What are the AI governance best practices that actually hold up in practice?

Every other practice on this list assumes you know which AI systems exist and how risky each one is. A discovery process that only checks IT-approved procurement will miss shadow AI adopted through personal accounts, vendor-embedded features, and department-level subscriptions, and every practice built on top of an incomplete inventory inherits that same blind spot.

Organizations that skip this step end up applying governance unevenly: heavy scrutiny on the systems everyone already knows about, and none on the tools adopted quietly by individual teams. Risk classification then determines how much of the rest of this list applies to a given system — a low-risk internal tool does not need the same oversight burden as a system making consequential decisions about people.

The best practice here is not "have a human-in-the-loop step," it is "have a human-in-the-loop step that is technically enforced." A documented review checkpoint that a bug or misconfiguration can skip is not a control, and real incidents confirm this repeatedly: a content-moderation system whose review step could be bypassed let thousands of wrongful bans through before anyone caught it. Oversight has to be built into the system architecture, not just the process documentation.

Scale the intensity of oversight to the risk tier of the system. Low-risk internal tools do not need the same review burden as a system making consequential decisions about people, and applying uniform oversight everywhere either creates unnecessary friction or, more often, gets quietly ignored for the low-risk cases and then never re-tightened when a system's risk profile changes.

Automated drift detection, anomaly alerting, and tamper-evident audit logs are baseline practice for any system beyond the lowest risk tier. The test of whether this is actually implemented well is simple: could you produce a complete, credible record of a specific AI system's decisions over the last six months on short notice? If the honest answer involves manually reconstructing logs from multiple systems, monitoring and logging are not yet at a defensible maturity level.

Most organizations do not train their own foundation models, they build on top of a vendor's. Best practice treats that vendor relationship as a governance surface in its own right: what was the underlying model trained on, what license or consent covers that data, and what containment does the vendor provide if the system misbehaves. Assessing this after an incident, rather than before deployment, is the pattern behind a large share of the AI governance failures that make the news.

Traditional model governance assumes a system produces an output a human acts on. Agentic AI that takes autonomous action needs distinct baseline practices: documented permission boundaries, credential scoping specific to each agent, and a kill switch that can halt autonomous action without waiting for a full incident response cycle to spin up. Treating agentic systems as a variant of existing model governance, rather than as a distinct control domain, is currently one of the most common gaps we see across otherwise mature programs.