DEEP DIVE · 03 · THICKEST
Claude Cowork
the full research
GA since 2026-02-24. An autonomous desktop agent aimed at non-engineer knowledge workers. Crosses local files, Gmail, and Drive. OSWorld jumped from under 15% in late 2025 to 72.5% by February — a fivefold gain in a year. The optional part of the workshop, but the one with the biggest potential impact on exec work.
GA
Mac / Win
Team / Enterprise RBAC
1. What Claude Cowork is
Claude Cowork — official spelling Cowork, not CoWork or Co-Work — is Anthropic's desktop automation agent. It's built for knowledge workers doing non-technical jobs: analysts, lawyers, finance, researchers.
How it sits in the product line
- Runs inside Claude Desktop — a capability of the macOS and Windows app
- Not the same as Claude Code. Code is for developers (VS Code, terminal). Cowork is for everyone else (local files, apps, browser).
- Design philosophy. "Outcome-centric" rather than "prompt-centric." You describe what you want. Claude plans the steps and runs them.
- Permission scope. Folder-based. Only the folders you allow, only what's in them.
Core capabilities
| Capability | What it does |
|---|---|
| File organization | Rename, sort, dedupe documents |
| Document generation | Merge multiple sources into a structured draft |
| Data extraction | Contracts and reports turned from prose into structured data |
| Research synthesis | Search, summarize, and combine across sources |
| Process automation | Excel models, decks, multi-step workflows |
2. Release history
Research Preview (2026-01-12)
Announced on 2026-01-12 as a research preview. macOS only, Pro and Max subscribers.
General Availability (2026-02-24)
GA on 2026-02-24. Same day, Anthropic shipped:
- 13 new enterprise connectors — Google Calendar, Google Drive, Gmail, DocuSign, FactSet, and others
- Cross-application context — tasks spanning Excel, PowerPoint, and other apps
- Private plugin marketplace for in-house custom plugins
- Industry templates for finance, legal, HR, operations
OSWorld benchmark trajectory
| When | Score | Model | Note |
|---|---|---|---|
| Late 2024 | <15% | Claude Opus 4.5 | Early Computer Use |
| Feb 2026 | 72.5% | Claude Sonnet 4.6 | At GA |
| Apr 2026 | 78.0% | Claude Opus 4.7 | OSWorld-Verified |
OSWorld is a 369-task benchmark covering file management, web, Office apps, multimedia, and OS operations. Human baseline: 87%.
3. Architecture
Hybrid by design
Cowork runs as a local + cloud hybrid.
Local side
- File operations — read, edit, create, delete, only in allowed folders
- Screen operations — window control, mouse, keyboard
- Browser automation — a controlled Chrome instance
- OS commands — a restricted set
Cloud side
- Inference — Claude Opus / Sonnet for planning and judgment
- State management — context preserved across multi-step tasks
- Connector brokering — API calls to FactSet, DocuSign, etc.
Internet is required
Cowork talks to Anthropic's cloud constantly. If the connection drops, the local VM keeps running but the cloud loses context — which means you can end up in an inconsistent state.
4. Connector catalog (as of 2026-04-21)
13 connectors GA today. Jira and ServiceNow are on the enterprise roadmap in beta, not yet on the GA list.
| Connector | Capabilities | Use case |
|---|---|---|
| Google Drive | Read, generate files | Document search and synthesis |
| Gmail | Search mail, draft replies | Triage and response management |
| Google Calendar | Query schedule, create events | Calendar ops, coordination |
| DocuSign | Read contracts, extract key terms | Risk review, redline proposals |
| FactSet | Company and market data lookup | Financial analysis, comp tables |
| Slack | Channel and message search | Knowledge extraction, sharing |
| Salesforce | CRM read and update | Sales data ops |
| LegalZoom | Legal template search | Contract templating |
| Apollo | Lead database search | Sales and marketing automation |
| Clay | Data integration and enrichment | Sales intelligence |
| Outreach | Sales sequences | Sales automation |
| MSCI | ESG data | ESG reporting |
| WordPress | Site management | Blog publishing automation |
5. Plans and pricing
Individual plans
| Plan | Monthly | Cowork limits | Context | Best for |
|---|---|---|---|---|
| Pro | $20 | Small | 200K tokens | Base users |
| Max 5x | $100 | Medium | 200K tokens | Power users |
| Max 20x | $200 | Large | 200K tokens | High volume |
Team plans
| Plan | Min seats | Standard seat | Premium seat | Features |
|---|---|---|---|---|
| Standard | 5 | $20/month | N/A | Claude chat only |
| Premium | 5 | $20/month | $100/month | Code + Cowork |
Enterprise
| Item | Detail |
|---|---|
| Seat price | Sales quote (custom) |
| Extended context | 500K tokens |
| RBAC | Groups, role definitions, per-feature control |
| HIPAA | Available on agreement |
6. Comparisons
Claude Cowork vs. Microsoft Copilot Cowork
| Attribute | Claude Cowork | Microsoft Copilot Cowork |
|---|---|---|
| Release | 2026-02-24 (GA) | 2026-03-09 (research) / 03-30 (Frontier) |
| Base model | Claude Opus / Sonnet | Claude, supplied by Anthropic |
| Host | Claude Desktop (standalone) | Inside Microsoft 365 |
| Local files | Full | M365 files only |
| Browser automation | Yes | Outlook / Teams only |
| App coverage | General (any desktop app) | Outlook, Teams, Excel, Word, PowerPoint |
| Price | $20–200/month (individual) | $99/month (E7 inclusive) |
Claude Cowork vs. ChatGPT Operator
| Attribute | Claude Cowork | ChatGPT Operator |
|---|---|---|
| Interface | Claude Desktop | Web / API |
| File access | Local filesystem | Cloud storage |
| Browser automation | Yes | Yes — stronger for web tasks |
| Security model | Local sandbox + cloud inference | Cloud agent |
7. Twenty exec use cases
Five each for the CEO, CFO, Legal, and IR. Time savings are industry-benchmark rough estimates.
