Microsoft published its 2026 Responsible AI Transparency Report on September 1, and Chief Responsible AI Officer Natasha Crampton frames the year’s central change plainly: governance had to move beyond individual models to cover how AI systems behave once they start acting as agents. The company restructured its Responsible AI Standard by layer of the technology stack, models, platform services, and applications, pairing “core requirements that always apply” with “scenario-specific requirements that can evolve as capabilities and risks change,” rather than applying one static checklist to every AI system regardless of what it does.
What actually changed
The report’s biggest shift is agentic AI governance: new controls address agent identities, the permissions an agent holds over the tools it can call, and monitoring of the actions an agent actually takes once deployed, a response to what Crampton describes as risk that “can evolve as systems interact with their environments” rather than staying fixed at the point a model is trained. On the evaluation side, Microsoft introduced an AI Red Teaming Agent to speed up risk identification, agent-specific evaluators to measure the quality and safety of agentic applications, a tool called RAMPART that converts red-team findings into repeatable automated tests, and a runtime policy-enforcement layer built from what the company calls ASSERT and an Agent Control Specification. Microsoft also reports ISO 42001 certification, the international standard for AI management systems, across Microsoft 365 Copilot, Foundry, and GitHub Copilot.
The layered structure itself is the more durable change. Rather than one Responsible AI Standard applied uniformly, Microsoft now separates requirements that hold everywhere, at the model layer, the platform layer, and the application layer, from requirements that flex depending on what a specific system actually does and what risk it actually carries. A narrow internal tool and a customer-facing agent that can take real-world actions do not get the same checklist under this structure, even if both are built on the same underlying model. That distinction matters operationally: it is an admission that a single fixed governance policy, sized for the highest-risk use case, either over-governs low-risk internal tools into uselessness or under-governs high-risk agentic deployments by treating them the same as everything else.
Why it matters to the enterprise AI buying committee
A vendor’s own governance standard is not binding on a customer’s deployment, but it is a preview of what that vendor will expect an enterprise customer’s own controls to look like, and it is a template a CAIO or Chief AI Officer can benchmark an internal program against. Three practical takeaways:
- Ask a model or platform vendor for their equivalent of a layered standard. Microsoft’s model applies different requirements at the model, platform, and application layer rather than one blanket policy. An enterprise evaluating vendors should ask whether each one differentiates governance requirements by where in the stack a given risk actually sits, or applies a single generic policy regardless of deployment context.
- Treat runtime monitoring as a checklist item, not a nice-to-have. Microsoft’s ASSERT and Agent Control Specification exist because pre-deployment testing alone does not catch what an agent does once it is live and interacting with real tools and data. Any enterprise deploying agents in production should be able to answer what its own runtime monitoring actually catches, not only what its pre-launch red-teaming covered.
- ISO 42001 certification is becoming a credible proof point, not just a badge. Microsoft naming specific certified products, rather than claiming certification for the company broadly, is a more verifiable claim a procurement team can actually check against the certifying body’s public registry.
The read this report doesn’t spell out
Microsoft’s restructuring around “risks that evolve as systems interact with their environments” is a direct acknowledgment that governance built for static models does not transfer cleanly to agentic ones, and it lands the same month multiple enterprise surveys have flagged agent governance as the gap holding back scaled deployment. For a buying committee still governing its AI programs with a model-era policy, checked once before deployment and left alone, Microsoft’s own restructuring is a concrete argument for building continuous, runtime-level oversight now rather than waiting for a governance failure to force the issue.
Source: Microsoft On the Issues
