Microsoft launches AI transparency effort with more than 40 organizations
The initiative aims to let businesses check agent activity without relying on a model’s own account, with planned integration into Microsoft products.
Loading page…
The initiative aims to let businesses check agent activity without relying on a model’s own account, with planned integration into Microsoft products.
Listen to this story
Microsoft launched the AI Transparency Initiative at Ignite on November 17, 2026, as a standards-development and product-integration effort—not a live independent-verification system. Its proposed approach would let organizations compare an agent’s account of its actions with evidence from runtime, identity, and policy systems. More than 40 organizations are involved, and Microsoft plans to apply the work to Microsoft AI, Foundry, and Copilot. If developed as intended, portable evidence and controls could help customers audit agents across providers and retain governance when switching models.
The work has three routes into practice: NIST standards research, cross-industry collaboration, and planned Microsoft product integration.
Proposed records cover what an agent was authorized to do, which model and software it used, what data and tools it accessed, and what actions it took.
Microsoft wants the approach to cover closed and open-weight models, with organizations retaining evidence and control over data access when they change providers.
Microsoft wants businesses to verify what AI agents do without taking the models at their word. At Ignite on November 17, 2026, it launched the AI Transparency Initiative, a collaboration it says includes more than 40 organizations. The announcement sets out a standards-development effort and planned product integration, rather than a completed verification system.
The problem, as Microsoft describes it, is a gap between observing activity around a model and reconstructing how a request became an outcome. Agents can use company data, tools, credentials and infrastructure without seeking approval at every step. Yet existing controls cannot see what happens inside models or independently validate their explanations. Microsoft says that gap creates security, compliance, audit and accountability risks.
The participating organizations include Accenture, Adobe, KPMG, Mastercard and Volvo Cars. Microsoft says the goal is a common, open approach to transparency, security and governance across providers, rather than controls that apply only to one company’s models.
In his published Ignite keynote, CEO Satya Nadella described the work as an ambition to establish an open industry standard. He named MAI models, Foundry and Copilot as intended destinations. Those are commitments to develop and apply the standards, not a claim that independent verification is already operating across those products.
And what the model tells us about itself is not independent evidence.
Satya Nadella, Microsoft chairman and CEO, in his published Ignite keynote
Microsoft organizes its proposed approach around three tasks: see, verify and control. Visibility starts with recording what an agent was asked and authorized to do, which model and surrounding software ran it, what data and tools it accessed, and which actions it took. The company also calls for continuous monitoring of agent activity and risk signals against organizational policies and thresholds.
Verification depends on distinguishing where each piece of evidence came from. A model’s account, a record from the software running it, and an observation from an identity or policy system are not interchangeable. Microsoft says records should identify their source so organizations can check model-reported signals against independently observed activity, using their own security and compliance systems rather than relying solely on the model provider.
The control layer is intended to act before something consequential happens, not merely document it afterward. Microsoft says evidence about a workflow’s purpose, owner, data and delegated authority should help organizations enforce their policies. The same records should support investigations, containment, recovery and audits when teams need to reconstruct an incident.
Portability is another explicit design goal. The initiative is intended to cover both closed and open-weight models, allowing developers to replace or update models without undermining security governance or audit access. Microsoft says organizations should retain their evidence and control over agent data access when they change models or providers.
For now, Microsoft is inviting organizations to help shape the initiative or follow its progress. Accountability, its announcement stresses, remains with the organization deploying the agent—not the AI.
Loading discussion...
Join the conversation
Explain whether portable records should be a buying requirement.
Be the first to share a perspective or an experience.
Reader comments
Newest comments first. Replies stay oldest first.