AI Lead Scoring: Prompts, Pipelines, and the Pitfalls Nobody Warns You About

Rule-based lead scoring is brittle. You get one weight wrong (say, "title contains Director = +10") and you ship junk to sales for a month before catching it. LLMs handle the fuzzy reasoning humans use intuitively — "this lead's job title is weird but the company profile and email signature pattern screams enterprise budget."

The catch: LLMs also confidently score garbage when you set them up wrong. Here's the path that works.

The minimal pipeline

New lead → enrich → score with Claude → assign tier → push to CRM

The score prompt is the part that matters. Everything else is plumbing.

The score prompt that doesn't suck

You are a B2B sales-development analyst scoring inbound leads for [company]
that sells [product] to [ICP].

Given the lead data below, return JSON with:
- score: 1-100
- tier: "hot" | "warm" | "cool" | "junk"
- reasoning: 2-3 sentences
- top_risk: the one thing that would make this lead not convert

Lead data:
[enriched profile]

Scoring rubric:
- Title fit (0-25): does this person have authority + budget for our price point?
- Company fit (0-25): is the company the right size, stage, industry?
- Intent signal (0-25): what evidence do we have they're actively shopping?
- Engagement quality (0-25): what is the quality of how they reached us?

Be honest in `top_risk`. A great-looking lead with a tell that they're a tire-kicker should score below an okay lead with strong intent.

Pitfall #1: The model loves your leads too much

Default LLM behavior is to find reasons everything is good. After two weeks, you'll notice 70% of your leads are scoring "warm" or higher. That's calibration drift.

Fix: ground the prompt with examples. Append:

Reference cases (these score from your past data):
- [paste 3 past hot leads that closed]
- [paste 3 past warm leads that needed nurture]
- [paste 3 past junk leads that wasted SDR time]

The few-shot examples pull the model's calibration toward your actual reality.

Pitfall #2: Hallucinated signals

The model will sometimes invent things it "saw" in the data. "Their funding round suggests they're scaling" — but no funding data was in the input. This is hallucination dressed as inference.

Fix: require the model to quote the source data for each scoring dimension. If it can't quote, it can't score that dimension.

Pitfall #3: The model gets worse over time (silently)

Three months in, score quality drifts. Could be the model version updated. Could be your audience changed. Could be that your ideal-customer definition changed and nobody updated the prompt.

Fix: weekly drift check. Sample 20 random leads from the past week, have a human re-score them, compare to the LLM's score. If average gap > 10 points, retrain — meaning: update the few-shot examples in the prompt.

What ships better than this

Honestly, nothing if your lead volume is under 500/month. The pipeline costs about 8 cents per lead in API calls and 30 minutes to set up. If it cuts your SDR's time on bad leads by 30%, the math works for any team doing more than a hundred leads a week.