CEO (5)
| # | Task | Output | Time saved |
|---|---|---|---|
| 1 | Weekly exec brief | Executive summary rolled up from multiple division reports | 4h → 30min |
| 2 | Board deck QA | Cross-version number and layout consistency checks | 2h → 15min |
| 3 | Media mentions | Google Alerts rolled into a Google Doc | 15min/day → 1min |
| 4 | Advisor-interview extraction | Transcript to key quotes and action items | 2h → 20min |
| 5 | Earnings Q&A draft | Prior-quarter answers and IR drafts into new-quarter candidate Q&A | 3h → 40min |
CFO (5)
| # | Task | Output | Time saved |
|---|---|---|---|
| 6 | Expense categorization | Multi-currency receipt images to CSV | 3h/month → 10min |
| 7 | Vendor contract review (first pass) | DocuSign integration, deltas vs. standard terms | 2h → 20min |
| 8 | Three-month rolling forecast | Last month's actuals + division projections into an updated Excel model | 5h → 30min |
| 9 | Subsidiary P&L roll-up | Merge and summarize subsidiary sheets | 4h → 45min |
| 10 | Bank and institutional investor Q&A prep | Prior Q&A plus drafts into candidate answers | 3h → 40min |
Legal (5)
| # | Task | Output | Time saved |
|---|---|---|---|
| 11 | NDA and standard contract risk screen | DocuSign + Harvey, risk flags | 30min/contract → 3min |
| 12 | Regulatory update memos | Regulator publications to an internal-impact memo | 2h → 25min |
| 13 | Litigation email evidence | Gmail search, timeline, summary | 4h → 40min |
| 14 | Template-update notification | LegalZoom new version to internal diff doc | 2h → 20min |
| 15 | Compliance Q&A log | Employee training questions and answers | 1h/month → 10min |
IR (5)
| # | Task | Output | Time saved |
|---|---|---|---|
| 16 | Analyst call summary | Analyst questions grouped by topic with draft answers | 3h → 35min |
| 17 | Competitor benchmark refresh | Competitor IR disclosures into a comparison table | 4h → 1h |
| 18 | ESG scorecard (annual) | MSCI / S&P data plus company data to a draft report | 8h → 2h |
| 19 | Monthly citizen investor list update | Crunchbase / Pitchbook aggregation | 2h → 15min |
| 20 | Ten-year Q&A encyclopedia maintenance | Past earnings transcripts plus FAQ tool refresh | 6h/decade → 40min |
8. MIXI deployment scenarios
Note. These are hypothetical. Any real rollout needs a proper scoping conversation with Anthropic Sales. Not built from MIXI internal data — just from how corporate strategy, legal, and IR functions typically run.
Scenario A: Strategy — weekly brief automation
Current state
Strategy team pulls a weekly brief for the exec team every Monday morning. Data from several systems and Google Sheets, hand-assembled in Google Docs, shared on Slack. Time: 3–4 hours/week
With Cowork
- Wire up market and competitor APIs (Similarweb, MSCI, etc.)
- Schedule Cowork to run Friday 17:00:
- Pull last week's numbers from Google Sheets
- Pull competitor movement from APIs
- Generate a new Google Doc from template
- Post summary to Slack
- Monday morning: 30-minute review and tweak by the strategy team
Result
Weekly work: 3–4h down to 30min. Annual savings ≈ 120–150 hours.
