Microsoft made Copilot Cowork generally available worldwide today, introducing usage-based pricing that company testing shows runs 30-40% cheaper than Claude's Microsoft 365 connector.
More than half the Fortune 500 already uses the agentic AI system after a three-month preview. Now every Microsoft 365 Copilot subscriber can access multi-step, autonomous workflows—but the shift to consumption-based billing means CFOs need new budget models and CTOs need governance frameworks before deployment.
For CTOs: Copilot Cowork runs tasks end-to-end with no human intervention. You delegate work, it executes using M365 apps, tools, and data. Built on Anthropic models (Opus 4.8, Sonnet 4.6). Security inherits your Entra ID permissions. All actions audit-logged in Purview.
For CFOs: Pricing is per-task consumption on top of $30/user/month M365 Copilot base license. Light tasks cost $0.50-1, medium $2-5, heavy $10-25 (estimates based on Anthropic Opus 4.8). Budget model spreadsheet available. Spending caps at tenant/group/user levels prevent runaway costs.
What Copilot Cowork Actually Does
Copilot Cowork handles complex, long-running tasks across multiple M365 applications without continuous human oversight. Unlike standard Copilot (which generates drafts and recommendations), Cowork executes complete workflows and returns finished results.
Real-world examples from Fortune 500 preview customers:
Engineering automation: One team taught Cowork to edit batch-job spreadsheets, generate dependency flow charts, and validate changes—work that previously required manual intervention and domain expertise.
Version control at scale: One development team used Cowork to compare 4,000+ files across two product versions. Manual comparison: weeks. Cowork: hours.
Sales pipeline recovery: A sales leader pointed Cowork at a stalled pipeline. Output: ranked list of at-risk opportunities with exact follow-up gaps on each deal. Manual review time: one week. Cowork: one morning.
The technical difference: Cowork uses multi-model reasoning, calls M365 APIs, accesses SharePoint/Teams/Outlook/Excel, retrieves context from Microsoft Graph, and chains operations autonomously. It's cloud-hosted (keeps running when your laptop's off) and operates within your M365 security boundary.
The Pricing Model That CFOs Need to Understand
Copilot Cowork requires a Microsoft 365 Copilot User Subscription License ($30/user/month) plus usage-based charges denominated in Copilot Credits.
Four factors determine task cost:
- Model use — Which AI model processes the task (Opus 4.8 most expensive, upcoming Cowork 1 model cheapest)
- Context retrieval — How much data Cowork pulls from Graph, SharePoint, email
- Tool calls — How many M365 app actions (send email, update spreadsheet, schedule meeting)
- Runtime — How long the task runs
Microsoft's cost framework identifies three task types:
Light tasks ($0.50-1 estimate):
- Small knowledge base (1-2 sources)
- Limited reasoning
- 0-1 outputs
- Example: "Summarize this email thread and draft a 2-sentence reply"
Medium tasks ($2-5 estimate):
- Multiple sources
- Structured reasoning
- 2+ outputs
- Example: "Analyze Q2 sales data, identify underperforming regions, and create action plans for each VP"
Heavy tasks ($10-25 estimate):
- Broad aggregation across many sources
- Deep reasoning with multiple decision points
- Many outputs
- Example: "Compare 4,000 files across product versions, document differences, flag breaking changes, and generate migration guides"
Four user personas drive different cost profiles:
- Power users: Frequent heavy tasks (10-15% of user base in preview)
- Regular users: Mix of light/medium tasks (40-50%)
- Occasional users: Primarily light tasks (30-35%)
- Experimental users: Sporadic use, testing workflows (5-10%)
Microsoft provides a budget estimator spreadsheet for modeling costs. The formula: (users per persona) × (task volume by type) × (cost per task type) = monthly spend estimate.
