An AI email agent that learns from your edits

What the learning loop does

An AI that drafts from instructions alone will always sound a little generic. inboxr closes that gap with your own replies. Every time you approve a draft, it stores an example made of three things: the email that came in, the draft the AI wrote, and the reply you actually sent. When it drafts the next email for that mailbox, it reads the most relevant examples first.

Nothing about this needs setting up. If you are reviewing drafts, you are already teaching it.

Sample data. Names and messages are fictional.

How it works, step by step

1. You approve a reply. It can be the draft as written or your edited version.
2. inboxr compares the two. If you changed the draft, it records that the reply was edited and writes itself a short summary of what changed, for example "shortened the opening and removed the apology".
3. The example is saved. It is stored against the mailbox, tagged with the email's topic, along with any note you left.
4. The next draft reads it. When a new email arrives, inboxr pulls a small set of recent examples for that mailbox, taking ones with the same topic first, and gives them to the AI alongside your knowledge base and agent instructions.

Because examples are matched by topic, a correction you made on a billing reply shapes the next billing reply before it shapes anything else.

Two kinds of signal

Clean approvals

When you send a draft without touching it, you are telling inboxr that the draft was right. That reply becomes a positive example, and it adds one to the mailbox's count toward unlocking autopilot.

Edits and notes

When you change a draft, the difference is the lesson. The optional note to the AI lets you say why in your own words. Keep notes short and specific. "Always confirm the order number before answering" is more useful than "be more careful".

What it does not learn from

• Rejected drafts. A rejection means no reply was needed, so there is nothing to imitate.
• Your signature. The signature is added when a reply is sent and is kept out of the examples, so the AI never starts writing it into drafts.
• Other mailboxes or other customers. Examples belong to the mailbox they came from and are only used there.

Personal details in examples, such as email addresses and phone numbers, are masked before an example is shown to the AI as a reference for a different email.

How to tell it is working

The analytics page shows an "edited before send" rate: how often a reviewed draft needed changes. If that number falls over time, drafts are getting closer to what you would have written. It is also the clearest sign a mailbox is getting ready for autopilot.

Learning is not the same as knowledge

The learning loop teaches style and judgment: how long your replies are, how you open, what you offer next. It is not where facts should live. If a draft got a fact wrong, fix the fact in your knowledge base and regenerate, so every agent and every future draft has it right. If the tone was wrong for a whole category of email, adjust the agent's role or tone.

Who it is for

Anyone whose replies have a recognizable voice: a founder who answers their own email, a support team with a house style, a studio that writes to clients in a particular way. The more consistently you review, the faster the drafts stop needing edits. For the bigger picture, read the guide to AI email agents.