How InboxAgent learns from you

Every time you correct InboxAgent, that correction is recorded and put to work. Here’s exactly what “learning” means — and what it doesn’t.

Where corrections come from

• Re-classify on an email in Inbox activity — “this was actually a lead”. These count with extra weight, because you actively fixed the record.
• Resolving items in Manual review — every Lead / Ticket / Ignore decision you make is feedback.
• Confirming on the Learning screen — an admin marks recorded corrections as verified, which doubles their weight.

What the corrections do

1. They accumulate into suggested rules. When the same correction keeps happening — “InboxAgent says Marketing, you say Sales” for a certain kind of mail — InboxAgent generates a suggested rule capturing the pattern. Confirmed corrections get there twice as fast.
2. Sender history informs future classifications. InboxAgent builds a per-sender track record (how often their mail turns out to be sales or support). That history — along with your sender labels — is context the AI sees the next time that sender writes in.
3. Proven senders can skip the review stop. With the optional fast-path enabled, high-confidence mail from senders with a strong track record goes straight to lead/ticket instead of pausing for review.

What learning does not do

• Suggestions don’t apply themselves. A suggested rule changes nothing until an admin reviews and acts on it — you stay in control of every standing behavior change.
• The AI model is not retrained on your mail. Learning works through the mechanisms above — suggested rules, sender track records, and prompt context — not by modifying the underlying AI.
• Your mail stays in your workspace. Your emails, your corrections, your rules and your sender labels are visible only to your workspace. No other organization can see them, and no other organization’s own rules or labels are ever applied to your mail.

What InboxAgent learns across organizations

Alongside the per-workspace learning above, InboxAgent keeps one small platform-wide layer so that common, generic mail gets classified better for everybody. Being straight about what that is:

• Nothing is recorded until at least three separate organizations show the same signal. A subject word, a sender domain or a category correction seen in only one or two workspaces is never written down at all. That threshold is what keeps the layer generic: a term that turns up across three unrelated businesses is a commodity word like “renewal” or “invoice”, not your customer’s name.
• We don’t keep the words or domains themselves. Each one is turned into a fixed code using a secret key that is stored separately from the database. Someone who obtained a copy of the database on its own could not turn those codes back into the original text. We hold the key, which is what lets the system recognize when the same value appears again, and we do not attempt to reverse the codes or use them to identify anyone. We don’t keep whole subject lines, message bodies, attachments, email addresses or contact names in this layer.
• What it actually does for you. When a sender domain has been seen by three or more organizations, the layer can pass a general hint into the AI’s context, for example “mail from this domain is usually support”. Our team also reviews the recorded patterns. Only category-correction patterns, which are just a pair of InboxAgent’s own category names and contain no text from anyone’s email, can be turned into a rule.
• What it never carries. It never moves your rules, your sender labels, your categories, your contacts or any text from your email into another workspace.

If your workspace is deleted, the corrections and sender history that fed this layer are deleted with it, and any pattern that then falls below the three-organization threshold is removed.

Getting the most from it

• Correct promptly and consistently — ten consistent corrections beat fifty contradictory ones.
• Have an admin visit the Learning screen occasionally and Confirm the corrections that were right — it’s the accelerator pedal.
• For senders you already know matter, skip the learning curve entirely: a sender label plus a rule is instant and deterministic.