1. Write the authority envelope
List what the agent may read, draft, send, write and approve. Give it least privilege: the minimum access needed for the task, on its own identity, not a shared login.
Executive answer
How do I govern AI agents that can send emails or change records?
Give each agent a written authority envelope, keep drafts separate from sends and writes, require a human approval tied to the exact message or change before anything leaves, check that the information behind the action is current and sourced, and keep a receipt for every action, including the ones the system blocked.
2. Separate the action states
A recommendation, a draft, a send, a record write and an activation are five different states. Each one gets its own gate, so a good draft never turns into an unapproved send.
3. Approve the exact action
Outbound email and record changes that are hard to undo need approval tied to that exact message or change, from a named person, before it commits.
4. Check what the agent knew
Before the action commits, verify the information behind it is current, traceable to its source and strong enough for the stakes. That check is Context-Verified Decision Intelligence (CVDI), and it runs outside the model.
5. Keep receipts for everything
Record every send, write, block and no-change outcome with who approved it and what the agent used. The blocked list is how you prove the controls work.
Where to start
List every agent or automation that can send or write without a click from a person. Mark which of the five controls exist. The AI Trust Audit starts with the riskiest gap.
Questions
Is a prompt instruction enough to stop an agent from sending?
No. A prompt is a request to the model. The control has to live in the workflow: the agent's permissions, the send gate and the approval step, which hold even when the model ignores its instructions.
What is the minimum control for outbound email?
Approval each time, by a named person, tied to the exact message. WWAIOS™ treats sends, posts and activations as gated actions that wait for that approval.
What should the record of an agent action include?
What the agent did or proposed, the information it used and how current it was, who approved or blocked it, and the outcome. WWAIOS™ calls this record a receipt.
Follow the source path
Related public sources
AI governance for agents Agent Trust Review Six-question agent checklist What is WWAIOS™ CVDI definition AI Trust Audit
