background, with its own identity and permissions, to keep work moving even when your attention is elsewhere. All of the outputs are grounded in Work IQ to understand the context that shapes work inside your organization, including people’s roles, projects, priorities, and institutional knowledge—so they are uniquely relevant to your business.

We’ve seen a step change not just in the products we deliver to customers, but also in the way our engineering teams build them. While many employees have contributed to Cowork, the core team remains intentionally small—growing from three engineers at the start to just nine when the product became generally available. And yet the speed-to-market was remarkable, with Cowork moving from inception to preview to being used by half the Fortune 500 all within the space of six months. The breakthrough came from rebuilding the software development process around agents, allowing a tiny team to achieve levels of speed, scale, and automation that would have been impossible with traditional engineering workflows.

Microsoft Scout also started with a small group of people working alongside autonomous agents with shared context, evaluations, and clear specifications, and again moved from an initial concept to a working product in a matter of weeks rather than months. “In the past, I spent a lot of time helping teams scale: aligning large groups, coordinating work, and making sure everything moved forward together,” says Omar Shahine, Corporate Vice President of Microsoft Scout. “Now it’s more about creating the conditions for expert teams to move quickly, experiment safely, and continuously capture what they learn so every project starts further ahead than the last.” Frontier teams like these—nimble, focused, and defined by human-agent collaboration—provide a strong signal for how knowledge work will transform in the months and years to come.

Depth and breadth of use

As AI products and the approach to building them mature, Copilot usage across organizations is also deepening and broadening. In the past year alone, the number of conversations per user has nearly doubled,³ while the number of users engaging with multiple Copilot features grew by triple digits.⁴ And the fact that average weekly engagement with Copilot is now on par with Outlook and Teams⁵ speaks to the speed at which AI is becoming a part of everyday work.  

When we look at what people ask Copilot Cowork to do most in a given week, we’re seeing a rise in requests that span multiple steps, like analyzing a set of data and drafting the email that explains it or researching a problem and writing the code to solve it. When we account for these multi-step workflows, analysis-related work accounts for 49% of all tasks, up from 29% when analysis is performed as a standalone task.⁶ In other words, the data suggests AI is moving beyond helping people complete tasks to helping them complete entire workflows.

Zooming out, early adopters like manufacturing, banking, and software/technology remain among the industries with the highest number of active agents in the Microsoft 365 ecosystem. But what’s surprising is that industries like automotive and healthcare are emerging as some of the fastest-growing agent adopters,⁷ signaling that agent value is spreading well beyond the early leaders.

The clearest sign of this momentum may be how quickly organizations are scaling Copilot across their workforces. The number of our customers with more than 50,000 seats has increased more than 7x year over year. Meanwhile, the number of enterprise customers deploying Copilot to the majority of their information workers grew nearly 75% quarter over quarter, a signal of how central Copilot has become to their operations. And the time it takes to reach high usage, meaning monthly active usage above 80% across a customers’ user base, has fallen from months to just days over the past year.⁸ Taken together, these shifts indicate that organizations are increasingly embedding AI and agents into the flow of everyday work.

What AI makes possible

At its most basic, AI-enabled speed creates efficiencies. But the organizations that are getting this right are moving beyond efficiency for efficiency’s sake and starting to see how AI changes the work itself—and can even expand what an organization is capable of doing.

For example, Microsoft’s own Cloud Supply Chain team took a “lean before agents” approach. It first mapped and simplified six end-to-end workflows and built a shared data foundation, then deployed more than 70 purpose-built agents across planning, sourcing, fulfillment, and logistics. The result was a 75% reduction in cycle time across selected workflows, with the lesson baked in: simplify the work itself before you automate rather than layering agents onto a broken process.

Cycle time is one measure of progress. The quality of the decisions underneath it is another. At intelligent power management company Eaton, root cause analysis had long relied on time-intensive investigations. To lighten the load, the company has begun to use an AI-powered quality agent that’s helping experts review reports on manufacturing quality issues to quickly surface potential root causes, trends, and supplier insights. Applied across approximately 5,000 reports so far, it has accelerated analysis, strengthened data-driven decision making, and enabled earlier risk identification and prevention actions while keeping Eaton’s expert team at the center of every decision.

Other organizations are discovering what’s possible by empowering those closest to the problem to build the solutions. Premera Blue Cross set out to let employees create their own AI agents to take repetitive, manual work off their plates and tackle growing backlogs and manual workflows. The health plan, which serves members across Washington and Alaska, started with putting Microsoft 365 Copilot and Copilot Studio into the tools their people already used every day, backed by training and clear guardrails so employees could experiment without putting member trust at risk. The team has since built more than 900 agents and counting, including a contract exhibit agent whose intelligent document processing cut a manual workflow from 30 to 45 minutes down to about three—freeing up employees to focus on members who need access to care or answers to critical questions.

Further along the spectrum, AI can unlock the ability to do work that previously couldn’t be done at all. At Levi Strauss & Co., where AI is being woven across the business from consumer insight to human resources (HR), an agent analyzed 1,100 finance standard operating procedures and cataloged 18,000 individual tasks in a day, creating a detailed view of work that wouldn’t have surfaced otherwise. “The agent sized an opportunity I’d assumed was impossible to quantify,” said Lisa Sterling, Vice President of Finance at Levi’s.

And at global consulting organization EY, an Autonomous Sourcing Agent (ASA) automates low-cost, low-risk transactions across the procurement process, from requisition and negotiation to purchase order creation. It can negotiate with one or many suppliers at once across cost, delivery timelines, and warranty terms while keeping humans in the loop for validation and escalation. Since its October 2025 pilot, the ASA has supported more than 200 transactions and is expected to handle 1,500 over the next year. The result is a faster, more efficient procurement process that helps teams get more value from suppliers while ensuring important judgement calls remain with people.