Scenario B: Legal — contract risk triage
Current state
Legal reviews every contract coming in from business units. 30–45 minutes per contract.
With Cowork + DocuSign
- DocuSign connector pulls unsigned contracts
- Cowork runs the first pass:
- Compare against standard terms
- Flag risk items
- Assign a level — GREEN / YELLOW / RED
- Log results in Excel
- RED and YELLOW pings Legal on Slack
Result
First-pass screen: 45min → 3min/contract. Monthly savings ≈ 7–8 hours, ≈ 100 hours/year.
Scenario C: IR — ten-year Q&A encyclopedia refresh
Current state
Ten years of investor meeting transcripts. Full refresh every 5–10 years takes 6–8 hours.
With Cowork
- Drop the last decade of transcripts (150 files) into Google Drive
- Cowork searches, categorizes, updates, and emits a markdown FAQ
- IR reviews the output (1–2 hours)
Result
Manual work: 6–8h down to 2h. 4–6 hours saved per refresh.
9. Security model
RBAC
Enterprise plan only. Not available on Pro, Max, or Team.
Audit logs
Serious gap. Cowork activity isn't written to the audit log (as of 2026-04-21). Financial supervision, legal discovery, medical workflows — anywhere a trail is legally or regulatorily required — Cowork is the wrong tool today.
| Activity | In audit log? |
|---|---|
| Claude Chat | Yes |
| Claude Code execution | Yes |
| Cowork tasks | No |
| File read/write | No |
Local file permissions
Cowork runs inside a Linux sandbox inside Claude Desktop. Read / write / delete are scoped to folders the user explicitly allowed.
10. Japan operational considerations
Data residency
Where it is now. All Cowork inference runs on Anthropic US infrastructure. No region selection. If your work has residency requirements, Cowork isn't a fit today.
Bedrock workaround
Integrating Claude via AWS Bedrock lets you process in the Tokyo region:
ANTHROPIC_BASE_URL=https://bedrock.ap-northeast-1.amazonaws.comBut Cowork doesn't support Bedrock routing. This workaround only applies to Claude API and Claude Code CLI.
APPI (Japan's data protection law)
| Requirement | Cowork status |
|---|---|
| Subject consent | Yes — folder access prompt |
| Purpose specification | Not clear |
| Safe management | Partial — sandbox isolated, but transferred to the US |
| Third-party transfer restriction | Data sent to Anthropic may qualify as "transfer" |
Recommendation for MIXI. Keep APPI-covered data (employee info, customer lists, identifiable IDs) off Cowork. Scope Cowork to non-personal work — market analysis, public-document summarization, competitor research.
AUTOMATION GOVERNANCE
Deciding how far to let automation go
"OK, Cowork can automate these things. Now — how far does the company let it go?" This is the question that burns the most clock during a Cowork rollout. What's technically possible and what the organization should allow aren't the same list.
Three Trust Levels
Every action gets sorted into one of three tiers: autonomous, semi-autonomous, manual. The whole point is to kill grey areas.
| Level | Meaning | Good for | Bad for |
|---|---|---|---|
| Watched (human in the loop) |
Agent proposes, human approves every step | Sending mail, deleting files, external charges, HR access | Nothing — but if everything lands here, you lose the productivity |
| Batched (review the output) |
Agent runs the full task, human reviews results | Document drafting, data extraction, renaming, draft emails (held before send) | Anything that goes external, anything irreversible |
| Autonomous (trust it) |
Routine work that doesn't need human eyes | Local read-only ops, standard roll-ups, calendar tidying, minutes formatting | Unbounded-cost workloads, external API calls |
Action type × default Trust Level
A starting draft for MIXI. Ready to hand to legal and IT at this level of detail.
| Action | Recommended Trust Level | Why |
|---|---|---|
| Local file read | Autonomous | Failure cost is near zero |
| Local file write (new) | Batched | Disk exhaustion, overwrite risk |
| Local file delete | Watched | Irreversible. Bad delete is expensive. |
| Gmail search / read | Batched | Scope needs a human check |
| Gmail draft | Batched | Send is a separate approval |
| Gmail send | Watched (always) | Wrong send is irreversible |
| Drive search / read | Autonomous | Read-only is fine |
| Drive new file | Batched | Verify sharing settings |
| Drive edit existing | Watched | Affects co-editors |
| Slack message send | Watched | The whole channel reads it |
| Calendar event add | Batched | Self-only vs. inviting others are different |
| Web search / scrape | Autonomous | Read-only |
| External SaaS write | Watched | Failures are public |
| Payments / billing APIs | Forbidden in Cowork | Failure mode is existential |
| Git push | Watched | History is permanent |
Blast radius thinking
Before automating anything, ask one question. "If this misfires 100 times, what happens?"
