Fairness in AI is a context-dependent engineering and governance objective that can be specified, measured, and evidenced for a stated use case, not an aspiration to be asserted. This guide turns that principle into a practical operating model for the systems organizations deploy today: predictive models, generative foundation models, and AI agents with varying degrees of autonomy. It treats harmful bias as an important, measurable pathway to unfairness, addressed as close as practicable to its causal source across data, labels, models, context, and workflow rather than corrected only after the fact, and it builds on the framework of AI Fairness: from Principles to Practice [1]. Two features of the current landscape shape the model. First, many organizations now consume foundation-model capability rather than train it, so deployer leverage shifts from training data to model selection, inference context, and the surrounding value chain. Second, systems increasingly act, not just answer, which places fairness risk in tool use, resource allocation, and inter-agent interaction at runtime and creates distinct feedback risks, including recursive synthetic-data effects, output homogenization, and inter-agent propagation. The guide provides a step-by-step workflow, a rule for choosing evaluation bundles, a runtime escalation ladder, a worked example, and a mapping from evidence to the EU AI Act, the NIST AI Risk Management Framework and its Generative AI Profile, and ISO/IEC 42001. It standardizes the evaluation process and required evidence while leaving operational tolerance an explicit, bounded, owned, and documented decision. Its aim is not to prove a system fair, but to make fairness claims explicit, testable, and auditable.