How an email travels through InboxAgent

Understanding the order things happen in explains almost every “why did it do that?” question. Here’s the journey of one email, in the sequence it really runs.

1. Duplicate check

If InboxAgent has already processed this exact message, it stops here — the copy is marked as a duplicate and skipped. Nothing is ever handled twice.

2. Known-conversation check

If the email is a reply on a thread that already has a helpdesk ticket, InboxAgent adds it to that ticket as a comment — instantly, without re-classifying. Same for threads it already knows: no duplicate records. See Connect Zoho Desk and Connect HubSpot.

If that ticket has since been deleted in your helpdesk, the reply isn’t lost: InboxAgent notices the ticket is gone and opens a fresh one carrying the reply.

3. Noise filters (no AI involved)

Cheap, deterministic checks clear the obvious cases first:

• Your own team’s mail — if the “Ignore emails from your own team’s domains” rule toggle is on, internal mail is skipped. Exception: if an internal email is actually a forwarded customer inquiry (say, a colleague forwards a lead to sales@), InboxAgent detects the real prospect inside and processes it as them — and if it can’t tell for sure, it sends the email to review rather than guessing.
• System senders — no-reply@, postmaster@, mailer-daemon@ and friends are always filtered.
• Bulk mail — newsletters and mass marketing (recognized by standard bulk-mail markers and known marketing domains) are skipped.
• Auto-replies — out-of-office and automatic responses are skipped.

Mail filtered here shows as No action needed with 100% confidence, a rule decided rather than a guess.

4. AI classification

Everything that survives the filters is read by the AI, which assigns a category and a confidence score. The AI also knows your sender labels (customer, priority, blocked) and the conversation’s history, which inform — but never dictate — its read.

Where that happens. The classification runs on a third-party AI service, so the message is sent there to be read and classified. It is not used to train anyone’s AI model, and only what the classification needs is sent. The full detail of what leaves InboxAgent, and what we commit to, is in our AI Notice and Privacy Policy.

5. Your rules

Your global rules are checked next, then any per-mailbox overrides. The first matching rule wins outright — it overrides both the AI’s suggestion and your confidence thresholds. No rule matched? The AI’s suggestion stands.

6. Confidence thresholds

Unless a rule already decided, InboxAgent compares the confidence score against your routing mode. Below the bar → Send to review, whatever the category. Above it → the decision proceeds.

7. The outcome

The final decision executes — lead, ticket, review, or ignore — your alerts fire, and the email appears in Inbox activity with its full audit trail.

The safety net

If the AI errors or can’t answer, the email goes to Manual review — never into a guess and never lost. If an action fails partway (CRM down), the email lands on Failed messages for retry. The design principle throughout: when in doubt, a human decides.