Case 1. File rename automation. 100 misfires = 100 badly-named files. Recoverable. Batched works.
Case 2. Automated customer email. 100 misfires = 100 customers got a weird message. Brand hit, potentially a legal matter. Watched, no exceptions.
Case 3. Automated invoicing. 100 misfires = 100 wrong invoices, refunds, trust destroyed. Don't run this on Cowork. Dedicated system plus dual approval.
Company policy template — MIXI draft
An A4-sized policy ready to hand to legal and IT.
# MIXI Cowork / AI Agent Automation Policy v0.1 (draft) ## 1. Scope This policy applies to Claude Cowork and equivalent autonomous AI agents used inside MIXI. ## 2. Trust Levels All automation is classified as Watched, Batched, or Autonomous. The classification is documented for each workflow. ## 3. Forbidden areas The following are not run through Cowork. They stay on existing systems with dual approval. - Payments, billing, transfers - Writes to HR or payroll data - Sending contracts with legal consequence - Customer PII sent to external APIs ## 4. Audit - Batched and Autonomous action logs retained 180 days - Monthly anomaly review by IT, reported to Strategy - Incidents escalated to the CTO within 24 hours ## 5. Review cadence Revisit this policy every six months. New features (new connectors, new capabilities) trigger an interim revision. ## 6. Violations - First occurrence: warning plus retraining - Second: Cowork access suspended - Third: disciplinary review ## 7. Revision history v0.1 (2026-04-22): initial draft
Operational honesty. Policies tend to get written and forgotten. Without a twice-yearly review and a retrospective loop from incidents, this goes stale in a year. Cowork ships new connectors quarterly — the policy has to keep up.
11. Ten adoption pitfalls
| # | Pitfall | Risk | Mitigation |
|---|---|---|---|
| 1 | No audit log for Cowork | Regulated activity can't be traced | Keep it out of regulated workflows |
| 2 | Internet required | Half-broken state when offline | Offline work uses the old method |
| 3 | NDA and training | Anthropic training on data is the default posture | Enterprise contract plus the No Training option |
| 4 | Destructive ops don't undo | Wrong delete is gone | Test environments, version control |
| 5 | No data residency | Japan PII processed in the US | Restrict to non-personal workflows |
| 6 | RBAC is Enterprise-only | Pro / Max: every user has equal power | Move to Enterprise earlier than you think |
| 7 | Plugin data leak risk | Compromised API keys become company-wide incidents | Key rotation SOP |
| 8 | File ops are irreversible | Bad renames, bad deletes | Read-only folder permissions where possible |
| 9 | OSWorld vs. reality | 72% is the ideal-app score. Real-world: 50–60%. | Start on low-stakes tasks |
| 10 | Anthropic dependency | Vendor lock-in | Keep eyes on alternatives |
12. 30-day pilot plan
Anthropic doesn't offer a formal 30-day trial program. This is a suggested self-run pilot structure.
Phase 1 — plan and permissions (Day 1–3)
| Task | Owner | Detail |
|---|---|---|
| Select participants | Strategy | 5–10 division heads plus 2 IT admins |
| Prepare folders | IT | Isolated pilot folders |
| Training materials | HR + tech | 30-minute onboarding video, FAQ |
| NDA / policy update | Legal | Explicit language on data sent to Anthropic |
| Licensing | Procurement | Pro / Team trial — 10 licenses, 30 days |
Phase 2 — initial onboarding (Day 4–7)
- Day 4. Kickoff meeting. Leadership sets the expectations.
- Day 5–6. Individual training. One hour per user, on their own use case.
- Day 7. Technical check. Security settings, connector tests.
Phase 3 — real tasks (Day 8–25)
| Team | Task | Weekly cadence | Target |
|---|---|---|---|
| Strategy | Weekly brief | 1x | Manual 3h → 30min |
| CFO | Expense categorization | 1x (month-end) | 80%+ initial accuracy |
| Legal | Contract risk review | 2–3 contracts | 50% less review time |
| IR | Analyst Q&A draft | 1x | Cut 2h off the draft |
Phase 4 — feedback and tuning (Day 26–28)
- Day 26. User survey — usability, accuracy, pain points.
- Day 27. IT and Anthropic support evaluation.
- Day 28. Report for leadership. Hours saved, rough ROI.
Phase 5 — decision (Day 29–30)
| Outcome | Threshold | Next step |
|---|---|---|
| Success | 75%+ target completion, satisfaction 4.0+/5.0 | Company-wide rollout, Q3 target |
| Promising, needs work | 50–75% completion | 2–4 more weeks |
| Miss | Below 50% or safety concerns | Stop. Evaluate alternatives. |
13. Official sources
The primary sources are in the Sources block below.
Last verified: 2026-04-22
Sources