The Truth / scorecard

Cowork: the future arrived early, and it's gated

"The first Copilot that acts instead of answers — visibly the future of the suite, currently fenced behind Frontier preview with rough edges to match."

8/10

What's genuinely good

  • It's agentic for real: give it an outcome and it works through Word, Excel, PowerPoint, PDF, email, scheduling, and search skills on its own
  • Custom SKILL.md skills — drop a markdown playbook in OneDrive and Cowork executes your team's procedure — the biggest adoption lever since prompts existed
  • The approval flow is genuinely well-designed: risk-leveled gates before consequential actions, so autonomy never means surprise
  • Scheduled prompts plus pause/resume turn it into a colleague with a standing to-do list, and outputs land tidily in a OneDrive Cowork folder

What sucks

  • Frontier preview only — most of the people reading this can't have it yet, and that's the single biggest thing wrong with it
  • Preview rough edges are real: connections drop mid-task, and some steps execute with less visibility than the approval model promises elsewhere
  • Five scheduled prompts maximum — you'll hit that cap in your first enthusiastic week
  • Custom skills silently cap at 50 per user (1MB each, 20 companion files) — fine today, a wall the moment a team starts sharing a library
  • It ruins the rest of the suite: one day of delegating outcomes makes plain Copilot Chat feel like a search box

The honest assessment

Every other scorecard on this site reviews a tool that answers. Cowork is the first one that does. You hand it an outcome — “turn these three reports into a board deck and draft the cover email” — and it plans, opens the built-in skills it needs (Word, Excel, PowerPoint, PDF, email, scheduling, enterprise search, deep research, and more), executes the steps, pauses at gates for your approval, and files the outputs in your OneDrive Cowork folder. The first time you watch it work, the question stops being “is this good?” and becomes “why does anything else in the suite still work the old way?”

Two design choices earn the 8 on their own. The approval model is the one Microsoft got most right: actions carry risk levels, and consequential ones — sending, scheduling, anything that leaves the sandbox — stop and ask. It’s the difference between autonomy you can deploy and autonomy you have to babysit, and it reads like it was designed by people who’ve been burned before.

The second is custom skills, and it’s bigger. Drop a SKILL.md file — markdown instructions plus up to 20 companion files — into OneDrive under /Documents/Cowork/skills/, and Cowork executes your procedure: your QBR format, your offboarding checklist, your proposal template, done your way every time. This is the biggest adoption lever since prompting itself, because it moves the encoding of “how we work here” from prompt-engineering folklore to a file a team lead can write, version, and hand out.

Now the deductions. Cowork is Frontier preview only — most readers cannot get it regardless of budget, which makes every paragraph above a preview of someone else’s present. The preview texture is real too: connections drop mid-task (pause/resume softens this, but you’ll meet it), and occasionally a step executes with less narration than you’d like — a step-opacity gap that sits oddly next to the otherwise excellent approval design. The quotas pinch fast: five scheduled prompts maximum, skills capped at 50 with no warning until you’re there, and a 250,000-character input limit that big source documents will find. None of these are scandals; all of them say “preview.”

The workarounds that change the score

  1. Write skills before prompts. The unit of value here isn’t a clever prompt, it’s an encoded procedure. Take your three most repeated multi-step chores, write each as a SKILL.md, and Cowork goes from impressive to indispensable in a week.
  2. Spend the five scheduled slots on digests, not tasks. A daily briefing, a weekly status compile, a Monday inbox triage — recurring synthesis jobs amortize the scarce slots best. One-off work doesn’t need a slot at all.
  3. Chunk big inputs. The 250,000-character limit is generous until it isn’t; split monster documents and let a skill define the assembly step.
  4. Treat approval gates as review points, not speed bumps. The gate showing you a drafted email before sending is your quality-control moment. Teams that click through it reflexively are opting out of the safety model on purpose.

What Microsoft won’t tell you

  • The skills folder is the governance story nobody’s written yet: SKILL.md files are just markdown in OneDrive — easy to share, easy to copy, and currently outside any review process your org has. Decide who curates the team library before it curates itself.
  • The 50-skill and 5-schedule caps fail silently-ish — you find them by hitting them. Inventory before you evangelize.
  • The real strategic tell: Cowork makes Copilot Chat feel legacy on day one. If you run adoption, plan for the morale problem of showing people a tool most of them can’t have yet.

Bottom line

Score the trajectory, not just the build: Cowork is the clearest statement Microsoft has made about where this whole platform is going — from answering questions to executing work under supervision. The preview roughness and the quotas are temporary; the Frontier gate is the only thing keeping this from being the most important scorecard on the site. If you’re in the program, build skills now — the head start is real. If you’re not, this page is your preview of next year’s job.

← All scorecards