AI Guardrails
How the LLM layer is scoped, and what it's deliberately not allowed to do.
Advisory, not authoritative
Every lead gets a transparent, rule-based fit_score
first — a plain weighted breakdown across industry, size, revenue,
tech stack, geography, and title seniority, fully auditable with no
LLM involved. The LLM layer adds a qualitative
conversion_likelihood and rationale on top, meant to
surface context the rules can't (recent funding, hiring surges,
competitor mentions) — not to replace the rule-based number or make
a final call on its own.
Guardrails built into the product
- No autonomous outreach. Drafted emails/LinkedIn messages are generated for human review and manual send only. This is a deliberate scope decision, not a missing feature — auto-sending risks real spam/compliance exposure (CAN-SPAM, GDPR, CASL) that a human sender is accountable for and a script isn't.
- No silent CRM writes. Scores, rationale, and drafts only get pushed to Salesforce when you explicitly click the push action for a specific lead — nothing syncs in the background.
-
Mock output is always labeled. Without an
Anthropic API key configured, the scoring layer runs on
deterministic mock data instead of failing — and every mock
rationale is prefixed
[mock LLM]so it's never mistaken for a real assessment. - Self-hosted data path. When you do configure a real API key, the enrichment data for each lead goes directly from your own instance to Anthropic's API — it doesn't pass through us, and we have no visibility into what's sent or returned.
Limitations you should know about
Conversion-likelihood estimates are probabilistic, not guarantees — treat them as one input among several, particularly for high-stakes or high-value accounts. The rationale text is generated, and generated text can occasionally be wrong or miss context a human rep would catch. Nothing here replaces your own judgment about a specific prospect.