What do you use Copilot Cowork for?

From Chatting to Delegating: Why 'Fire and Forget' Changes Everything

If you've ever spent an afternoon copying an AI answer from Teams into Outlook, reformatting it in Excel, then going back to ask the chatbot three follow-up questions — you already know the problem with chat-based AI. It answers; it doesn't finish.

Microsoft's Copilot Cowork, announced March 9, 2026 as the centerpiece of the company's "Wave 3" update, is designed to close that gap. As Microsoft describes it, Cowork turns a request into a plan and executes it across your apps — Outlook, Teams, Excel, Word, PowerPoint, and SharePoint — grounded in your actual work data through Microsoft's Work IQ system.

The mechanics matter here. Cowork runs tasks in the background with a plan-to-action loop: it drafts a plan, checks in when it needs clarification, and waits for human approval before applying any changes. Microsoft says you can run a dozen tasks in flight at once, each moving forward while you focus on other work.

Charles Lamanna, Microsoft's EVP for Copilot and agents, put the distinction bluntly: "With chat, you're babysitting every step — this is much more like 'fire and forget' with Cowork to get the job done." His own calendar triage illustrates the point. As CNET reports, Cowork analyzed his meetings over 40 minutes and saved him and his executive assistant "hours worth of time."

The market is watching this shift closely. Anthropic's Claude Cowork releases helped trigger a $285 billion selloff in enterprise software stocks, as investors repriced what agentic tools mean for traditional software, and some observers see Copilot Cowork as Microsoft playing catch-up. Either way, the stakes are real:

• Tasks run for minutes or hours, not single chat turns, producing real outputs along the way.
• Work is grounded in your enterprise graph — mail, files, calendar, and chats — not generic web context.
• Human approval gates keep the agent inside your governance boundaries.
• The ecosystem is extending fast: Microsoft's skill-recorder project already converts recorded work sessions into reusable Skills for Cowork.

That last point is where agent skills come in — and where caution is warranted. Published research has found roughly 26% of agent skills carry at least one vulnerability, so extending an autonomous agent with unvetted instructions carries real risk. SkillzBench is building a free answer: a directory and MCP connector that verifies, security-scans, and sandbox-benchmarks skills before your agent ever sees them.

Delegation only works when you can trust the delegate. Fire and forget only changes everything if you know what you're firing.

7 Real-World Workflows: What Cowork Actually Does Well

The clearest way to understand what Copilot Cowork actually does is to look at the work people already hand off to it. Between Microsoft's own documented examples and a first-hand practitioner account from Stoneridge Software, seven workflows stand out — each with concrete inputs and outputs.

The four Microsoft-documented workflows

Microsoft EVP Charles Lamanna's calendar triage is the flagship demo: Cowork analyzed a 40-minute meeting, drafted decline notes, and saved him and his executive assistant "hours worth of time," according to CNET's coverage. The input is a calendar conflict; the output is AI-written decline messages plus a reorganized schedule.

Meeting prep goes further. As described in Microsoft's announcement, Cowork produces end-to-end packets: a briefing document, supporting analysis, a client-ready deck, and a draft status email — all grounded in your Outlook, Teams, and SharePoint data via the "Work IQ" system.

Deep company research yields an executive summary, a research memo, and a labeled Excel workbook with citations. Product launch planning follows the same pattern: describe the outcome, and Cowork builds the plan across apps.

The three practitioner workflows

Tracy Ottenstroer of Stoneridge Software, one of the first Frontier preview users, reports three real uses:

• Blog research and drafting with live edits — Cowork aggregates sources, drafts, then adjusts as she reviews
• Business analysis producing both a Word report and an Excel model from the same request
• 1:1 meeting summaries that capture commitments and follow-ups automatically

Her guidance for getting good results: describe the outcome, not the steps, using the 16,000-character input for context — audience, format, priorities — and be specific about scope.

The honest boundary

Ottenstroer's framing is the most useful takeaway: "Copilot prepares — you advise." Cowork excels at preparation, aggregation, and first-pass drafting. It should never be the final voice on pricing, strategy, or client recommendations.

That boundary matters as the ecosystem grows. Microsoft's skill-recorder converts recorded work sessions into reusable Skills for Cowork, and published research found roughly 26% of agent skills carry at least one vulnerability — which is why SkillzBench is building free security scanning and sandbox benchmarking for the skills you attach to any agent. Delegate the prep; keep the judgment.

How to Prompt Cowork So It Delivers: Practitioner Best Practices

Most advice about "prompting agents" is recycled chatbot lore. But the only first-hand Cowork account available — Microsoft partner Tracy Ottenstroer's Frontier preview notes — offers practices that actually match how a delegated agent behaves.

Her first rule: describe the outcome, not the steps. Cowork builds its own plan-to-action loop, so tell it what "done" looks like — audience, format, priorities — rather than narrating a procedure. The 16,000-character input limit is generous; use it to specify scope explicitly, including date ranges, particular documents, and which emails matter.

Specificity pays off because Cowork runs against your real data. A vague "research this client" produces a vague memo; a scoped request grounded in named files and a defined window produces the briefing packet, analysis, and draft deck Microsoft showcased at launch (per the official announcement).

Second, exploit parallelism. Microsoft notes it's "easy to have a dozen tasks in flight at once, each one moving forward while you focus on what only you can do." Cowork checks in for clarification and requests approval before applying changes, so you supervise rather than babysit.

Third, know the boundary. Ottenstroer's verdict, from days of real use: Cowork is "excellent at preparation, aggregation, and first-pass drafting. It should never be the final voice on client recommendations, pricing, strategic advice, or anything that requires professional judgment." Her model guidance follows the same logic — Claude Sonnet 4.6 handles roughly 80% of use cases; Opus 4.6 earns its keep on high-stakes, multi-document synthesis.

