How to build a lead scoring model with LLMs (that actually beats your rules engine)

Every CRM sells 'lead scoring.' Almost nobody uses it because rule-based scoring is garbage. Here's the hybrid LLM + rules approach that actually works.

Why rule-based scoring fails

HubSpot lets you build scoring rules like '+10 for VP title, +5 for company size >100, -20 for Gmail address.' These rules are easy to write and sound reasonable. They're also almost useless in practice.

Here's why: the signals that actually predict purchase intent are buried in the content of what leads write, not the fields they fill in. 'I'm evaluating 3 vendors and we need to decide by end of month' is a HOT lead even if the title is missing. 'Just researching' from a VP of a Fortune 500 is a COLD lead. Rules can't read that difference. LLMs can.

The hybrid architecture

The prompt that does the heavy lifting

You are scoring inbound leads for [COMPANY]. Analyze the
lead data below and return JSON with three scores:

1. INTENT (0-100): how actively are they looking to buy?
2. FIT (0-100): how well do they match our ICP?
3. URGENCY (0-100): how time-sensitive is their need?

ICP EXAMPLES:
- Best customers: [3 examples of your ideal customers]
- Bad-fit customers: [2 examples of customers who churned]

LEAD DATA:
{lead_object}

BUYING SIGNALS TO LOOK FOR:
- Specific timelines ("need by Q2")
- Budget mentions ("we have $X allocated")
- Competitor evaluation ("comparing you to X")
- Pain language ("our current tool is broken")

OUTPUT JSON:
{
  "intent": 0-100,
  "fit": 0-100,
  "urgency": 0-100,
  "reasoning": "1-sentence explanation",
  "buying_signals": ["list of quotes from the lead"],
  "red_flags": ["list of concerns"],
  "recommended_action": "route_to_sales|nurture_sequence|disqualify"
}

Two things make this prompt work in production: the ICP examples (few-shot learning ties the model to your specific customer patterns) and the quote extraction (forces the model to cite actual lead text instead of hallucinating reasoning).

Validation is the step everyone skips

A lead scoring model is only as good as the validation data behind it. Here's the 3-step validation we run for every deployment:

FAQ