SailPoint chief executive Mark McClain has framed identity security as the discipline that determines whether AI delivers advantage or exposure, arguing that when software can act thousands of times faster than a human, the controls around what it is permitted to do become the constraint that matters. For identity governance practitioners, the statement is less a prediction than a description of work already overdue.
The problem: velocity breaks human-paced governance
Enterprise identity governance was built on the assumption that access changes are infrequent, human-initiated and mutually distinguishable. A person joins, a manager approves, an entitlement is provisioned. Reviews happen on a calendar. Anomalies are visible because they stand out against a slow baseline.
Autonomous agents invert every one of those assumptions. An agent may request credentials, act on data and delegate to another agent many times per minute. The volume of identity events outpaces any review campaign, and the concept of a periodic recertification becomes meaningless when the entitlement set changes hourly. Worse, the agent operates on behalf of a human or a business process, so the blast radius of an over-permissioned agent is indistinguishable from the blast radius of the person it represents, only faster and harder to attribute.
What identity governance has to become
Runtime authorisation, not just provisioning. Governance has to move from deciding access at grant time to constraining it at request time. That means policy evaluated on every action, scoped to the specific task, with the ability to deny in flight. Static entitlements granted for a role are too coarse for an actor that can attempt thousands of operations in a minute.
First-class non-human identities. Agents, service accounts, workloads and automation pipelines need owners, lifecycle rules, credential rotation and review just as employees do. In most estates they exist in greater number and with weaker oversight. Bringing them into the governance inventory is the precondition for any credible AI control story.
Attribution and auditability. When an agent acts, the audit record must show which agent, on whose behalf, under which policy and with what approval. That is an identity governance problem, not a logging problem. Without it, incident response cannot answer the first question a regulator or board will ask.
Governed delegation chains. Agents calling agents create delegation paths that no existing access review models. Governance must represent those chains explicitly, bound their depth and expire them automatically, rather than treating each hop as an independent service account.
The practical implication
McClain’s framing points at a sequencing decision. Organisations that treat AI adoption as an application rollout and govern it afterwards will discover that their identity platform cannot observe, bound or evidence agent behaviour. Those that extend identity governance first, with inventory, ownership, runtime policy and audit trails in place, can let agents move quickly precisely because the boundaries are defined. Speed without that foundation is not acceleration. It is unmanaged privilege at machine tempo.