It begins with a plain inventory: which AI systems do you build or use, what role do you play for each, and which teams and sites are involved? That inventory fixes the scope.
Three assessments then carry the weight. An AI risk assessment looks at what could stop you meeting your objectives. A risk treatment chooses controls, compares them with the standard's reference controls and records what applies and why. An impact assessment asks what each system could mean for the people it touches.
Policies, roles, competence, data and supplier checks, monitoring, internal audit and management review complete the system before the audit by an accredited certification body.
The inventory is often the surprise. Teams discover chatbots, scoring tools, recommendation engines and document-reading software that nobody had listed. Once they appear on one page, you can decide which deserve close oversight and which need only light control. Proportion matters: the aim is responsible use that your teams can sustain, not a system that stops sensible experiments.