What are you seeing your users use Microsoft Copilot Cowork for??

The Adoption Paradox: Millions of Seats, Uneven Real-World Use

Every Fortune 500 company you can name probably has a Copilot license somewhere in the building. Microsoft reports Microsoft 365 Copilot has surpassed 30 million paid seats, with customers holding 50,000+ seats growing more than 7x year over year. By the company's own account, its newer agentic product, Copilot Cowork, went from inception to use by half the Fortune 500 within six months — a vendor claim, but a striking one.

Here's the tension: seats sold don't equal work transformed. One training vendor's adoption analysis estimates that only about 35.8% of employees who hold Copilot licenses actively use them — meaning roughly two-thirds of licensed users may be, in that vendor's blunt phrasing, "costing money without generating results." That figure is attributed only to unspecified "industry data" and comes from a company selling training, so treat it with caution. But even at half that rate, the gap is enormous.

The picture from the ground is similarly mixed. Microsoft describes the current phase as "seat-add and expansion," with most organizations still in pilots and phased rollouts rather than daily-deep usage, per governance-focused analysis. Meanwhile, the standout successes look nothing like general-purpose rollouts:

• Premera Blue Cross built 900+ role-specific agents, cutting a 30–45 minute contract workflow to roughly 3 minutes.
• Kantar's HR "people agent" handles about 40% of HR queries, shrinking an employment verification letter from 3 days to 120 seconds.
• Infosys, TCS, and Wipro each deployed Copilot to 100,000+ employees and report 86–95%+ active usage rates.

So what separates a license that sits idle from a deployment like those? The research points to a consistent answer: tailoring beats general deployment. Microsoft itself notes that organizations seeing the most value shape AI to specific roles rather than pushing a general-purpose tool to everyone, and adoption guidance consistently recommends starting with a few focused, high-impact workflows.

That leaves the core question this article explores: what are people actually doing with Copilot Cowork when it works — and why do some users thrive while identical licenses go untouched down the hall? The same pattern shows up in the broader agent-skills ecosystem, where a tool is only as valuable as the specific, verified capability attached to it — the problem SkillzBench is building its free, benchmarked skills directory and MCP connector to solve. The short answer, as we'll unpack: the winners aren't using Copilot as a chatbot; they're using it as a workflow engine.

The Real Usage Data: Multi-Step Workflows, Not One-Off Tasks

If you want to know what Cowork is actually for, ignore the feature list and look at what people type into it. According to Microsoft's own usage analysis, users rarely ask Cowork for a single task. They hand it an entire job: analyze a dataset and draft the email that explains it, or research a problem and write the code to fix it.

That shift shows up clearly in the numbers. When multi-step workflows are counted, analysis-related work accounts for 49% of all Cowork tasks — up from 29% when analysis is measured as a standalone activity. Microsoft reads that gap as a signal that AI is moving from completing tasks to completing workflows, and the data supports it.

The broader Copilot picture reinforces the pattern. An analysis of 100,000+ Copilot chats found 49% of conversations supported cognitive work — analysis, problem-solving, evaluation, and creative thinking — compared with 19% for working with people, 15% for finding information, and just 17% for producing work. Users are treating these tools as thinking partners, not typists.

The daily workflows where this shows up are remarkably consistent:

• 86% of users say it's easier to catch up on missed meetings
• 85% get to a good first draft faster
• 64% process email more efficiently
• 84% take action after meetings more quickly

Notice what those have in common: each one is a compressed workflow, not a one-off trick. Meeting catch-up means summarizing a transcript, extracting decisions, and drafting follow-ups. First drafts mean gathering context, structuring an argument, and producing the document in one pass.

One caveat worth keeping in mind: nearly all Cowork-specific usage data comes from Microsoft itself, a vendor with obvious incentive to present favorable numbers. But the named customer examples are concrete. Levi Strauss & Co.'s finance agent analyzed 1,100 SOPs and cataloged 18,000 tasks in a single day, and Kantar's HR agent cut employment verification letters from three days to 120 seconds, according to Microsoft's customer reporting.

