AI Governance, Consent & Approval Gates
AI proposes, judges validate, policies authorize, executors act and humans supervise — the seven gates every action passes before a real channel is touched.
Five stages, every single action.
This is the part that makes autonomy defensible. An agent’s proposal is validated by something other than itself, authorized by policy rather than by persuasion, and executed only by a component whose entire job is execution — with a person in the loop wherever judgement is genuinely required.
AI proposes and reasons
Agents discover markets, accounts and contacts, build the strategy, draft the outreach and work the conversation — and they show the reasoning that got them there rather than emitting a result.
Judges validate
A separate judging pass checks the proposal before anything acts on it. The model that wrote the thing is not the model that approves it.
Policies authorize
Consent, suppression, jurisdiction, channel and budget policy decide whether the validated action is permitted at all. A policy refusal is not overridable by a better-sounding argument.
Executors act
Only then does an executor touch a real channel — and only if the provider is configured for it. Everything it does becomes a persistent task, approval, event or audit row.
Humans supervise
Sensitive decisions escalate to a person by design. Autonomy levels are set explicitly, per surface, and the approval trail is the record of who allowed what.
What has to pass before a real channel is touched.
| Gate | Surface | What it enforces |
|---|---|---|
| Consent | Consent Ledger | No channel action proceeds without recorded consent of the right kind, with the evidence behind it. |
| Suppression | Suppression Center | Global and per-channel suppression is applied before a send is attempted, not reconciled afterwards. |
| Preference | Preference Center | A recipient’s stated preferences are honoured across every channel simultaneously. |
| Policy | Policy engine | Jurisdiction, channel, budget and content policy authorize — or refuse — the validated action. |
| Judge | Judging pass | A separate validation pass reviews the proposal before any executor sees it. |
| Approval | Approval Command Center | Sensitive decisions escalate to a person, and the approval trail records who allowed what. |
| Provider | Providers | Real email, voice, paid media and webhooks only execute where a provider is actually configured for that channel. |
Execution is gated, deliberately. Real email, voice, paid-media and webhook actions require a configured provider plus passage through every gate above; where a channel is not configured, the surface still works and renders governed simulation data so the programme can be designed and reviewed before anything can reach a real person.
Set how much the system may do without asking.
Autonomy is explicit and per-surface, set in the Autonomy Control Center. Raising it never removes the gates — it changes which decisions escalate to a person and which the platform is permitted to take on its own, and the approval trail records the difference either way. The watchdog reports whether the agents and providers are actually healthy, with alerting behind it.
