Before AI agents act: govern them like enterprise integrations
The practical path from AI pilot to production is not more autonomy. It is a clear control plane for identity, access, testing, observability, ownership, and retirement.
The last seven days did not produce a sufficiently consequential Workday product announcement to justify another news-led note. A more useful leadership question is already in front of most enterprises: what must be true before an AI agent is allowed to act on Workday data or business processes? The answer looks familiar to integration leaders. Every production agent needs an identity, a bounded purpose, controlled access, test evidence, operational monitoring, and an accountable owner.
Treat the agent as a production integration
An AI agent may reason differently from a traditional interface, but it still crosses trust boundaries, reads sensitive data, invokes services, and can change enterprise records. The same delivery disciplines therefore remain essential: least-privilege access, explicit data contracts, environment promotion, exception handling, audit evidence, support ownership, and a controlled path to disable the capability.
Workday is building the control plane
Workday's generally available Agent System of Record provides a unified view of Workday-built and third-party agents, records interactions, applies identity and permissioning, and supports a lifecycle from registration through deactivation. Workday's announced Agent Passport adds a complementary assurance layer: pre-production testing, signed attestations against public standards, continuous monitoring, and the ability to allow, block, route, limit, or revoke agent actions. Agent Passport is planned for early access in the second half of 2026, with general availability projected before year-end.
A practical production-readiness gate
Before launch, leaders should require six answers: who owns the agent; which users, data, and actions it can access; which decisions remain deterministic or require human approval; how its behavior was tested; what telemetry and service levels will be monitored; and how it will be restricted or retired. This converts responsible AI from a policy document into a repeatable delivery gate.
The AMS opportunity starts after go-live
Agent operations create a recurring managed-services agenda. Teams must review usage and value, investigate unexpected outcomes, retest after model or prompt changes, tune permissions, maintain knowledge sources, manage consumption, and retire agents that no longer perform or justify their cost. The support model should connect AI governance, Workday security, integrations, functional ownership, and enterprise risk rather than place the agent in an isolated innovation queue.
Why this matters in the GCC
The regional implication is an inference from these capabilities and the NIST AI Risk Management Framework: multi-entity GCC organizations can scale innovation faster when every agent has consistent global controls plus clear local ownership for data, policy, language, and regulatory context. A sensible starting point is one bounded workflow with reversible actions and visible human review—not a broad licence for autonomous access across HR and finance.
- Give every production agent a named business owner, technical owner, identity, purpose, and retirement path.
- Separate AI reasoning from deterministic rules, approvals, and controls that must remain authoritative.
- Fund ongoing agent operations—monitoring, retesting, permission reviews, value measurement, and incident response—not only the initial build.