Microsoft has launched an AI Transparency Initiative with more than 40 organizations, arguing that enterprises need a way to independently verify what AI models and agents do—not merely record what the systems say they did.
Announced at Microsoft Ignite on November 17, the initiative is aimed at a growing enterprise problem. Organizations are connecting AI agents to business data, credentials, applications and infrastructure. Those agents can then take actions at machine speed, sometimes with limited human intervention. Existing monitoring can show activity around a model, but it may not explain the path between a request and an outcome.
Microsoft’s “see, verify and control” framework
Microsoft’s proposal is organized around three requirements:
- See: record what an agent was asked and authorized to do, which model and runtime it used, what data and tools it accessed, and which actions it took.
- Verify: distinguish model-reported signals from evidence collected independently by identity, policy and security systems. An agent should not be the sole source of truth about its own behaviour.
- Control: use evidence about an agent’s purpose, owner, data and delegated authority to enforce policies before consequential actions occur and to reconstruct events during investigations or audits.
The distinction matters because an AI agent is not just another software feature. It can interpret instructions, select tools and pursue a task across several systems. If its permissions are broad, a compromised or poorly constrained agent could behave like a privileged insider while operating continuously and at a scale that manual review cannot match.
Open standards are central to the proposal
Microsoft says it will expand its participation in the National Institute of Standards and Technology’s AI Agent Standards Initiative. NIST describes that program as an effort to encourage industry-led technical standards and open protocols for autonomous agents, including work on agent security, identity and authorization.
The Microsoft initiative is also intended to work across models and providers. That would allow an organization to change the model behind an agent without losing the evidence needed for security operations, compliance reviews or post-incident investigations. The company says participating organizations include Accenture, Adobe, KPMG, Mastercard and Volvo Cars.
Microsoft plans to incorporate the resulting work into Microsoft AI, Microsoft Foundry and Microsoft Copilot. The announcement describes a direction and standards effort; it does not establish that a finished, cross-provider transparency standard is already available.
Why this matters to IT and security teams
For enterprises, the most important idea is portability. AI governance that works only inside one vendor’s platform becomes fragile when organizations use multiple models, agent frameworks and security tools. Evidence that can be checked by an organization’s own systems is more useful than a provider-generated activity summary that cannot be independently tested.
The proposal also reinforces a point raised in Guava Insights’ earlier coverage of Microsoft Project Perception and Guardian Agents: identity, least privilege and containment need to apply to non-human operators as well as employees. That is especially relevant as businesses experiment with agents that can send messages, modify records, retrieve files or invoke administrative tools.
It also connects to the broader security risk described in our report on AI agents being used in espionage operations. The lesson is not that agents are inherently unsafe. It is that organizations need reliable records of authorization and action before they can investigate misuse or demonstrate compliance.
The hard part is implementation
Microsoft’s framework is sensible, but the difficult work is still ahead. Industry participants must agree on what evidence is required, how it is generated, who can attest to it, how long it is retained and how it remains comparable across vendors. Standards also need to avoid creating a false sense of assurance if the underlying logs can be altered or if important model and runtime behaviour remains unobservable.
Businesses deploying agents should not wait for a final industry standard to begin. They can start by maintaining an inventory of agents, assigning accountable owners, separating agent identities from user identities, restricting access to the minimum required tools and data, and retaining logs that connect delegated authority to actions.
Microsoft’s initiative is therefore best viewed as an attempt to move enterprise AI governance from broad promises of “responsible AI” toward evidence that security and compliance teams can inspect. Its credibility will depend on whether the proposed standards become genuinely open, portable and independently verifiable across competing models and platforms.


