Microsoft’s Project Opal Puts AI in the Driver’s Seat of Work

Reading Time : 3 minutes

Microsoft has quietly moved the goalposts on what a digital assistant is supposed to do. In a briefing led by chief executive Satya Nadella, the company unveiled Project Opal, a system built to plan and carry out long-running tasks inside an environment the user still controls. The framing is deliberate: where Copilot once answered questions and drafted text, Opal is positioned to take ownership of work that unfolds across hours or even days, marking a clear departure from the conversational help most users have grown accustomed to.

 

 

The mechanics behind Opal reveal how seriously Microsoft is taking the autonomy question. Rather than relying on a single model, the agent operates as a computer-using system that navigates web portals, fills out forms, and completes multi-step workflows much as a person would. Each task runs on a dedicated Windows 365 Cloud PC, and a separate AI supervisor watches every action, enforces guardrails, and pauses to request human input when a judgment call arises. Users can watch the work in real time, replay what the agent did, or step in and take direct control whenever they choose.

 

 

From an enterprise perspective, Microsoft has built in the kind of controls that compliance officers will want to see before trusting a machine with sensitive tasks. Opal is strictly opt-in and is managed through the Microsoft 365 Admin Center, where administrators can provision Cloud PC pools, define website allow lists that constrain where the agent can go, and decide which user groups gain access. It entered public preview in November 2025 and remains available as an early-access offering through the Frontier program for Microsoft 365 Copilot customers.

 

 

Opal also fits into a wider pattern in which Microsoft is pushing Copilot beyond chat and into territory once reserved for human specialists. Earlier this year the company added multi-model intelligence to its Researcher agent, letting users choose between OpenAI and Anthropic models for deep research. Microsoft is also assembling a unified Copilot application, targeted for broader availability by October 2026, that folds chat, coding, research, and autonomous agents into a single interface. Taken together, the moves suggest a strategy of gradually handing Copilot the ability to run tasks rather than merely suggest them.

 

Bénédicte Lin – Brussels, Paris, London, Beijing, Seoul, Bangkok, Tokyo, New York, Taipei, Hong Kong
Bénédicte Lin – Brussels, Paris, London, Beijing, Seoul, Bangkok, Tokyo, New York, Taipei, Hong Kong

 

#Microsoft #ProjectOpal #Copilot #AutonomousAI #ArtificialIntelligence #Microsoft365 #AITechnology #FutureOfWork #TechNews #EnterpriseAI