Practical guidance for safety, risk and governance professionals responsible for AI deployed in occupational, health, safety and environmental contexts.
Design human oversight to actually intervene
Meaningful oversight requires that the overseeing professional has the training to understand the system's behaviour and limitations, real-time visibility into what the system is doing, and the practical authority and time to intervene before harm occurs. A "human in the loop" who cannot realistically override a decision in time is oversight in name only — and increasingly, regulators are testing for the difference.
Treat AI incidents like any other safety incident
An AI-related near-miss or failure should trigger the same investigation discipline as any other safety event: preserve logs and model version data before they roll over, establish a timeline, identify contributing factors (data, model, process, human), and close the loop with corrective action. Waiting for a formal AI-specific incident process to be built from scratch is how evidence gets lost.
Separate "does it run" from "is it still right"
Uptime monitoring tells you a system is running. It says nothing about whether its outputs are still accurate as real-world conditions drift from training conditions. Build in independent accuracy checks, not just availability checks.
Make competence, not just process, auditable
A documented governance process is necessary but not sufficient — auditors and regulators increasingly want evidence that the individuals running it are actually competent. Independently verified credentials such as AISP, AIRP and AIIP give organisations a defensible answer to "how do you know your people know what they're doing?"
For the standards these practices are grounded in, see AI Safety Standards; for regulatory context, see Regulations & Guidelines.