Since 2016, Microsoft has dedicated hundreds of engineers, lawyers, and policy experts to establish a foundation for its responsible AI practices.
Building on that foundation, EY teams have rigorously evaluated and tested Microsoft 365 Copilot AI features against ISO 42001 requirements, ultimately resulting in Microsoft 365 Copilot successfully achieving ISO 42001 certification in March 2025. This exercise allowed EY to systematically validate how Microsoft’s responsible AI principles were embedded into the design and day-to-day operations of Microsoft 365 Copilot, demonstrating not just compliance, but resilience and readiness at scale.
EY’s evaluation surfaced key themes that demonstrated Microsoft’s preparedness for ISO 42001 certification offering actionable insights on how to advance their own responsible AI journeys:
To bring responsible AI policy to life, its principles must be translated into clear, actionable guidance that engineering teams can apply in practice. At Microsoft, this translation happened through a structured impact assessment process for AI features supporting Microsoft 365 Copilot. Following Microsoft’s Responsible AI Standard, these assessments prompted teams to anticipate potential risks to stakeholders and define appropriate mitigations. The process is grounded in policy-aligned questions and supported by practical tools such as software development kits for user feedback, safety filters to block harmful content, and secure AI Application Program Interfaces for approved large language models (LLMs) that help teams meet responsible AI requirements. EY teams examined how these assessments were used to embed policy into product development, offering a lens into how Microsoft operationalizes responsible AI in ways that are measurable, repeatable and scalable.
Evaluating harms in the context of the AI features
Understanding how AI systems behave under pressure is essential to building resilient and trustworthy solutions. For Microsoft 365 Copilot, this involved evaluating potential harms, such as the generation of ungrounded or harmful content or vulnerability to jailbreak attempts, within the context of each feature’s intended use. Microsoft conducted simulated harms evaluations to anticipate how these risks could emerge and to design layered defenses that mitigate them. EY teams validated that harms evaluation was embedded into the development lifecycle, ensuring AI features were proactively stress-tested against adversarial risks before release to safeguard users from real-world threats upon go-live.
Implementing safety systems to perform responsible AI monitoring at scale
LLMs can occasionally generate content that poses risks to users or organizations. To proactively address this, Microsoft 365 Copilot incorporated multiple layers of AI safety systems. These include classifiers that detect potentially harmful prompts or outputs and trigger mitigations such as suppressing unsafe responses or redirecting the user. Microsoft also uses metaprompting to shape system behavior in line with responsible AI principles, such as avoiding speculation or emotional inference when summarizing meetings. EY teams assessed how AI safety systems were embedded into the product’s architecture, verifying that Microsoft 365 Copilot’s design included layered safeguards like classifiers and metaprompting confirming that Microsoft’s approach to responsible AI is both intentional and technically robust.
Continuously monitoring AI features in production
Responsible AI doesn’t end at deployment. It demands continuous oversight. Microsoft 365 Copilot is actively monitored in production to validate that it performs reliably and safely, aligned with policy expectations. Engineering teams track a range of metrics, including success rates, uptime, accuracy and indicators of misuse such as jailbreak attempts. These metrics feed into intelligent alerting systems that detect anomalies in real time, enabling rapid response from on-call teams. EY teams examined telemetry pipelines and alerting mechanisms to ensure continuous monitoring was not only in place but actionable, enabling responsiveness to deviations in AI behavior.
Keeping humans at center in the responsible AI equation
Even the most advanced AI systems require human judgment to responsibly guide their development. At Microsoft, this role is championed by designated responsible AI leads embedded within product teams and empowered to oversee risk management throughout the AI lifecycle. These individuals collaborate with engineers to assess how features will be used, anticipate potential risks and apply lessons learned from past deployments. They also serve as a bridge to Microsoft’s Office of Responsible AI, ensuring consistent governance across teams. EY teams confirmed that responsible AI champions were strategically embedded across product teams, empowering them to guide risk-aware development and uphold governance throughout the AI lifecycle. Research conducted in collaboration between Ernst & Young LLP and the Saïd Business School at the University of Oxford, shows that placing humans at the center of major transformations increases the likelihood of success by a factor of 2.6 times compared with transformations that do not prioritize a human-centric approach.
While ISO 42001 certification was a key milestone, it marked the beginning—not the end—of Ernst & Young LLP’s collaboration with Microsoft. Drawing upon the experience of EY teams operationalizing responsible AI at scale and Microsoft’s leadership in the responsible AI arena, Microsoft’s M365 Trusted Platform Team and EY professionals worked closely with Microsoft’s compliance, legal, responsible AI and engineering teams.
Together, they explored strategies to reinforce existing responsible AI practices, future-proof their approach and embed responsible AI into the pace of innovation.
With these insights and feedback, EY teams defined the following guiding principles to sustain and scale responsible AI practices:
- Embedding responsible AI into day-to-day workflows to build operational muscle memory
- Continuously validating the design and effectiveness of controls
- Equipping teams with tools and templates to accelerate adoption
- Capturing real-world feedback from users to inform improvements
- Delivering ongoing training to keep pace with evolving risks and requirements
These principles were translated into tangible initiatives, to make sure responsible AI wasn’t just a policy, but a practice. By investing in these areas, Microsoft strengthened its ability to deploy AI responsibly, adapt to emerging regulations, and build lasting trust with users.
“Operationalizing trust in AI requires more than a checklist—it demands a system of accountability that evolves with both the technology and the people who use it,” said Andrea Craig, Principal in Technology Consulting at Ernst & Young LLP. “Our work with Microsoft focused on building that system—one that’s resilient, scalable, and designed to adapt as AI capabilities and customer expectations continue to evolve.”
