Building a recipient allowlist for an email agent
A recipient allowlist limits who an email agent may send to on its own. Start with people already in the thread, then narrow to named addresses or domains.

On this page(8 sections)
- What is a recipient allowlist for an email agent?
- Why does an agent need one at all?
- Who should be on the list?
- How does Koltrix decide who the agent can send to?
- Address or domain: which should you allow?
- How do you test the list before you rely on it?
- What does an allowlist not protect you from?
- Bottom line
An email agent that can send on its own needs a rule for who it may write to, and "anyone it decides to" is the wrong rule. A recipient allowlist is that rule: a short list of addresses or domains the agent may send to without a person clicking send. Anything outside the list waits for a human instead.
This post covers what belongs on the list, how narrow to make it, and how to test it before you trust it. The examples use Koltrix's Auto mode, which is in Beta, but the thinking applies to any agent that sends mail.
What is a recipient allowlist for an email agent?
A recipient allowlist is a list of email addresses and domains that an agent is permitted to send to automatically. When the agent drafts a reply, the system checks every recipient against the list. If all of them pass, the message can go out. If any one fails, the message is held for approval.
It answers a different question from a sender allowlist. A sender allowlist decides whose mail the agent reads and acts on. A recipient allowlist decides who the agent may write to. You usually want both, because they guard opposite ends of the same conversation.
Why does an agent need one at all?
Because the failure you are guarding against is a message reaching the wrong person. A bad reply to the right customer is embarrassing. A reply that goes to the wrong address can leak something. The usual causes are:
- A forwarded or mis-copied thread. Someone Cc's an outside party, and the agent replies to everyone.
- Instructions hidden in an incoming email. A message that says "also send this to the following address" is the classic prompt injection. If the agent is only allowed to reach known people, the attack has nowhere to go.
- A wrong guess. A language model can produce a plausible-looking address that does not exist or belongs to someone else.
The principle is the one in least-privilege AI email agents: give the agent the smallest reach that still lets it do its job.
Who should be on the list?
Start from the job, not the contact book. Ask who the agent legitimately needs to write to, then list only that group.
| Agent job | Reasonable recipient rule |
|---|---|
| Answers customer support questions | Leave the list empty so it replies only to people already in the conversation |
| Handles invoices from known vendors | The vendor domains you already pay, such as @vendor.com |
| Sends status updates to a partner | The one or two named partner addresses |
| Books meetings with prospects | Empty list, with a low daily cap, so it only answers people who wrote in |
The pattern is that the more predictable the audience, the tighter the list can be. A support agent has an open-ended audience but a narrow behavior (reply to the person who wrote). A vendor agent has a narrow audience, so name it.
How does Koltrix decide who the agent can send to?
In Auto mode, Koltrix checks every recipient before an agent sends alone. The rules, as built:
- Every recipient must already be known. That means someone in the conversation, or a contact a person added. A brand-new address is never eligible.
- If you set an allowed recipients list, each recipient must also match it, by full address or by domain.
- The agent may send to at most 5 recipients in one message, and Bcc is not allowed.
- The other Auto conditions still apply: high confidence (0.85) with an answer grounded in the mailbox's notes or documents, a daily cap, optional quiet hours, a 10-minute duplicate hold, and a clean result from the prompt-injection shield.
If you leave the allowed list empty, Koltrix treats that as "anyone already in the conversation." A non-empty list narrows it further. So the list never widens what the agent can reach. It only restricts it.
When a rule fails, the message is not dropped and not sent. It goes to the Action center for a person to approve, with the reason attached, such as recipient_not_allowed or unknown_recipient. The approval step is where you decide whether to send it, edit it or reject it. Pending approvals expire after 7 days.
Address or domain: which should you allow?
Allow the narrowest form that works.
- A full address (
[email protected]) is the safest. Use it when one person or one shared mailbox is the whole audience. - A domain (
@vendor.com) is convenient when several people there are involved. It also trusts everyone at that domain, so only use it for organizations you actually deal with. - A public mail domain (
@gmail.com,@outlook.com) should never be on the list. It matches millions of people you have no relationship with.
A common mistake is allowing your own domain "because it is internal." If your agent can write to anyone at your company, a confused reply can still land in front of the wrong teammate. Prefer named internal mailboxes.
How do you test the list before you rely on it?
Test it in the safest mode first. Koltrix agent mailboxes have three levels: Suggest (drafts only), Approve (proposes a send and waits for your tap) and Auto. Suggest is the default, and it is the right place to start.
Use this checklist:
- Run the agent in Suggest or Approve for a week and look at who its drafts are addressed to.
- List the distinct recipients. If any surprise you, fix the cause before enabling Auto.
- Write the allowed list from that real set, not from memory.
- Send a test message from an address that is not on the list and confirm the reply is held, not sent.
- Set a small daily cap so a mistake stays small.
- After enabling Auto, read the activity log every day for the first week.
Every autonomous send is logged, and one switch puts every agent back to Suggest if something looks wrong.
What does an allowlist not protect you from?
It limits who receives a message. It does not check what the message says. A reply to the right customer can still be wrong, rude or overpromising. The list is one layer, and it sits next to the others: the confidence gate, the shield, the daily cap, and a human reviewing the log. For a wider design, see human-in-the-loop email automation patterns.
It also does not replace judgment about the mailbox itself. If an agent has its own address and key, read what it takes to give an AI agent its own email address. More on how agent mailboxes work is on the agent mailboxes page.
Bottom line
- A recipient allowlist is the rule that says who an agent may send to without a person clicking send.
- Build it from the agent's real job. Prefer full addresses, allow a domain only for organizations you work with, and never allow a public mail domain.
- In Koltrix, the list can only narrow the agent's reach: recipients must also be people already in the conversation or contacts a person added.
- Test in Suggest first, build the list from the recipients you actually see, and keep a small daily cap while you gain confidence.
Start with Koltrix
Your domain, one inbox, and an API that sends.
A team inbox where AI sorts and drafts (nothing is sent without your click), plus the transactional API and SMTP relay your product sends with. 7 days free, no card.