The lesson for business owners is that the highest-value deployments are purpose-built for specific workflows, not general-purpose assistants bolted onto everyone's desktop. That's exactly the gap agent skills address — small, focused instruction packages that extend an AI tool into one job done well. SkillzBench is building a free directory where you can browse, verify, and benchmark agent skills before wiring them into your own agents, with security scanning and published performance lift. Join the waitlist to be first in line when it launches.

What Tailored Deployments Look Like in Practice: Named Enterprise Examples

A 45-minute contract workflow that now takes three minutes. An employment verification letter that once sat in an HR queue for three days, delivered in 120 seconds. These aren't projections — they're results from named enterprises running Microsoft Copilot agents, and they share a common thread: each organization tailored AI to a specific role, not a generic rollout to everyone.

Microsoft's own analysis of Copilot deployments makes the point directly: the organizations seeing the most value tailor AI to specific roles and workflows rather than giving everyone access to a general-purpose tool. The named examples bear that out. Premera Blue Cross built more than 900 agents and cut a contract exhibit workflow from 30–45 minutes to roughly 3 minutes. At Kantar, an HR "people agent" now handles about 40% of HR queries, shrinking employment verification letters from three days to two minutes. Levi Strauss & Co.'s finance agent analyzed 1,100 SOPs and cataloged 18,000 tasks in a single day — work that finance VP Lisa Sterling said "sized an opportunity I'd assumed was impossible to quantify."

The pattern repeats across industries:

• Providence clinicians respond to patient messages up to 35% faster with Copilot support.
• PG&E addressed up to 40% of helpdesk demands, saving more than $1 million annually.
• Microsoft's Cloud Supply Chain runs 70+ purpose-built agents and reports a 75% reduction in cycle time across selected workflows.
• Microsoft People Operations saw a 5x improvement in successful task completion with tuned agents versus a general-purpose baseline.

That last data point is the most instructive. A tuned, task-specific agent outperformed the general-purpose alternative by a factor of five — evidence that the payoff comes from narrowing scope, not broadening it. It's also why adoption guidance consistently recommends starting with three to five focused, high-impact workflows instead of enterprise-wide access, and why the numbers can look so lopsided: Microsoft 365 Copilot has passed 30 million paid seats, yet one (unverified, vendor-attributed) analysis claims only about a third of licensed employees actively use theirs. Broad access without purpose tends to produce shelfware; narrow deployment with a clear workflow produces results like Premera's.

For teams building these role-specific agents, the same principle applies beyond Copilot: the quality of what an agent can do depends on the skills you give it. SkillzBench is building a free directory of agent skills — with security scanning, sandbox benchmarking, and version pinning — so you can see which skills actually lift performance before wiring them into your own workflows. Join the waitlist to get access at launch.

The lesson from these deployments is simple: pick one painful workflow, tailor the agent to it, measure the result. That's how a three-day letter becomes a two-minute one.

How to Put Cowork (or Any AI Agent) to Work in Your Business

The gap between buying AI and actually using it is wider than most businesses expect. Microsoft 365 Copilot has surpassed 30 million paid seats, yet one analysis suggests only about 35.8% of employees with access actively use it — and while that figure comes from unverified industry data, the pattern it describes is one many organizations recognize firsthand.

The lesson from companies getting real results is consistent: start narrow, not broad. Rather than rolling out a general-purpose assistant to everyone, adoption guidance recommends picking three to five focused, high-impact workflows. The best candidates are the ones users already rely on Copilot for daily — meeting catch-up (86% found it easier), post-meeting follow-up (84%), email processing (64%), and first-draft creation (85%).

