At Build 2026 in June, Microsoft made Copilot Cowork generally available worldwide — an agentic system that plans, executes, and delivers completed work rather than drafts — and unveiled Microsoft Scout, its first "Autopilot" agent, in preview.
What Cowork actually does
Where Copilot Chat answers questions, Cowork takes a defined task and executes it start to finish across multiple apps, running in a secure cloud-hosted environment that keeps working even when the user's laptop is off. It ships with 13 built-in skills spanning Word, Excel, PowerPoint, PDF, email, scheduling, meetings, and enterprise search, and is grounded in a company's own systems through what Microsoft calls Work IQ.
The pricing story
Cowork is off by default; an admin enables it, decides who gets it, and sets spending limits at the tenant, group, and user level. Billing is consumption-based at $0.01 per Copilot Credit, with each task's cost derived from four inputs: the model used, the context pulled in, the tools called, and how long the task runs. This is the clearest example yet of AI software abandoning per-seat pricing for per-work pricing — a shift with real consequences for how IT budgets are built.
Multi-model by design
Cowork selects models per task: OpenAI's GPT-5.5 Thinking for deep research with citations, Anthropic models for visual work like PowerPoint. New plugins announced at GA include Harvey, Miro, monday.com, Moody's, Morningstar, and S&P Global.
The bigger picture
Between Cowork's GA and Scout's preview, Build 2026 marked the moment Microsoft stopped positioning agents as an experiment. The question enterprises now face is not whether to pilot agentic AI, but how to govern spend on software that bills like a contractor.