Specialization at scale

The difference between the organizations seeing changes like these and those still waiting isn’t how widely they’ve deployed AI—it’s how well they’ve tailored it to the work. Yet the most common deployment pattern is still a general-purpose one: give everyone the same tool and let them figure it out. The organizations creating the most distinct value are doing something different—starting with how specific roles actually work, then building AI around those patterns.

One example: S&P Global is showing how AI can make deep, specialized expertise accessible to more people by bringing trusted intelligence directly into the flow of work. As demand for AI-ready data grew, the company used Microsoft Copilot Agent Builder and Microsoft Graph connectors to create a Copilot agent in Microsoft Teams, giving customers a simpler way to access and interact with its commodities research and market insights. By delivering 95% faster data extraction and 98% faster comparative analysis, the solution helps customers turn vast amounts of information into actionable intelligence more quickly, empowering faster, more informed decisions across global energy and commodities markets.

Microsoft’s sales team mapped distinct seller personas to dedicated AI agents—Researcher, Analyst Agent, Sales Agent, Deal Agent, and MSXi Copilot (a sales intelligence and reporting agent)—each designed to fit specific moments in how sellers spend their time. With agents taking on more administrative and coordination tasks, seller time dedicated to customer-facing activities doubled, from 25% to 50%. The pilot group saw 9.4% more revenue per head and 20% faster deal closure, and in a survey reported a 3x increase in adoption of priority use cases.⁹

Kantar took a ground-up approach. At the marketing data and analytics firm, a “people agent” for HR queries (really a system of about 10 agents with 60 topics and related flows working together) is being built and expanded through collaboration by people that actually handle and support these requests across its People team. It already resolves a growing share without human intervention: an employment verification letter that once took three days is now returned in 120 seconds. Kantar reports the share of HR queries handled this way has risen from zero in February to about 40%, against a year-end target of 95%.

Microsoft’s People Operations team runs more than a million employee interactions a year across more than 100 systems, doing work that carries real weight for people. They are using Frontier Tuning to scale HR expertise across their processes and employee moments rather than improving one task at a time, capturing the expertise of practitioners as tasks, skills, and rubrics that define quality standards. The result is an AI-powered system that can take on operational work with the context, standards, and institutional knowledge necessary for success, transforming their practitioner know-how into organizational capability. The team can see how the agent performs on real work and tune the levers—tools, context, skills, and the model itself—to ensure consistent performance. The agent developed with Frontier Tuning is now live in People Operations, delivering 5x improvement in successful task completion compared to a baseline general-purpose model.¹⁰ As routine activities are increasingly handled by AI, Microsoft HR’s People Operations experts can focus on the complex employee interactions, nuanced decisions, and strategic work where human insight matters most.

What’s next

As AI technology advances and organizations everywhere begin to see real value emerge, the measures that matter are changing as well. We’ll see increased focus on depth of use, measurable change in the work itself, the growth and governance of agents, the redesign of roles and workflows, and the emergence of capabilities that didn’t previously exist. The organizations building durable advantage are the ones where technology, operating models, and human agency are evolving together. That is the Frontier Firm taking shape.

Footnotes 

¹Microsoft Q4 Earnings call.  

²This analysis was conducted by internal Microsoft teams. The analysis compared 125 test runs across a total of 12 light, medium, and heavy work data prompts in Copilot Cowork and Claude Cowork with their Microsoft 365 connector, both using model Opus 4.8. For Copilot Cowork, the costs were calculated using variable rates based on relevant models, context, tools, and runtime from our internal logs. For Claude Cowork, the costs were calculated using publicly available API rates based on token usage and Microsoft 365 connector usage. We used these costs to get a total price per run and compared the prices across light, medium, and heavy prompt sets between Copilot Cowork and Claude Cowork with their Microsoft 365 connector. Cost savings results based on tests conducted in June 2026. Actual costs and potential savings may vary depending on usage, configuration, time, and other factors.  

³Microsoft Q4 Earnings call.  

⁴Based on anonymized, aggregated Microsoft 365 Copilot productivity signals, count of monthly-active M365 Copilot users who use 2 or more unique Copilot-based features over a 28-day period. Includes all countries and enterprise customers in June 2025 and June 2026.  

⁵Microsoft Q4 Earnings call. 

⁶Based on anonymized, aggregated Microsoft 365 Copilot productivity signals, Cowork telemetry from July 19, 2026 to July 26, 2026, analyzing the most common tasks customers delegated to Copilot Cowork during the reporting period.  

⁷Based on anonymized, aggregated Microsoft 365 Copilot productivity signals, Monthly Active Agents measures the count of unique active agents observed through telemetry across Declarative Agents, Custom Engine Agents, and SharePoint Agents. The metric is calculated over a rolling 28-day window. An agent is counted as “active” if, within that 28-day period, it meets at least one of two criteria: (1) it has at least one day of user-initiated usage, or (2) it completes at least one autonomous run. Each qualifying agent is counted once regardless of how many times it was used, ensuring a de-duplicated measure of the active agent base. Growth is assessed over a 12-month period spanning June 2025 through June 2026, comparing the active agent count at the start and end of the period.  

⁸Microsoft Q4 Earnings call.  

⁹Internal Microsoft sales team data based on 687 sellers of Microsoft 365 Copilot from January to June 2024, as compared with sellers with low usage of Copilot.  

¹⁰Success was defined as a “perfect run”: the task achieved a 100/100 score against human-defined evaluation criteria. Tasks were evaluated across multiple executions to account for variability in completion paths, task sequencing, and tool selection, with only consistently perfect outcomes counted as successful.