Junk filtering

A business inbox is full of mail that isn't a customer: sales pitches, newsletters, receipts, system notifications, out-of-office replies, spam. Before spending effort on extraction and scoring, Leadveyor decides whether each message is a customer enquiry at all.

What arrives (example, 100 messages)
  • Customer enquiries68
  • Ads and sales pitches11
  • Newsletters8
  • Receipts, notifications7
  • Auto-replies3
  • Spam3
Sender domain rules
Code
Always-junk domains stop here, for free. Never-junk domains skip the filter.
Fast model classifies
AI
lead · advertisement · spam · newsletter · transactional · auto-reply · regular mail, with a confidence.
Confident “not a lead”
Set aside as junk / not a lead. Visible in the Inbox, one click to rescue.
Lead, or unsure
Goes on to extraction. Doubt always counts as a lead.
Left: an illustrative business inbox — a third of it isn't enquiries. Right: how a message is judged.

The bar on the left is an illustrative inbox: about two thirds customer enquiries, the rest ads and sales pitches, newsletters, receipts and notifications, auto-replies and spam. On the right, a message passes sender domain rules, then a fast model's classification; only a confident "not a lead" is set aside, while leads and unsure cases go on to extraction.

Step 1: domain rules

Your profile can list sender domains:

  • Always junk — messages from these domains are set aside immediately, without any model call. Use it for persistent senders: an SEO agency, a supplier's marketing list.
  • Never junk — messages from these domains skip the filter entirely. Use it for lead platforms and partners whose notifications look automated but contain real enquiries.

Step 2: the classifier

A fast, inexpensive language model reads the message with your business description and sorts it into one of seven categories:

CategoryResultExamples
leadLeadA request for a quote, a booking, a question about your services, a form submission, an enquiry forwarded by staff
advertisementJunkSomeone selling to you: SEO, marketing, software, lead lists
spamJunkScams, phishing, gibberish
newsletterNot a leadNewsletters, campaigns, social notifications
transactionalNot a leadReceipts, invoices, shipping, password resets, system notifications
autoresponderNot a leadOut-of-office and automatic replies
regularNot a leadReal mail that isn't an enquiry: suppliers, staff, job applications

Each verdict comes with a confidence (high, medium, low) and a one-line reason, which you see on the message in the Inbox.

The confidence rule

Only a confident "not a lead" stops a message. If the classifier says junk with low confidence, or says lead, the message goes on to extraction. Short and vague enquiries ("how much for 50 people?") count as leads. The reasoning is simple: reading one more ad costs a fraction of a cent; losing a customer costs a job.

Extraction is a second safety net: if the main model finds no customer enquiry in a message that passed the filter, the message is marked "not a lead" with the reason no customer enquiry found.

Your own rules

Businesses have their own kinds of junk. Add extra rules in plain words to the profile — "Messages from suppliers about deliveries are not leads", "Applications for jobs are not leads" — and the classifier follows them. Your business description in workspace settings matters too: "a quote for 80 people" is an enquiry for a caterer and probably not for a law firm.

Nothing is deleted

Messages set aside are kept, with their category and reason, in the Inbox. Not junk — process again sends a message back: the filter is skipped and the message is extracted and scored as a lead. Switch the filter off entirely in the profile when every message you send is known to be an enquiry (for example a form that only customers can reach).

Was this page helpful?