Translate principles into use-specific commitments
For each use case, identify who benefits, who may be disadvantaged and which outcomes are unacceptable. Define practical commitments: information users receive, decisions AI must not make alone, data that must not be entered and situations requiring specialist review. Give each commitment an accountable owner.
ISO/IEC 42001 provides a framework for responsible AI development, provision and use. It does not remove the need to interpret legal, contractual and sector requirements. Record those obligations separately and obtain appropriate advice where their application is uncertain.
Make human oversight meaningful
Specify what reviewers must check, what information they can access and when they must refuse or escalate an output. Allow sufficient time and authority to intervene. Reviewers should understand relevant limitations, including plausible but incorrect answers and performance differences across the populations encountered.
Test the workflow with realistic failures. Can a reviewer spot an unsupported recommendation, override it and record the reason without unreasonable friction? Monitor rubber-stamping and excessive reliance. A signature is not strong evidence of oversight when the person lacks the means to question the system.
Provide transparency and routes to challenge
Decide what affected people need to know about AI involvement, its purpose and the responsible organisation. Explain limitations in language suited to the audience. Establish a route for questions, correction and challenge, with clear ownership where an outcome may have significant consequences.
Use impact assessments and feedback to examine unequal or unexpected effects. Where evaluation depends on sensitive information, involve privacy and legal specialists before collecting it. Record evidence limitations openly; vendor-level fairness claims may not represent your users, language, accessibility needs or context.
Review whether commitments work in practice
Bring complaints, incidents, overrides and evaluation findings into governance review. Reconsider safeguards when the model, task or affected population changes. If evidence shows unacceptable outcomes, be prepared to restrict, redesign or suspend use rather than preserve approval because implementation was expensive.
Connect these reviews to AI risk management and the existing ISO 42001 service. Horizon organises risks, controls, owners, evidence and reviews. It does not determine fairness, decide acceptable harm or certify that an AI use is responsible.