Before any rollout, business leaders must address three deployment realities:

• Permissions oversharing — Cowork inherits your SharePoint permissions, so loosely shared files can surface in deliverables. Ottenstroer's warning: "The biggest enterprise risk with Cowork is permissions oversharing," and "governance-first deployment is not optional." Audit sharing before enabling wide use.
• Geographic restrictions — EU/EFTA/UK tenants have the Anthropic sub-processor toggle off by default.
• Government clouds are blocked entirely.

One more caveat worth knowing: Cowork is still in Research Preview via the Frontier program, with no committed GA date, so treat these practices as evolving.

If you're extending Cowork with reusable skills — for example, Microsoft's skill-recorder converts recorded work sessions into Skills the agent consumes — remember that published research found roughly 26% of agent skills carry at least one vulnerability. SkillzBench is building a free directory with security scanning and sandbox benchmarking to close exactly that gap; browse and test verified skills there before equipping any agent with them.

Extending Cowork with Skills: Where Coding Speed Really Comes From

If you came here looking for Cowork's killer coding use cases, here's the honest answer: there aren't any documented yet. Every workflow Microsoft and its partners describe — calendar triage, meeting packets, company research — is Microsoft 365 productivity work, not software development.

The coding-adjacent value lives one layer out, in the skill ecosystem growing around Cowork. The most concrete example is microsoft/skill-recorder, a desktop app with 4.2k stars that records your on-screen work session, reconstructs it as an intent plus ordered steps via the GitHub Copilot SDK, and builds a reusable Skill or Automation for agents like Cowork. That turns "how I actually did this task" into something an agent can repeat.

Copilot Studio extends this further. A September 2026 preview added a Copilot node that lets workflows invoke a Cowork task, grounded in the connected user's mail, files, calendar, and chats. Third-party agents from Adobe, Monday.com, and Figma also connect through open standards including MCP, per VentureBeat's coverage.

This is the pattern that actually accelerates work:

• Record real work once, replay it as a reusable Skill
• Chain Cowork tasks into larger Copilot Studio workflows
• Pull in specialized third-party agents via MCP

Here's the catch: skills are instruction packages agents execute, and that makes them a security surface. Published research found roughly 26% of skills carry at least one vulnerability — a meaningful risk when a skill can read your mail and files.

That's where verification matters. SkillzBench is building a free, independent layer for exactly this: commit-SHA version pinning, security scanning, and sandbox benchmarking that publishes each skill's Lift, plus a free MCP connector so agents only see the skills you've vetted. If you're building out a Cowork-adjacent skill stack, browse and test verified skills free on SkillzBench rather than pasting in whatever tops GitHub search.

Should Your Team Adopt Cowork Now? Pricing, Availability, and a Sanity Check

Before you budget a single dollar for Cowork, it's worth pausing on what "adopt" actually means right now. Copilot Cowork was announced March 9, 2026, but it is not broadly available: it's in a Research Preview with limited customers, with broader access through the Frontier early-access program that licensed tenants can opt into via the M365 Admin Center. Microsoft's own announcement is internally inconsistent on this point, and Stoneridge Software's practitioner account confirms there is no committed general-availability date — analyst estimates point to Q2 2026, but that's a guess, not a promise.

The pricing picture is clearer, though it requires a decision about timing. Cowork sits within Microsoft 365 Copilot, which costs a $30 per user per month add-on on top of existing licenses. Microsoft is also launching a new E7 bundle at $99 per user per month on May 1, 2026, which packages Copilot with Agent 365, Entra Suite, and the full E5 security stack. If your renewal calendar lines up with May, the bundle math may favor waiting; if you're already paying for Copilot, the Frontier opt-in costs nothing extra.

A few claims circulating in coverage deserve a flag before they reach your leadership deck. Reports of multi-model "Critique" and "Council" features — one model drafting while a second reviews, or parallel models compared by a judge — plus specific model names like Claude Sonnet 4.6, come from a single secondary analysis, not Microsoft's documentation. Treat them as reported, not verified.

Then there's the honest counterpoint. Alongside the enthusiasm, CNET reports that one study found AI may actually lengthen workdays and make them less enjoyable, and workers worry about job replacement. Microsoft's own time-savings claims — like Charles Lamanna's 40-minute calendar triage saving "hours worth of time" — are vendor statements without independent verification.

So here's a neutral framework for the go/no-go call:

• Confirm eligibility: licensed Copilot tenant, admin opt-in to Frontier, and awareness that EU/EFTA/UK tenants have the Anthropic sub-processor toggle off by default.
• Audit permissions first — Stoneridge calls permissions oversharing the biggest enterprise risk, and "governance-first deployment is not optional."
• Pilot on the seven documented workflows: calendar triage, meeting prep packets, cited company research, launch planning, blog drafting, business analysis with Excel models, and 1:1 summaries.
• Measure results yourself — time spent, output quality, rework — rather than trusting vendor anecdotes.
• Extend only with skills you've verified, since published research found ~26% of agent skills carry at least one vulnerability.

That last point is where the ecosystem matters. Cowork's value grows through skills and automations — Microsoft's own skill-recorder converts recorded work sessions into reusable Skills via the GitHub Copilot SDK. SkillzBench is building a free directory that verifies, security-scans, and benchmarks agent skills with published Lift, so you can equip agents with skills that have actually been tested. Join the waitlist now, pilot the documented workflows in the preview, and let your own measurements — not the hype cycle — decide when to scale.