DPO Radio

The hardest design decision in an AI compliance integration is the write gate. An open write surface — where a trusted external key can mutate production records autonomously — is the kind of architecture that loses inspection-readiness. A no-write surface — where external integrations can only read — limits adoption to research-only use cases.
The approval queue is the third option. External Claude clients can propose writes, but every proposal lands in the queue, every proposal is previewed by a human administrator, and every approve, decline, or expiry is recorded in the audit trail. The integration gets write-capable adoption; the compliance team keeps the human gate; the auditor gets a complete lifecycle trail.
The queue's load-bearing constraint is the same constraint the rest of Forseti enforces: there is no bypass. There is no "trusted key" mode. There is no "auto-approve low-risk writes" toggle. There is no "skip the queue for internal integrations". Every external write is human-gated.
A write tool call enqueues a row with the tool name, the arguments, the proposing key prefix, the proposal preview, the rationale, and the expiry timestamp (24 hours from enqueue). The calling client receives a pending-write identifier and uses it to poll for status.
An administrator opens the pending writes surface under Forseti. The page lists every active proposal with the tool name, the credential and proposing client, the proposed change preview, and the time-to-expiry. The administrator clicks into a proposal to see the full arguments and the rendered preview.
The administrator's three options. Approve runs the tool with the requester's current resolved permissions at the moment of approval, not at the moment of proposal. Decline closes the row without executing; the calling integration receives a "declined" response on its next poll. Expire is the default if no action is taken — an hourly sweep moves expired rows to a terminal "expired" state.
Every proposal emits five lifecycle events as it moves through the queue: proposed, approved, executed, declined, expired. All five are stitched together by a shared proposal identifier so the audit trail can walk the lifecycle end to end.
When the administrator approves, the platform resolves the API-key holder's permissions at that moment. If the key holder's permissions have changed since proposal — for example, the administrator who issued the key has been offboarded and lost write permissions — the approval fails with a clear reason. The proposing integration sees the failure in its poll.
No Forseti token allowance stands between a proposal and the queue, because there is none. A proposal lands in the queue however much the organisation used that month, and the administrator decides on it whenever they review it. The queue never becomes a tripwire that locks the team out during a busy period.



Five lifecycle events per proposal, stitched together by a shared proposal identifier, flow through the same immutable audit trail as the rest of the platform. The chain of custody is unbroken from the external-client call to the executed (or declined or expired) outcome.

Permission resolution happens at approve time, not at proposal time. An offboarded administrator's key cannot execute a proposal even if the proposal was created while they still held permissions. The audit event records the resolved-permission state.

The 24-hour expiry is enforced by an hourly sweep. Expired proposals do not execute. The expiry is itself an audit event, so an administrator who ignores a queue for two days cannot later argue that the proposal "could have executed" — the expired state is in the trail.

Open the proposal and inspect the current record effect before acting. The MCP audit view lets a reviewer follow the proposal and the resulting decision.
Ready to see an external write proposal land in the queue, get previewed, and execute on operator approve? Request a personalised demo.


Issue a write-capable MCP key and run the queue against a Claude Code DPIA-drafting integration. Measure approval cadence against your review-team tolerance.

Talk to our team about how to scope queue ownership and approval responsibilities across your administrators.
No. The approval queue is the load-bearing human gate for every external write proposal. There is no "trusted key" mode, no "auto-approve low-risk writes" toggle, no "skip the queue for internal integrations" exception. Every external write is human-gated by construction.
The hourly sweep moves the proposal to an expired terminal state and emits the "write expired" audit event. The calling integration receives a "expired" status on its next poll. Clients can re-propose by calling the tool again; there is no auto-revival.
The API-key holder's permissions, resolved at the moment of approval. If the key holder's permissions have changed since proposal — for example, the administrator who issued the key has been offboarded — the approval fails with a clear reason. The proposing integration sees the failure in its poll.
Yes. The approval queue shows line-by-line preview for bulk proposals with per-row approve / decline. The administrator can approve, decline, or selectively edit the list before approve.
Five linked events: proposed, approved, executed, declined, or expired. For an approved-and-executed proposal the trail carries proposed → approved → executed, all stitched together by a shared proposal identifier. The Audit Trail page can filter on the MCP source to isolate external traffic.