Microsoft launched Copilot Cowork on June 16, 2026, permanently shifting its agentic AI assistant for Microsoft 365 to a usage-based billing model. The move ends unlimited AI access for business users, with the company now charging per task or token consumed. Internal testing reportedly demonstrated that flat-rate subscriptions would bleed cash at scale, forcing the pivot to a metered approach. Copilot Cowork isn’t another chatbot—it’s an autonomous digital coworker that executes complex, multi-step workflows across Outlook, Teams, Excel, and the entire Microsoft 365 ecosystem.

What Makes Copilot Cowork Different

Copilot Cowork represents the next generation of agentic AI. Unlike its predecessor, Microsoft 365 Copilot, which primarily generated text and answered questions, Cowork takes initiative. It can monitor an inbox for vendor emails, cross-reference them with contracts in SharePoint, draft responses in Word, and schedule follow-up meetings—all without human hand-holding. Each completed workflow, or “task,” becomes a billable event. Microsoft defines a task as a unit of autonomous work: drafting a document, updating a spreadsheet, or coordinating a Teams meeting. The pricing structure is granular; simple actions like summarizing an email cost a fraction of a full task, while multi-app orchestrations consume multiple tokens.

The assistant’s autonomy relies on large language models (LLMs) and orchestration logic hosted in Azure. Microsoft’s internal testing, according to sources familiar with the matter, saw some power users triggering thousands of tasks per day when unmetered. At scale, the compute costs for unlimited plans threatened margins. Usage-based billing aligns cost with value, ensuring organizations pay only for what Cowork does, not for idle potential.

Why Agentic AI Demands a Meter

Agentic AI breaks the traditional SaaS pricing model. A subscription to Office 365 assumes a relatively fixed cost per user, regardless of how many emails they send or formulas they calculate. But an AI that can launch a dozen background processes simultaneously doesn’t fit that mold. Large model inference is expensive; every call to a model like GPT-4 or DeepSeek V4 burns cloud compute time. When an agent chains dozens of prompts to complete a complex task, the meter ticks rapidly.

Microsoft needed a billing mechanism that reflected real resource consumption. Usage-based pricing is already common in IaaS and PaaS clouds—think Azure Functions or AWS Lambda. Bringing this pay-as-you-go philosophy to productivity software is a logical, if controversial, step. It forces businesses to think about AI utilization as an operational expense rather than a flat overhead.

DeepSeek V4: The Engine Behind the Curtain

Part of the cost calculus involves model choice. Microsoft has integrated DeepSeek V4, a highly efficient open-weight model developed by Chinese AI lab DeepSeek, as one of the engines powering Copilot Cowork. DeepSeek V4 boasts benchmark scores rivaling GPT-4o but with a reported 50% reduction in inference cost. Its mixture-of-experts (MoE) architecture activates only a subset of parameters per task, slashing compute cycles. By combining DeepSeek V4 with Azure’s infrastructure, Microsoft can offer agentic features at a lower per-task price.

The inclusion of DeepSeek V4 also signals a pragmatic shift. Instead of relying solely on OpenAI’s models, Microsoft is diversifying its AI supply chain. DeepSeek V4’s open-source nature allows for fine-tuning on enterprise data without the premium of proprietary API calls. For tasks like data analysis in Excel or email triage, the model performs admirably, while more creative or nuanced tasks may still fall to GPT-4o or a proprietary Microsoft model. The metering can even differentiate: tasks labeled “standard” use DeepSeek V4 at a lower unit cost, while “premium” tasks call on larger models for a premium price.

How the Pricing Model Works

Microsoft’s pricing for Copilot Cowork is based on “task units” (TUs). Each TU corresponds to a fixed amount of compute—roughly equivalent to one agentic action or a single model call. A simple query like “summarize my last three emails” might consume one TU. A complex task like “find all pending contracts, draft responses, and schedule reviews for next week” could consume twenty. Organizations purchase TUs in packs, with volume discounts kicking in at 10,000, 100,000, and 1 million TU tiers. Overage is possible but capped to prevent runaway bills. The baseline per-TU price is $0.15, placing the cost of a typical daily workflow at a few dollars per user.

This stands in contrast to Microsoft 365 Copilot’s flat $30 per user/month fee. For heavy users, Cowork’s metered plan could quickly exceed that, but for average users, it might be cheaper. Microsoft’s bet is that customers will appreciate only paying for what they use, while the company protects margins from super-users. Early access customers reported mixed feelings: some appreciated the flexibility, while others missed the predictability of a flat fee.

Microsoft isn’t alone in exploring metered agentic AI. Google’s Duet AI for Workspace, launched in early 2026, also employs a usage-based model, billing per “action item.” Slack AI charges per resolution for its autonomous workflows. The broader SaaS industry has been inching toward consumption-based pricing for years, and AI is accelerating the shift. A Gartner report from January 2026 predicted that by 2028, 70% of enterprise AI tools would use some form of metering, up from 20% in 2025.

For Microsoft, the meter also serves as a competitive moat. By tying Copilot Cowork so tightly to Azure, it creates a feedback loop: more AI usage drives more Azure consumption, fattening Microsoft’s cloud revenue. The integration of DeepSeek V4 helps keep costs down, making the service viable for price-sensitive small and medium businesses.

Reactions and Business Implications

Corporate IT buyers reacted cautiously. Some applaud the transparency: “We can finally tie AI spending to actual business outcomes,” said one CIO of a mid-sized logistics firm. But many finance departments fear unpredictable budgets. Microsoft has attempted to mitigate this with built-in cost controls and analytics dashboards in the Microsoft 365 admin center, showing real-time TU consumption per user and per department.

The shift also has implications for employee productivity tracking. Since Cowork logs every autonomous action, it paints a detailed picture of which tasks are being automated. Managers could use this data to identify repetitive work and redesign processes—a double-edged sword if tied to performance metrics.

The Road Ahead

Microsoft plans to extend Copilot Cowork’s agentic reach beyond Office. Teasers at Build 2026 hinted at integration with Power Platform, Dynamics 365, and even Azure DevOps. Each expansion will bring its own metering nuances. The company also hinted at “bring-your-own-model” capabilities, where enterprises can plug in their own fine-tuned DeepSeek V4 or other open models, paying only for the AI runtime.

Usage-based billing may become the norm for AI-powered software. As models grow more capable and autonomous, the line between software license and compute consumption blurs. Microsoft’s Copilot Cowork is the first major productivity suite fully embracing that reality. For businesses, the message is clear: the AI coworker is a powerful ally, but you’ll pay by the task—whether you’re ready or not.