---
title: A personal AI email assistant that is not just a draft button
metaTitle: Building a personal AI email assistant with its own address
description: Most AI email assistants sit inside your mailbox and draft. An assistant with its own address can be delegated to instead: cc it, forward to it, and let it handle the thread.
date: 2026-07-29
author: MCPmailer
tags: Use cases
---

The AI email assistant most people have tried lives inside their mailbox and offers to write the reply they were already writing. It saves keystrokes and nothing else, because the work was never the typing. The work is the twenty conversations you have not opened, the scheduling ping-pong, and the follow-up you owe someone from Tuesday.

![An assistant with its own address handling a thread it was copied into](/blog/personal-ai-email-assistant/hero.webp "Copy it in, and the conversation becomes its problem rather than another draft you approve.")

An assistant with an address of its own can be given that work instead of watching you do it.

## Delegation beats autocomplete

The difference is architectural, not cosmetic. An assistant inside your mailbox can only act when you open something. An assistant with its own address, say `ada-assist@agents.yourcompany.com`, is a participant: it can be copied in, forwarded to, and written to directly by other people.

That unlocks three patterns you cannot get from a draft button.

**Cc to delegate.** You copy the assistant on a thread with a line like "coordinating a time" and drop out. It takes the scheduling ping-pong from there, and the other person is talking to it rather than waiting on you.

**Forward to dispatch.** An invoice, a form to fill, a question that needs an answer from your documentation. Forward it and let the assistant do the extraction, the lookup, and the reply.

**Direct address for the routine.** Publish the assistant's address for the categories that never needed you: booking requests, status questions, document collection.

In all three the assistant is doing the conversation, not preparing one for you.

## What it needs to be useful rather than annoying

Two tool calls before every reply, which is the same discipline described in [the system prompt for an email agent](/blog/email-agent-system-prompt): read the whole thread with `get_thread`, and look the person up with `lookup_contact` so it knows what previous conversations established. An assistant that asks your accountant to re-explain who they are is worse than no assistant.

Then memory afterwards. One line to the contact record per meaningful exchange is what makes the fifth conversation with someone noticeably better than the first, and it is covered in [giving an email agent memory](/blog/agent-memory-contacts-email).

## Draw the line before you need it

A personal assistant sits close to things that matter, so decide the boundary while nothing is on fire.

| It can | It must not |
| --- | --- |
| Propose and confirm times | Accept or decline anything with money in it |
| Answer from documents you gave it | Speak for you on anything sensitive or personal |
| Collect documents and details | Share what is in one thread with another correspondent |
| Chase a reply once | Contact people who have not written to you |
| Draft and send routine replies | Send to your manager, your board, or your family |

That last row is the one people forget and then regret. A relationship allowlist, or simply whitelist mode on the identity per [identities](/docs/identities), keeps the assistant among the counterparties you actually meant.

> [!note]
> An assistant that says it is an assistant does better than one impersonating you. People adjust their expectations instantly and stop being annoyed by the small stiffness in its replies.

## Handing back

The assistant should hand a thread to you the moment it is out of its depth, and handing back should be as cheap as delegating. Because everything happens in real threads, that is just the assistant replying to you with the conversation attached, or leaving the thread for you to take over from the dashboard.

The signals worth wiring: anything the assistant cannot ground in something it read, anything emotional, anything a person explicitly asks you about, and any second attempt that did not land. Same triggers as a support agent, smaller stakes, described in [how to build an agent that answers support email](/blog/ai-agent-customer-support-email).

![Delegating by cc, dispatching by forward, and handing a thread back](/blog/personal-ai-email-assistant/delegate.webp "Three ways in, one way back. All of them are ordinary email.")

## Set it up in an afternoon

1. Create an identity with a name that reads as an assistant, not as a second you. [Quickstart](/docs/quickstart).
2. Put it in whitelist mode until you trust it, so only people you allow can reach it.
3. Give it a short prompt: who it works for, what it may decide, what it must escalate, and how to sign off.
4. Forward one category to it for a week. Scheduling is the usual first choice because the failure mode is a wrong time rather than a wrong commitment.
5. Read every thread that week, then widen or narrow the boundary based on what you saw.

The parts that make it work unattended, waiting for replies without a scheduler and reacting to inbound in seconds, are in [long-running email conversations](/blog/wait-for-reply-long-running-agents) and [webhooks or polling](/blog/webhooks-vs-polling-agent-email).

## Why not just use the assistant built into your mail client

Use both. The in-client assistant is good at what it does: summarising, searching, drafting while you are looking at the thread. It cannot be delegated to, because it has no identity of its own, no address anyone can write to, and no way to act while you are not there.

The distinction is exactly the one in [agent inbox or shared mailbox](/blog/agent-inbox-vs-shared-mailbox): an assistant acting *as* you belongs in your mailbox, and an assistant acting *for* you needs an address of its own.

## Questions

### What is a personal AI email assistant?

An agent with its own address that can be copied into threads, forwarded messages, or written to directly, and that handles conversations rather than drafting them for you to send.

### How is that different from AI in my mail client?

An in-client assistant only acts when you open something and always sends as you. An assistant with its own identity can be delegated to and can act while you are elsewhere.

### Will people know they are writing to an assistant?

They should. Give it a name that reads as an assistant and let it say so. People adjust immediately, and the alternative is someone discovering it mid-thread.

### How do I stop it emailing the wrong people?

Run the identity in whitelist mode while you build trust, so only allowed addresses can reach it, and keep first contact with strangers out of its job description entirely.

### What should it never handle?

Anything with money, anything personal or sensitive, and anyone you would not want receiving a slightly stiff reply. Those threads should route back to you with the context attached.

### Can it work with my existing inbox?

Yes, by forwarding categories to it. Nothing about your own mailbox changes, and you can stop forwarding at any point.

## Related

- [How to give an AI agent its own email address](/blog/email-for-ai-agents)
- [Giving an email agent memory](/blog/agent-memory-contacts-email)
- [Email for scheduling agents](/for/scheduling-agents)
- [The system prompt for an email agent](/blog/email-agent-system-prompt)