The short version
- Connected mailboxes are polled every 15 minutes for inbound mail.
- Replies are matched to the outreach that prompted them using the RFC 5322 In-Reply-To and References headers, not the sender's address.
- Each reply is classified into one of six kinds: positive, neutral, negative, auto-reply, unsubscribe, or out-of-office.
- Classification drives consequences. An unsubscribe suppresses the contact everywhere; a real reply stops the sequence; an out-of-office defers the follow-up instead of wasting it.
- A mailbox whose authorization has expired is marked for reauthorization rather than failing silently.
The hard part of reply handling is not reading the reply. It is knowing that a message is a reply at all, and to which of your emails.
The intuitive approach — match inbound mail by sender address against your recipient list — fails constantly in exactly the situations that matter most.
Why sender matching fails
- People reply from aliases. You wrote to
sarah@publisher.comand the answer comes froms.jones@publisher.com, or from a personal address entirely. - Threads get forwarded. The author passes your email to their editor, who replies. That is a success, and address matching records it as a stranger emailing you.
- Shared inboxes rewrite the sender. Mail to an editorial inbox comes back from whoever picked it up, or from the desk's generic address.
- Names collide. Two people at large publishers share a first name, and a fuzzy match attributes the reply to the wrong thread.
Every one of those is a normal outcome of outreach. A matching strategy that mishandles them is a matching strategy that fails precisely when someone was interested.
Matching on the header chain
Dice matches replies using the standard mail headers instead: In-Reply-To and References. Every message you send carries a unique message ID, and any conforming mail client that replies to it includes that ID in the reply's headers — regardless of who sends the reply, from which address, after however many forwards.
This is what mail clients themselves use to build threads. Adopting it means an alias reply, a forwarded reply, and a shared-inbox reply all attach to the right outreach message, with no fuzzy matching and no guessing.
The header chain is how email itself decides what belongs to a thread. Anything else is a heuristic reimplementing it badly.
Six classifications, six consequences
Once a reply is matched, it is classified — and the classification is not a label for a dashboard. Each one triggers something specific:
| Classification | What happens |
|---|---|
| Positive | The sequence stops and the reply is surfaced to you. Nothing automated goes near a live conversation. |
| Neutral | The sequence stops. A question or a “not right now” is a human response and deserves a human answer. |
| Negative | The sequence stops. Following up after a no is how a polite decline becomes a spam complaint. |
| Unsubscribe | The contact is suppressed across the entire workspace, immediately and permanently, and removed from reachability scoring. |
| Out of office | The follow-up is deferred rather than consumed. An autoresponder is not a reply, and burning a follow-up on one wastes the message. |
| Auto reply | Recorded, and the sequence continues. A ticket-received acknowledgement is not a conversation. |
The out-of-office case is the one that most tools get wrong, and it is worth dwelling on. If an autoresponder counts as a reply, the sequence stops and the person never hears from you again after they return. If it counts as silence, the follow-up fires while they are still away and lands in a pile of two hundred unread messages. Deferring is the only handling that respects what actually happened.
Alongside the classification, each reply gets a one-sentence summary of what was said, a flag for whether a follow-up is warranted, and — where a response makes sense — a suggested reply you can edit or ignore.
Polling, and why 15 minutes
Connected mailboxes are polled on a fifteen-minute schedule. A newly connected mailbox is also synced on demand, so its history is available immediately rather than after the next tick.
Fifteen minutes is a deliberate middle. The only decision that depends on speed is whether a follow-up fires, and follow-ups are days apart — a quarter hour of latency cannot cause a wrong one. Polling every minute would multiply provider API calls for no behavioural difference at all.
When a mailbox stops working
OAuth tokens get revoked. Passwords change. Accounts get suspended. When a sync fails with an authorization error, the mailbox is marked as needing reauthorization and you are notified.
The alternative — retrying quietly and hoping — produces the worst failure in the product: campaigns that appear to be running while every message silently fails, and replies that arrive and are never seen. A broken mailbox should be loud.
Editorial outreach
2 approved · your connected mailbox
Approved
Send
Follow up
Next message
A useful addition for your AI tools roundup
Hi Sarah,
Your comparison gives readers a practical view of the category. I noticed one relevant option is missing and pulled together the facts you would need to assess it…
Sarah Jones
growthlab.example
Marcus Lee
saasframe.example
A positive reply is usually the start of a link, not the end of the work. Once the placement goes live, backlink monitoring takes over and keeps checking that it stays there.