Our AI wrote a sales email swearing the product was SOC 2 Type II and HIPAA certified. It holds neither. The knowledge base behind it lists SOC 2 Type I. Not Type II. HIPAA isn't in there at all. But the sentence read clean, the tone was warm, and it was addressed to a VP of Engineering at a security buyer. Exactly the person who checks. What surprised us building this: a claim like that sails past the checks most AI SDRs run. Spelling, tone, no spammy words. It fails a different one. We route every factual claim through a grounding check against the approved source-of-truth, so "SOC 2 Type II" gets matched against what the company can actually prove. It can't, so a deterministic gate, not the model, blocks the send and hands it to a human with the reason. Agents advise, code decides. (The reps and the buyer here are synthetic. The failure mode is not.) A better model wouldn't have saved us. A perfect writer still can't know which certifications you hold today. Only a check against your own source-of-truth can, and it has to live outside the thing doing the writing. So the question we keep asking sales teams: after your AI drafts an email, does anything re-check the finished claim against your real certs and pricing before it sends, or does it go out on the model's word? #RevOps #AISDR #AICompliance #SalesAI