AI still does useful work.
Drafting, summarizing, comparing evidence, and proposing a safer action can happen without granting permission to execute.
Customer follow-up proof · no sign-in · about 60 seconds
Review a customer message before AI sends it.
See a customer update move from AI-prepared draft to evidence check, human decision, one approved action, and a reviewable result.
Powered by ARBITER Email Gate
Nothing is sentNo credentials or customer dataHuman decision stays explicitEvery outcome leaves a receipt
- 01DraftAI prepares
- 02CheckEvidence holds
- 03DecideHuman controls
- 04ReceiptOutcome proven
Inspect the ARBITER control recordRequest → policy → grant → synthetic action → receipt → expiry
The experience above explains the customer decision first. The technical record below shows how ARBITER limits the action to one message, one recipient, one approval, and one short-lived permission.
An AI agent is ready to send an email.
{
"actor": "agent:v1ra",
"action": "send_email",
"target": "customer@example.com",
"reason": "send approved outage update",
"riskLevel": "medium"
}The agent can prepare the message, but the Gate owns permission to send it.
Authority remains explicit.
Policy identifies what needs review. A person can revise, approve once, or deny without handing over permanent access.
The result stays reviewable.
The request, detected risk, human decision, final action, and closure of authority remain visible after the moment passes.