The identity your programme cannot see

Every control in a mature access programme assumes a person: a manager who attests, a start date, a leaving date, a disciplinary process. Service accounts, API keys and integrations arrive without any of that. They are created by whoever needed them, on the day they needed them, and they persist.

The joiner-mover-leaver process is the spine of access governance. It works because employment gives it structure: a contract begins, a role changes, a resignation is tendered. Each event triggers a review of access. For the non-human population there is no joiner, no mover and no leaver. There is only creation.

The consequence is not theoretical. When a person exits, the credentials they stood up remain in place — carrying their authority and none of their accountability. The register that would tell you this exists, in most institutions, as a spreadsheet maintained by the team that last needed it.

Test this against any service account in your estate. Ask what it is for, who owns it and when it was last reviewed. In most institutions, two of the three questions have no answer. The third is usually answered by whoever happens to be in the room.

The usual response is a discovery tool. Discovery has its place: it finds what is reachable. But a tool cannot tell you why a credential exists, what it was meant to do, or who should answer for it. Discovery is the beginning of the work, not the work.

Ownership is the harder problem. A service account belongs to a system, a team, a vendor, or a person who left two reorganisations ago. Until one named person answers for each identity, review is theatre: the same list is circulated, the same blanks are signed off, and nothing is ever removed because nobody can say what breaks.

Lifecycle follows ownership. A credential with an owner can have a purpose, a review date and an end. A credential without one has none of these, and no tool will supply them.

This is the ground on which agentic systems are now being deployed. The argument of this article is that the sequencing matters: an institution that cannot enumerate its service accounts cannot govern an AI agent, because an agent is only a service account that decides.

An agent inherits the estate it joins. Standing privileges, shared credentials and unreviewed access do not become safer because the actor is now autonomous. They are simply exercised faster, at machine speed, with the same absent accountability.

The remedy is not new. Enumerate the population. Assign an owner to every identity. Set a lifecycle and hold to it. Review on a schedule that means something. This is access governance applied to a population that was never employed — which is to say, applied where it was always needed and never done.

Institutions that do this work now will find agent adoption straightforward, because the questions agents raise are the questions they have already answered. Institutions that do not will govern their agents the way they govern their service accounts today: by assumption.