Skip to main content

​
Email content and signatures

Not on purpose. Before Magic AI categorizes an email-sourced ticket, Thread runs the message through a cleanup pass that strips common signature blocks: sign-offs (Regards,, Thanks,, Best,, Sent from my iPhone, etc.), quoted-reply markers (-----Original Message-----, From: / Sent: / To: / Cc: header blocks), title lines (CEO, Manager, Director), phone/fax/mobile lines, social handles, and bare URL lines. The cleaned version is what Auto-Categorization and Auto-Prioritization see, not the raw email.Cleanup is heuristic, though — it recognizes signature shapes, not the concept of a signature. A stray line in a signature block that doesn’t match any of those shapes (for example, a company or product name on its own line, like QuickBooks Services) can survive into the cleaned text. If your board has an Item that matches that word, Magic AI can end up choosing it even though the request has nothing to do with it.If you see this happen, the fix is on the categorization side: sharpen the Item’s definition so unrelated mentions don’t match. See Configure Auto Categorization and prompt tips.
Two ingredients usually combine on these:
  1. The spam message was short enough that most of the meaningful text left after signature cleanup was a keyword the sender’s signature happened to contain (a vendor name, a product name).
  2. Your board has an Item whose definition matches that keyword loosely.
Magic AI has to pick from the options you’ve defined, so once the signal survives cleanup and matches an Item, that Item wins. Two ways to reduce the chance next time:
  • Tighten the Item’s definition. Add a short prompt explaining what a real request for that Item looks like, and what shouldn’t count. See prompt tips for the “let it say I don’t know” pattern.
  • Test in the emulator. In Admin Panel → Assistive AI → Categorization: Edit, paste the offending email into the emulator on the right and iterate on the definition until Magic AI stops picking it.
If a specific partner keeps hitting this, share the ticket ID with support so we can look at the exact cleaned text Magic AI saw.
For an email-sourced ticket, it’s the ticket summary plus the cleaned first message from the contact. Quoted replies below the sender’s message, forwarded headers, and recognized signature blocks are removed before Magic AI reads it. Internal notes and technician replies are not used to categorize.

​
Corrections and accuracy

Change the Type / Subtype / Item on the ticket the same way you would any other field, either in Thread Inbox or in your PSA. Overrides are tracked on the Assistive AI Accuracy Dashboard so you can see which boards and Items get corrected the most.
In Admin Panel → Assistive AI → Categorization: Edit, check the Ignore box next to any Type, Subtype, or Item you don’t want Magic AI to assign. See Configure Auto Categorization.
Whenever Auto-Categorization runs on a ticket, Thread appends an internal note explaining the choice. Open the ticket in Inbox and look at the internal notes for the explanation, then use it to adjust your Type / Item definitions.