Skip to content
Bryan Silva
02clay-ai-scoringAll case studies

GTM Engineering · Clay · LLM guardrails

Email prospecting with Clay, AI scoring and a guardrail

Every lead enriched in Clay gets a score, a tier and, when it's worth it, a personalized email checked against facts.

Status
Scoring in production · sending in pilot
Role
End-to-end design and development
Stack
  • n8n
  • Clay
  • OpenAI
  • Supabase
  • Gmail API
  • Webhooks

Results

lead quality
+45%
tiers (A–D)
4
emails/day
500+

Problem

Deciding by hand which leads deserve a personalized approach doesn't scale and ends in generic emails.

Solution

Leads arrive from Clay by webhook; public signals (active ads, website technologies, traffic) become verified facts; a hybrid rules + LLM score sets tiers A–D, and the best ones get an AI-written email, validated by a guardrail and sent at a human pace.

How it works

  1. 01Trigger

    Clay webhook

    Each enriched lead arrives in real time, without relying on the CSV export.

  2. 02Process

    Signals → facts

    Active ads, website technologies and traffic become a closed list of facts.

  3. 03Process

    Rule-based score

    Opportunity 0–20, digital maturity 0–12 and reachability 0–8: base 0–40.

  4. 04AI

    LLM analysis

    Normalizes to 0–100 and adjusts by ±25 at most, citing only facts from the list.

  5. 05AI

    AI copy

    Tiers A and B get a personalized email via structured output.

  6. 06Process

    Guardrail

    Blocks placeholders, amounts, unsupported figures and claims about unknown data.

  7. 07Output

    Locked sending

    Status lock, time window and probability gate: 500+ emails a day.

Inside the system

clay-ai-scoring/analysis
n8n analysis workflow: Clay webhook, signals, base score, LLM analysis, copy, facts guardrail and record

Analysis and ranking flow

clay-ai-scoring/sending
n8n sending workflow: business-hours gate, next lead fetch, validation, status lock and Gmail send

Email sending flow

Technical decisions

  • Fact vs. interpretation. Deterministic code produces the facts; the LLM only interprets.
  • Hybrid score. Rule-based 0–40 (opportunity 0–20, digital maturity 0–12, reachability 0–8); the LLM normalizes to 0–100 and adjusts by ±25 at most.
  • Closed evidence. The LLM may only cite facts from the list; invented evidence sends the lead to manual review.
  • Post-copy guardrail. Blocks placeholders, monetary amounts, unsupported figures and claims about unknown data.
  • Safe sending. Status lock before sending (no duplicates) plus a time and probability gate (500+ emails/day).
  • Real time. Webhook integration that works around Clay's CSV export limitation.

Result

Zero manual triage: every lead gets a score and a tier, and the best ones go out with a personalized, fact-checked email.