AI agent governance is the set of organisational decisions that determine who may design, share, deploy and operate AI-driven workflows, what information those workflows may access, and who is accountable when they fail. It should be designed before a pilot agent is connected to sensitive business systems.
## Who should be allowed to create AI agents? Separate experimentation from production deployment. Permit bounded experiments with approved data; require named ownership, a business purpose, review and testing before sharing agents widely. Creation permissions and deployment permissions need not be identical. Confirm the precise administrative controls supported by your subscription and connected products.
## How should data and tool permissions work? Use least-privilege permissions, scoped identities and a documented connector register. Evaluate whether the tool acts with the user's permissions or a service account. Record the data classification, retention assumptions, cross-system transfers and how access is revoked. Never assume an agent inherits all the controls of the underlying ChatGPT workspace.
## What approvals should agents require? Define risk tiers. Drafting a private summary can require user review; publishing externally, changing customer records, initiating payments or sending communications should require explicit authorisation and independent controls. Include rollback, incident escalation and an emergency disable mechanism.
## Who owns monitoring and accountability? Each deployed agent needs a named business owner, technical operator and review frequency. Monitor errors, unexpected tool calls, policy exceptions, cost and quality. Retain proportionate audit evidence and regularly review access after staff departures or role changes.
## How does this fit a full rollout? Agent permissions are one component of the ChatGPT Business implementation playbook. See the IT governance playbook, company-tool connections and AI Maturity Assessment. AI Build Academy's intended specialist training pathway covers agent creation and testing, creator-versus-deployer permissions, connector access, human approvals, monitoring, audit records and incident response. Detailed configuration exercises, control templates and role-based practice belong in those lessons rather than this public overview. Course availability and enrolment links must be verified against the Academy catalogue before being published; technical controls depend on the actual plan and connected tools.