The named reviewer already exists

Every governance framework published this year asks the same question: who is accountable for the AI agent?

The answers arrive in a familiar shape. Inventories. Identity registries. Delegated-authority documents. Escalation maps. Chain-of-authority diagrams with four tiers of owners, approvers, and stewards. The advice is sound, and almost none of it describes a working system.

The reason is simpler than the frameworks suggest. Most organisations are trying to build an accountability layer from scratch, on top of a tool that was deployed without one. They are drawing the org chart after the work has already started.

There is a different sequence available. It starts with the reviewer, not the register.

The accountability gap is a staffing problem

Read the guidance closely and a pattern emerges. The framework asks for a named business owner, an operational owner, an exception approver, and a system steward. Four roles, each with documented authority, each auditable.

Then it asks who should fill them, and the room goes quiet.

The quiet is understandable. These are senior judgement roles. The exception approver needs enough domain depth to overrule a system that is right most of the time. The operational owner needs enough context to know when routine output has drifted. Hiring for that depth takes months, and the people who have it are already expensive.

So the roles get assigned to whoever is nearest. The agent's accountability chain ends in a team alias, a shared inbox, or a manager three functions away who reviews a dashboard once a week. The framework is technically satisfied. The accountability is not real.

This is the gap the frameworks describe honestly and solve poorly: accountability runs to a named human, and named humans with the right judgement are the scarcest resource in the building.

The reviewer was never the bottleneck

Now consider the sequence in reverse.

In an attended model, the senior expert is already there. Before the agent's first output reaches a client, a consequential decision has already been routed to a credentialed human who approved it, corrected it, or sent it back. That human is named. Their credentials are verified. Their decisions are logged, because the correction is what makes the system better.

The chain of authority the frameworks describe is not something to be designed and then staffed. It is already operating. The expert who reviews the work is the accountable owner of the work. There is no separate question about who owns the agent, because the agent does not act on consequential matters without the reviewer's sign-off.

// accountability is not a diagram. it is a person who has already said no today.

This reframes most of the governance checklist. The inventory exists because every specialist is deployed against a named role. The audit trail exists because every correction is captured as training signal. The approval boundary exists because consequential work is routed, not flagged after the fact. The revocation path exists because attendance can be withdrawn as quickly as it was granted.

None of this requires a new hire. It requires recognising that the oversight layer and the accountability layer are the same layer.

Named is doing more work than accountable

The word that matters in "named human" is not accountable. It is named.

Accountability assigned to a function is accountability deferred. A risk committee cannot be paged when an agent is about to send a non-compliant filing. A department cannot hold the judgement that a specific clause is wrong for this jurisdiction. Only a person can, and only a person whose name, credentials, and decisions are on record.

Verification is what makes the name meaningful. An expert whose identity and credentials are attested is not a profile picture and a job title. They are a verifiable principal in the chain. When the framework asks "who approved this action," the answer is a person whose qualifications to approve it can be checked, not a role that could be filled by anyone with access.

This is the difference between oversight as compliance theatre and oversight as an operating property. The first produces documentation that describes control. The second produces a record of control being exercised, decision by decision, by a human whose name is attached.

The framework converges on the product

The interesting development this year is not that governance guidance exists. It is that the guidance, followed to its conclusion, describes attended deployment almost exactly.

Named accountable owner. Explicit, bounded authority. Human approval before irreversible actions. Runtime records that reconstruct what happened and who approved it. Fast revocation. Corrections that feed improvement.

An organisation can build this from internal hires and committees. It will take quarters, and the hardest roles to fill are the ones that require the most senior judgement.

Or it can start from the other end: specialists that arrive attended, with the named reviewer already in the loop, already credentialed, already accountable for the output that leaves the building.

The question the frameworks ask — who owns the agent — has an answer that is older than the frameworks. The person who reviews the work owns the work. That was true before the work was done by software. It is still true now.

// every agent, attended.