Intelligent commercial email processing
A B2B services company was receiving 80 emails per week. Leads were drowning in the noise. Nobody had time to respond fast enough. We put an AI pipeline between the inbox and the team.
Representative scenario based on real engagements. Names, figures and contexts have been adapted for confidentiality reasons.
The problem
A 6-person B2B services SME. No dedicated sales team - requests come into a shared email inbox, and everyone is supposed to monitor it.
In practice: nobody really does. Emails pile up. Urgent leads get buried among newsletters, supplier follow-ups, and support requests. A prospect who asks a question on Tuesday might wait until Thursday for a reply. Some never get a reply at all.
80 incoming emails per week. 20% are potential leads. 60% are noise. The rest is existing client follow-up. All arriving mixed together, without structure.
What we built
An AI pipeline connected directly to the inbox.
Real-time classification - every incoming email is analyzed and categorized: new lead, client follow-up, support, administrative, spam. Leads are surfaced immediately.
Key information extraction - stated need, sector, estimated urgency, contact details. A structured brief ready to read in 10 seconds.
Draft reply generation - for standard requests, a personalized draft is submitted for one-click approval. The team reviews, adjusts if needed, sends. No writing from scratch.
Automatic follow-ups - leads without a reply after 24 hours are flagged. Prospects who didn’t follow up after an initial exchange receive a follow-up at the right moment.
Getting the classification right took 3 weeks of adjustment. Early models regularly confused certain support requests with leads. We built a correction dataset specific to the company’s sector - since then, the false positive rate has dropped below 5%.
Results
- Average response time: 2.5 days → 6 hours
- Lead response rate: ~60% → 87%
- Time spent on email: down roughly 60% for the team
- Several opportunities identified that the team had missed entirely in the 3 months before
Stack
- Python - LLM pipeline for classification and generation
- IMAP connection to the existing inbox
- Draft validation interface (web, mobile)
- No tool migration - plugged into what they already used