Ungoverned AI replies
AI email replies need authority, not just better prompts.
A model can draft a reply. The harder question is whether it should send, from which address, under whose authority, and with what record. Gent puts delegated email replies inside scopes, approvals, and audit history.
The problem
A send button is a business action.
AI email can affect reputation, commitments, access, money, privacy, and client relationships. If reply authority only lives in a prompt or a shared mailbox credential, it is hard to explain where the boundary was.
Wrong authority
The reply appears to come from a person or team account even when software prepared it.
No approval path
Sensitive messages rely on ad-hoc review instead of an explicit held action.
Overbroad access
The agent may inherit mailbox permissions that are broader than the work requires.
Weak audit
After the reply goes out, it is hard to prove what was drafted, approved, rejected, or sent.
What changes
The agent can prepare work without owning final authority.
Separate
The agent works from its own inbox rather than a human sent folder.
Scope
Tokens define what the agent may read, send, label, delete, update, or create.
Hold
High-risk sends and destructive actions can be queued for approval.
Review
A human sees the intended action and approves or rejects it before execution.
Audit
The held action, reviewer, outcome, and message record remain attached to the inbox.
Why Gent
Governance belongs in the email layer.
Not blind autonomy
The agent's authority is defined by token scopes and approval rules.
Not a personal inbox
The agent does not need a person's mailbox to receive, draft, and act.
Not only logs
Approvals happen before the risky action, not only after it has already occurred.
API-first
Products and agents can integrate directly with messages, tokens, approvals, and audit events.
Next step
Give the delegated work its own address.
Start with one workflow that should no longer borrow a human inbox: vendor intake, client follow-up, support escalation, or agent-managed approvals.