Summary:
More than 40 organizations are working with Microsoft on AI transparency. The announcement sets a direction, not a finished audit format or implementation deadline.Microsoft announced the AI Transparency Initiative on November 17 with more than 40 organizations, including Adobe, Accenture, KPMG, Mastercard and Volvo Cars. Its central problem is practical: an agent's explanation of its actions is not enough evidence to establish what actually happened.
A standards effort, not a finished audit format
Microsoft plans to contribute to NIST's SI Agent Standards Initiative, develop a common approach with industry partners, and incorporate the work into Microsoft AI, Foundry and Copilot. The intended scope spans providers and both closed and open-weight models.
The announcement sets out goals, but provides no final technical specification, common log schema or delivery timetable. Organizations can register interest; that is not the same as downloading an interoperable audit implementation.
Separate reports from observations
Microsoft's proposed evidence should distinguish what comes from the model, its runtime and independent identity or policy systems. The aim is to let customers check actions using their own security tools and retain governance when switching models or vendors.
That distinction matters in a simple example. Suppose an agent says it sent an invoice only after approval. Three records answer different questions:
- The agent's report: what it believed it was doing.
- The authorization record: which recipient, amount and action a person or policy actually approved.
- The external system's receipt: whether that invoice was sent, when, and to whom.
If those records disagree, a fluent summary should not settle the dispute. A useful investigation starts with the mismatch and preserves the original evidence, rather than asking the same agent to rewrite its explanation.
Do not confuse the initiative with products already shipping
Microsoft's separate Ignite security overview describes Perception as generally available, while some agent governance controls remain in limited preview. It explicitly distinguishes those containment capabilities from visibility into the behavior of the underlying model. A product's availability label therefore does not establish that the new transparency standards have shipped.
Questions to ask before granting more authority
For a pilot, ask who owns each agent, how delegated authority expires, which records come from outside the model, and whether evidence can be exported without losing identifiers or timestamps. Test a failed action as well as a successful one: can your team tell a rejected request from an executed action followed by a misleading report?
Keep the scope concrete. Start with one workflow and the evidence needed to reconstruct it, rather than treating a new initiative as a blanket assurance. Our guide to embedded safety assessments examines a related distinction between access, evaluation and a completed verdict.
Sources and image
Microsoft's November 17 announcement and security overview, checked November 18, 2026. Featured image: original PromptsZone conceptual evidence-flow diagram, not a released specification or product interface. The invoice example and pilot questions are editorial analysis.