• Pick 3–5 workflows where the before/after is measurable — time per task, turnaround time, error rates.
• Train before you scale. Organizations that train first reportedly see adoption rates two to three times higher than those that simply hand out licenses.
• Assign owners to each workflow so results get reviewed and refined, not just generated.
• Expand only after the first workflows prove out — Microsoft itself describes the current enterprise phase as "seat-add and expansion," not mature deployment.

Before any of that, treat governance as a prerequisite, not an afterthought. Copilot reaches across SharePoint, Teams, OneDrive, and Exchange, which means messy permissions become AI-readable permissions. Roughly half of IT leaders lack confidence in managing Copilot's security and access risks, according to Gartner data cited by Lighthouse Global. The Echoleak vulnerability — which enabled silent exfiltration of sensitive email data before Microsoft patched it — is a useful reminder of why data remediation comes first.

Once the workflows are running, build in a simple review habit: Copilot drafts, humans finalize. This matters most in legal, finance, and HR work, where a confident-sounding draft can still be subtly wrong. Microsoft's own legal department pairs AI speed with human review and reports 32% faster task completion alongside a 20% accuracy boost — the combination works because nobody skips the final read.

Finally, remember that the highest-value deployments are tailored, not generic. Premera Blue Cross built 900+ purpose-built agents, and Levi Strauss used an agent to analyze 1,100 finance SOPs and catalog 18,000 tasks in a single day. The closer your AI gets to your specific workflow, the bigger the payoff — and the more important it becomes to know which skills and extensions are trustworthy. That's exactly the problem SkillzBench is building for: a free directory that verifies, security-scans, and benchmarks agent skills so you can extend your AI tools with evidence instead of guesswork. Join the waitlist to be first in line when it launches.

Going Further: Extending AI Agents with Skills You Can Trust

The pattern behind every standout Copilot deployment in the data is the same: the value doesn't come from a general-purpose assistant, it comes from agents purpose-built for a specific job. Microsoft's own cloud supply chain runs more than 70 purpose-built agents and reports a 75% reduction in cycle time across selected workflows. Premera Blue Cross built over 900 agents, including one that cut a 30–45 minute contract workflow down to roughly three minutes, according to Microsoft's customer data.

That's where agent skills come in. A skill is an instruction package that extends an AI agent into a specific workflow — the difference between an agent that can "help with contracts" and one that knows exactly how your contract exhibit process runs. Skills are already everywhere, especially for coding agents like Claude Code, Codex, Cursor, and Gemini CLI. The problem is trust: published research has found that roughly 26% of published skills carry at least one vulnerability, and there's rarely any way to tell which ones actually help and which introduce risk.

For a business audience, that matters. Governance is already the biggest adoption roadblock for Copilot itself — governance-focused analysts note that about 50% of IT leaders lack confidence in managing Copilot's security and access risks. Handing an unvetted third-party skill to an agent with access to your company data only compounds that concern.

So what does "trustworthy" actually look like when you're choosing a skill? A few signals are worth demanding:

• Security scanning and license checking on every skill, not just the popular ones
• Version pinning, so you know exactly which code your agent is running
• Sandboxed benchmarking that shows whether the skill actually improves results — measured with and without the skill, under identical conditions
• Ongoing monitoring, since skills change upstream and yesterday's safe version isn't always today's

This is the problem SkillzBench is being built to solve. It's a free, independent directory of AI agent skills with four pillars: Discover (continuously crawling and deduplicating skills from GitHub), Verify (pinning versions, validating, license-checking, and security-scanning), Benchmark (testing every skill in a sandbox and publishing the Lift in percentage points), and Equip (a free MCP server that connects any agent only to the skills and tools you choose). Every skill gets a 0–100 trust score built from security, quality, license, maintenance, and community signals — with benchmark Lift always shown separately, and no money ever touching the scores.

The lesson from the enterprise data is clear: tailored agents beat generic ones, and the difference between a great agent and a risky one often comes down to what you attach to it. Skills, proven — that's the standard worth holding every extension to before it touches your workflows. SkillzBench's waitlist is open now for teams that want skills they can actually trust.