“Who owns this AI?” sounds simple until the system uses a third-party model, runs through a shared platform and acts inside another team’s workflow. The useful answer is a small accountability model, not a single name in a register.
Separate three kinds of ownership
Business accountability
The business owner accepts the intended use, benefit and residual risk. This role decides whether the system should operate and whether its outcomes remain acceptable.
Technical operation
The technical owner maintains deployment, integrations, identities, monitoring and change. For agents, this includes the tools and credentials through which the agent acts.
Control assurance
Risk, security, privacy, legal and compliance functions define or test controls. They should challenge the first line without becoming the de facto owner of every AI system.
Every material decision should have one accountable role, named contributors and evidence that the decision was made.
Make accountability follow the lifecycle
Ownership should be tested at intake, approval, deployment, material change, incident and retirement. A stale owner field is not accountability. Useful platforms trigger reviews when a model, purpose, data source, permission or supplier changes.
Questions for suppliers
- Can roles vary by business unit, region and risk tier?
- Can the system escalate unowned or overdue records?
- Does it preserve who approved what, when and on which evidence?
- Can technical changes trigger a new business or control review?
- Can responsibility be delegated without losing the accountable owner?
The goal is not more workflow. It is a decision record in which authority, responsibility and evidence remain connected.