Direct the decision system
Executives may delegate operation, research, and implementation. They remain accountable for purpose, risk appetite, affected people, and the decision to act.
Use boundary: work from a fictional scenario or an approved, redacted decision register. Never place board materials, customer or employee records, credentials, contracts, or non-public financials into an unapproved service. Confirm purpose, rights, lawful basis, retention, access, and provider terms first.
Any production implementation additionally requires approved redacted inputs, rights and lawful basis, data minimization, least privilege, tenant separation, managed secrets, audit logs, representative, adversarial, and regression evaluations, proposal-only behavior, human approval, monitoring, a stop condition, an incident path, and tested rollback.
What leadership must retain
AI can surface patterns and alternatives, but fluent text is not evidence and a model has no mandate to set strategy. Keep five decisions with named people:
Classify the decision before using AI
Low-impact support
Formatting a public-source brief or generating questions can be suitable for a controlled draft workflow. A person still checks sources and releases the output.
Material business decision
Pricing, capital allocation, vendor selection, customer eligibility, or a public claim needs verified data, domain review, documented assumptions, approvals, and a reversible test.
Consequential people or rights decision
Hiring, dismissal, promotion, pay, credit, access, safety, or legal status must not be delegated to a model. Use qualified human review, applicable-law analysis, fairness testing, meaningful explanation, and an appeal path.
A fictional regional retailer wants a weekly “risk brief.” The team first uses synthetic events to test whether the system distinguishes cited facts from inference. It grants read-only access to an approved warehouse view, excludes personal data, and keeps every recommendation in proposal status. The CFO validates figures; the operating owner approves any experiment; a material citation failure stops the workflow.
Exercise: build an accountability map
- List several recurring decision types using titles only, without underlying confidential content.
- For each, name the decision owner, affected groups, evidence owner, and qualified reviewer.
- Mark allowed data classes and explicitly prohibited data.
- Define whether AI may summarize, challenge, or propose. Do not grant execution authority.
- Add an approval gate, monitoring signal, stop condition, incident contact, and rollback owner.
Completion gate
Move on when these are documented
The NIST AI Risk Management Framework provides a voluntary structure for governing, mapping, measuring, and managing AI risk. Apply the law and policies relevant to your organization with qualified advisers.