Why I built the scoring model from scratch instead of using GPT-4

Most people building AI tools in 2026 reach for an LLM first. I did as well.

I tested GPT-4 for lead scoring. Accuracy: 61%. That's 11 points above random guessing. Not good enough to build a product around.

The problem is structural. Dead CRM lead data is sparse : irregular touchpoints, incomplete fields, inconsistent hygiene. LLMs fill in missing signal with plausible-sounding confidence. They can't say "I don't know." They're trained to complete.

So I built a reinforcement learning model trained specifically on re-engagement outcome data. Accuracy: 94%.

That gap ? 61% to 94% , is the entire product. Everything else is presentation.

Question for the PH community: Have you ever tried to re-engage dead leads? What was the biggest friction : knowing who to contact, what to say, or just finding the time?