Microsoft auto-installed Copilot on eligible commercial Windows devices last month. Forty-plus updates shipped in July alone. Cowork, its agentic mode that plans and executes multi-step tasks, went generally available. Twenty million paid seats now sit inside enterprise organizations worldwide.

And Microsoft’s own research says most of those seats are producing marginal results.

The company’s 2026 Work Trend Index surveyed 20,000 workers across ten countries. Buried in the data is a split that should concern every leader paying for AI licenses right now. Fifty-eight percent of AI users say they produce work they could not have produced a year ago. That sounds impressive until you meet the other number: among a subset Microsoft calls “Frontier Professionals,” that figure jumps to eighty percent.

Frontier Professionals represent sixteen percent of the AI-using workforce. Microsoft distinguishes the group by how its members work with AI.

Their habits deserve a closer look.

What the sixteen percent actually do

Microsoft found three patterns in how Frontier Professionals approach their work.

First, they pause before starting work. Fifty-three percent of Frontier Professionals stop before a task and decide which parts the AI should own and which parts they should do themselves. Among everyone else, that number drops to thirty-three percent.

That collaboration habit resembles dividing work with a new team member on their second week: figure out who handles what before anyone starts working.

Second, they intentionally do some work without AI. Forty-three percent of Frontier Professionals deliberately keep their own skills sharp by choosing to do certain tasks manually. The broader population does this at thirty percent. They give AI a defined role while continuing to exercise their own judgment.

Third, they routinely rethink workflows. They look at how the work flows, identify where an agent can improve the quality or speed of a task, and restructure. They keep adjusting that division of labor as the work changes, much as a growing human team would.

The mental model gap

For anyone running a team, these findings suggest a useful training priority: help people define how AI participates in the work.

Frontier Professionals treat AI as a collaborator with a role in the work. Eighty percent say they produce work they could not have produced twelve months ago: new analyses, new formats, or faster work on complex tasks that used to take days.

The remaining eighty-four percent sit outside Microsoft’s Frontier Professionals category. Separate surveys show only twenty to thirty percent of purchased Copilot seats see weekly use. Seventy percent of the Fortune 500 pay for Microsoft 365 Copilot. Those adoption figures give leaders a reason to examine usage habits alongside seat provisioning.

Microsoft has pushed distribution harder through auto-install on devices, more features, and forty updates in a single month. The Work Trend Index data also gives leaders a reason to invest in the habits people bring to those tools.

What this looks like on your team

If your team has AI tools and results are falling short of expectations, examine how people divide and review the work with AI.

For a team with AI access, define how the AI participates in the work. Who decides what it handles? Who reviews what it produces? When does a human take over? When does the AI run without supervision?

Use those questions to agree on the division of work before starting a task.

Teach people to direct AI work, review its output, and adjust its assignment. Give the system a defined role and explicit limits on autonomy, with human oversight proportionate to the task.

Why providers are hiring deployment specialists

Salesforce hired zero new traditional engineers in fiscal year 2026. Instead, they planned more than a thousand forward-deployed engineers whose job is to sit inside client organizations and get AI working in production. OpenAI launched a deployment company. Microsoft created Frontier Company, a standalone unit with six thousand specialists.

The companies building the AI are spending billions on people whose job is to teach other people how to work with it. That investment in collaboration supports the technology they sell.

Teams can begin practicing those habits now: pause before starting, decide what the AI should own, and review the result together. Vendor support can build on a division of work the team already understands.

Twenty million paid seats make that work a practical training priority.