AI governance isn't a policy document: on Azure it's a set of concrete technical controls that get turned on via configuration, get audited, and can be tested on every deployment.
Microsoft structures its AI approach around four operational phases: map (identify use-case-specific risks before building), measure (evaluate the system with quantitative safety and quality metrics), mitigate (apply technical controls like content filters and reinforced system prompts), and operate (monitor in production with alerting and incident logging). This framework isn't just conceptual: each phase has specific tooling inside AI Foundry, it doesn't stay on slides.
Content Safety analyzes text and images across four categories (hate, violence, sexual content, self-harm) with four configurable severity levels per category. It's invoked as a standalone service or comes automatically integrated into every call to Azure OpenAI Service. It also includes detection of protected material (known copyrighted code and text) and a moderation service for end-user-generated content before it ever reaches the model.
Prompt Shields is a specialized classifier that detects two distinct attack types: direct jailbreaks (the user tries to manipulate the model into ignoring its system instructions) and indirect prompt injection (malicious instructions hidden inside a document, email, or webpage that the agent processes as context, without the user having written them). This second vector is especially relevant for agents with web search or file search tools, where retrieved content is a real attack surface.
The groundedness detection service compares the generated response against the provided source material (the documents retrieved in a RAG, for example) and flags claims not supported by that source, even offering a corrected version of the response. It's the most direct tool for measuring hallucinations in production without relying on constant manual review, and it integrates as an evaluator inside AI Foundry.
Every Azure AI service publishes a Transparency Note: documentation specific to intended capabilities, known limitations, and responsible-use considerations for that particular service. These complement the model catalog's model cards, which document performance benchmarks, evaluated biases, and use cases not recommended. These documents are direct input for the risk analysis required by frameworks like the EU AI Act.
Azure's AI services inherit the platform's certifications: SOC 1/2/3, ISO/IEC 27001 and 27701, HIPAA with a Business Associate Agreement available, FedRAMP High in Azure Government, and PCI DSS. For the EU AI Act, Microsoft publishes guides mapping its existing controls (Content Safety, evaluation, model documentation) against the obligations for high-risk AI systems, though final legal responsibility for classification and compliance still rests with the deployer of the system, not the platform provider.
On the operational side, Azure Policy lets you block, at the subscription level, the creation of AI resources outside approved regions or without Private Link configured, while the Responsible AI dashboard within Azure Machine Learning runs fairness and explainability analysis on models -- particularly relevant for scoring or automated-decision cases where a regulator may demand justification for why the system decided something.
Carlos Montiel is an enterprise AI solutions architect. He implements LLMs, Agents, RAG, and orchestrators for companies across Guatemala and Latin America. Reach out for a consultation.
Contact Carlos Montiel