Cost Management: The Controls That Matter
Microsoft ships three layers of cost control at GA:
1. Access Control
- Off by default — Admins enable Cowork per tenant
- Group-based licensing — Assign to specific teams/roles
- Per-user permissions — Not everyone needs agentic AI
2. Spending Limits
- Tenant-level caps — Hard limit for entire organization
- Group policies — Department budgets (e.g., Engineering $5,000/month, Sales $2,000/month)
- User-level limits — Individual spending caps within group budgets
- Credit request workflow — Users can request additional credits when needed (requires admin approval)
3. Visibility & Alerts
- Usage reports — Tenant, group, and user-level breakdowns
- Per-task pricing — See exact credit cost for each Cowork execution
- Custom thresholds — Set alerts at 50%, 75%, 90% of budget
- Real-time notifications — Admins notified when spend crosses thresholds
The governance risk: An over-permissioned agent with unlimited credits could send thousands of emails, download entire SharePoint sites, or generate massive compute bills. Microsoft's control plane addresses this with OAuth 2.0 delegated permissions (agent inherits user's access only), Purview audit logs (every action recorded), and adaptive authentication (if user session revoked, agent access cut immediately).
Model Choice: Anthropic Now, GPT and Cowork 1 Soon
At GA, Copilot Cowork runs on Anthropic models:
- Opus 4.8 — Frontier intelligence, highest cost, best for complex reasoning
- Sonnet 4.6 — Balanced cost/performance for most enterprise tasks
In Frontier preview (available to select customers):
- GPT 5.5 — OpenAI's latest, comparable cost to Opus 4.8
Coming in weeks:
- Cowork 1 — Microsoft's fine-tuned model optimized for Cowork tasks at "substantially lower cost" than Opus/GPT
The strategic angle: Multi-model design means Microsoft can route tasks to the most cost-efficient model for each workload. Light tasks → Cowork 1 (cheap). Heavy reasoning → Opus 4.8 (expensive but accurate). This cost optimization happens automatically as more models become available.
Microsoft claims Copilot Cowork averages 30-40% cheaper per prompt than Claude Cowork with Microsoft 365 connector. Testing compared equivalent tasks. Cost advantage comes from efficient context retrieval, multi-model routing, and native M365 integration (no external API calls).
Security & Compliance: What CISOs Get (and Don't)
Copilot Cowork operates within the Microsoft 365 trust boundary, meaning:
Security inheritance:
- Entra ID permissions — Agent inherits user's access scope, not super-user
- OAuth 2.0 delegation — No password storage, token-based auth
- Conditional access — If user session revoked, agent access terminates immediately
- Data Loss Prevention (DLP) — Cowork blocked from accessing/moving content labeled confidential or higher
Audit & compliance:
- Purview integration — Every agent action logged with full reasoning trace
- eDiscovery support — Agent interactions discoverable, retainable, deletable per policy
- Microsoft Defender integration — SOC can monitor agent behavior, set alerts for anomalies (e.g., downloading thousands of files)
- Sentinel integration — Auto-suspend agent permissions if unusual patterns detected
The compliance framework:
- GDPR: Data lifecycle management through Purview
- HIPAA: Agent actions part of organizational data estate
- EU AI Act: Constitutional AI (Anthropic) provides behavioral guardrails
- SOC 2: Agent audit trails meet control requirements
What's missing: Multi-tenant isolation details unclear. If Cowork processes data from multiple customers on shared infrastructure, how are workloads isolated? Microsoft hasn't published architectural details on tenant separation for Cowork runtime.
Fortune 500 Adoption: Who's Using It
More than 50% of Fortune 500 companies used Copilot Cowork during the three-month preview, making it "the fastest growing feature in the history of our Frontier program."
Named preview customers:
- Accenture
- Avanade
- Advance Local
- Capital Group
- Koch Industries
- LTM
- Ooredoo Qatar
- Zurich Insurance
Satisfaction metrics: Microsoft reports Cowork has "among the highest user satisfaction of any Copilot or agent experience we have shipped." No specific NPS or CSAT scores disclosed.
The adoption driver: Teams that tested Cowork found it collapsed multi-week manual work into hours or days. When one workflow saves 80+ hours of labor, ROI justifies the consumption cost even at $10-25 per heavy task.
What CTOs Should Do This Week
1. Assess current M365 Copilot coverage
- Cowork requires M365 Copilot USL ($30/user/month)
- If you're not on Copilot yet, Cowork won't change that decision
- If you are, identify which teams could benefit from agentic workflows
2. Model costs with Microsoft's estimator
- Download budget spreadsheet
- Map your users to personas (power/regular/occasional/experimental)
- Estimate task volume and mix (light/medium/heavy)
- Run scenarios with different model choices (Opus 4.8 vs upcoming Cowork 1)
3. Set governance policies before enabling
- Access: Which groups get Cowork? (Start narrow: pilot with 10-20 power users)
- Spending caps: Tenant limit? Group budgets? Per-user caps?
