- Your AI SDR doesn't have a writing problem. It has a governance problem, and a better model won't fix it. A perfect model still can't know your current certifications or self-certify it broke no EU rule. So we put a deterministic gate between the draft and the send. 🧵
- The setup: an AI drafts each email in a named rep's voice. Then a pure-Python verifier crew runs four independent checks before anything can send. Factual grounding, deliverability, EU AI Act Article 5, style fidelity. Agents advise, code decides.
- Take one prospect from the demo: Chris Tanaka at Vaultline, a renewal-pressure security buyer. The draft read clean. It also claimed "SOC 2 Type II certified and fully HIPAA certified." Our product knowledge base holds only SOC 2 Type I. HIPAA isn't held at all.
- The factual check verifies every claim against a source-of-truth, not the model's confidence. Verdicts are SUPPORTED (with a citation), UNSUPPORTED, or CONTRADICTED. Both certification claims came back unsupported. Facts check: red.
- The same draft added "Only 2 onboarding slots left this quarter, before Friday" and "Most of your competitors have already moved." Manufactured urgency and deceptive social proof. Both are named patterns under EU AI Act Article 5, enforceable since Feb 2025. Article 5: red.
- This is the part engineers ask about. The policy gate is deterministic and sits outside the LLM. An email is CLEAR only if there are no unsupported claims, deliverability is 0.7 or higher, and Article 5 is clean. Anything else BLOCKS and routes to a human with the reasons.
- The optional LLM verifier can only ADD unsupported findings. It can never CLEAR a claim the deterministic check flagged, and it cannot override the gate. An LLM grading an LLM isn't certifiable. Same input, same verdict, every run.
- Every email produces a signed send-receipt (JSON + HTML): model and version, the retrieved style sources with match scores, each claim with its verdict and citation, the sub-scores, and the gate decision. A compliance team gets a filable audit trail. We keep receipts.
- Does the gate hold up? On a 5-case labeled adversarial set it scored 5 of 5: the false cert, a sub-floor price, manufactured urgency, deceptive social proof, and the one clean email it let through. Deterministic, so reproducible. Those numbers are from that fixed set.
- Real question for RevOps and compliance folks: when an AI SDR over-claims a certification you don't hold, where should it get stopped, before send or after a prospect forwards it to their own legal team? #RevOps #AICompliance #AIGovernance #SalesAI
- Everything above runs in a demo you can click through: the clean case, the blocked case, the receipts, the benchmark. Every rep and prospect is synthetic. See it: https://veriprajna.com/demos/ai-sales-personalization
Published on X · July 31, 2026
On social media
See this post on its original platform
In our archive