Responsible AI in Client Engagements

Design AI use cases with clear ownership, testing, human oversight and failure handling.

Responsible AI begins with the business process, not a policy document.

Ask before implementation

• What decision or action will the system influence?
• Who could be harmed if it is wrong?
• What accuracy level is acceptable?
• Who reviews or overrides the output?
• What happens when the model is uncertain?
• What inputs can the system access?
• What tools or downstream actions can it trigger?
• How will performance be monitored after launch?

Human oversight

Human approval is especially important where:

• the output affects employment, credit, health, safety or rights;
• errors are costly or difficult to reverse;
• model confidence cannot be reliably calibrated;
• the organisation is still learning the workflow.

Do not add a meaningless “human in the loop.” Define what the human checks, when they intervene and what information they receive.

Testing

Test against representative real-world cases, including:

• common cases;
• edge cases;
• ambiguous inputs;
• adversarial or malformed input;
• failure of external tools or data sources.

Traceability

Record:

• intended use;
• owner;
• model/provider;
• data sources;
• key risks;
• testing outcome;
• approval status;
• monitoring plan.

Avoid automation bias

Users may over-trust confident AI output. Training and interface design should reinforce when independent review is required.