Agentic email uses AI agents with scoped tool access (via protocols like MCP) to read, triage, draft, and send mail under defined policies rather than one-off prompts.
What good looks like#
- Done: Grant the agent read scope first; add draft scope, then send scope only for defined categories.
- Done: Issue per-agent tokens so access can be revoked individually.
- Done: Require approval for messages to new external recipients.
- Done: Log every action with the reasoning the agent produced.
What bad looks like#
- Seen in audits: Granting full send access on day one.
- Seen in audits: No kill switch when the agent misclassifies a thread.
How to move from bad to good#
Work through the good list in order and re-verify after each change. Most teams find one or two items from the bad list already present; fixing those usually produces the largest improvement.
Frequently asked questions#
What is MCP in email?
Model Context Protocol lets AI clients call tools like 'search inbox' or 'draft reply' through a standard interface. MailMaid exposes an MCP server with scoped tokens.