Human accountability. Defined operating boundaries.

Document the intended use, permitted actions, prohibited actions and likely failure modes. Limit tool access and autonomy to the approved task. Identify the business owner and the people authorised to approve changes.

Require a designated reviewer for actions with material financial, legal, safety or operational impact. Show the supporting information and proposed action before approval. Keep a practical way to pause an agent or return the workflow to manual operation.

Classify data by purpose and sensitivity, grant access by role and minimise what is shared with models and service providers. Agree hosting, retention, transfer and deletion arrangements for the engagement. Use customer data for model training only with explicit written authorisation.

Define a representative evaluation set and acceptance criteria before release. Test accuracy, robustness, security, bias where relevant and foreseeable failure modes. Specify logs for model versions, input sources, significant outputs, tool actions and human decisions, with controlled access and retention.

Agree an incident contact and escalation path, pause criteria, rollback procedures and responsibilities before launch. Investigate failures, document corrective actions and retest before restoring affected functions. Qualified professionals retain responsibility in regulated or high-risk settings.

Users should know when they are interacting with AI, what it can do and when to seek human review. Singapore’s agentic AI guidance provides a reference for bounded autonomy, accountability, lifecycle controls and user understanding. Reference to it does not imply government endorsement or certification.