- Alerts: Who gets notified at 50%/75%/90% of budget?
- Approval workflows: Do heavy tasks require supervisor approval?
4. Enable Purview logging and Defender monitoring
- Turn on audit logging for all Cowork actions
- Set up Sentinel alerts for anomalies (file download spikes, external email volume)
- Integrate with SOC monitoring dashboards
5. Run pilot with defined use cases
- Pick 2-3 workflows with measurable manual effort (hours/week)
- Assign to pilot users, track task costs and time savings
- Measure ROI: (labor cost saved) ÷ (Cowork consumption cost)
- Scale to broader deployment if ROI positive
What CFOs Should Budget For
Base cost: $30/user/month M365 Copilot license (unchanged)
Variable cost: Cowork consumption based on task volume and complexity
Estimating framework:
- Conservative scenario (mostly light tasks): $5-10/user/month additional
- Moderate scenario (mix of light/medium tasks): $15-25/user/month
- Aggressive scenario (frequent heavy tasks): $40-60/user/month
Example: 500-user deployment, moderate usage
- Base: 500 × $30 = $15,000/month
- Cowork consumption: 500 × $20 = $10,000/month
- Total: $25,000/month ($300K/year)
ROI calculation: If each user saves 4 hours/month through automation at $75/hour loaded cost:
- Labor savings: 500 × 4 × $75 = $150,000/month
- Net benefit: $150K - $25K = $125,000/month
- Annual ROI: $1.5M savings on $300K spend = 5x return
The budget risk: Usage-based pricing means costs can spike unexpectedly. One team running thousands of heavy tasks could blow through monthly budget in days. Spending caps and alerts mitigate this, but require proactive monitoring.
The Competitive Landscape
Microsoft vs Anthropic (Claude Cowork):
- Microsoft claims 30-40% lower cost per prompt
- Microsoft: native M365 integration, no external APIs
- Anthropic: M365 connector requires API calls, higher latency
Microsoft vs OpenAI (ChatGPT Enterprise):
- ChatGPT Enterprise: seat-based pricing ($60/user/month), no usage charges
- Copilot Cowork: $30 base + consumption
- Crossover point: If users run >$30/month in Cowork tasks, ChatGPT Enterprise cheaper for heavy users
Microsoft vs Salesforce (Agentforce):
- Salesforce: CRM-focused agents, usage-based pricing similar to Cowork
- Microsoft: M365-native, broader productivity scope
- Different TAM: Salesforce = revenue teams, Microsoft = knowledge workers
The strategic question for CIOs: Do you standardize on one vendor's agentic platform (Microsoft, Anthropic, OpenAI, Salesforce) or adopt multiple for different use cases? Multi-vendor approach gives flexibility but increases governance complexity and training overhead.
Bottom Line
Microsoft Copilot Cowork GA marks the shift from AI copilots (assistants) to AI coworkers (autonomous agents) in enterprise productivity.
For CTOs: The technology works—Fortune 500 validation in preview proves it. The governance challenge is new: you're granting non-human entities delegated access to corporate systems. Start with narrow pilots, strict spending caps, and comprehensive audit logging.
For CFOs: Usage-based pricing creates budget uncertainty but aligns cost with value. Model conservatively, set caps aggressively, measure ROI religiously. If labor savings exceed consumption costs by 3x+, scale deployment. If not, limit to high-value use cases only.
For CIOs: Cowork represents a platform bet on Microsoft's enterprise AI stack. It requires M365 Copilot ($30/user/month), ties you deeper into the Microsoft ecosystem, and creates switching costs. Evaluate against alternatives (Claude, ChatGPT Enterprise, Salesforce Agentforce) before committing to broad rollout.
The market signal: Microsoft shipping agentic AI at enterprise scale with Fortune 500 adoption validates the category. Every productivity vendor will now race to match or undercut Microsoft's cost structure. Expect competitive pricing pressure through 2026.
