- A sales AI wrote to Werner Enterprises: "recently growing your One-Way Truckload fleet to 2,735 trucks." Every word is true, straight from Werner's real 10-K. And it still should never have been sent. One word is the reason 🧵
- The number is correct. The source is Werner's FY2023 Form 10-K, filed 2024-02-26 with the SEC. The problem is one word: "recently." That filing is over two years old now. A true fact wearing a recency claim on a stale source leaves a false impression.
- This is the failure mode almost nobody screens for. Not bad grammar. Not a fabricated number. A claim that is correctly cited but misleadingly used. We call it contextual misuse, and it is the bug most publicized AI SDR flameouts quietly share.
- Personalization is not verification. Tools that pull SEC filings to personalize outreach never re-check whether the resulting claim is still current for this prospect. They optimize for signal, not for whether the finished sentence is actually safe to send.
- So in the demo an AI drafts, then deterministic Python decides what ships. Not an LLM judging an LLM. One check: if a claim uses recency language, the cited source must be within 365 days. Werner's 10-K is older, so the line is marked stale and stripped.
- Click that line and you see the actual computed check: source age against the 365-day window, grounding, entity match, each one pass or fail with the real expression shown. Nothing hand-wavy. You can trace exactly why a sentence was allowed or killed.
- When it strips the stale line, the email still sends, just without it. The design guarantee: every claim in what actually goes out is source-backed, so sent integrity is 100%. Not "zero hallucination." The model still drafts. Unproven lines just never reach the buyer.
- Because the verifier is deterministic, the same input gives the same verdict every run. On a fixed 25-case labeled golden set it scored 25/25: every bad claim caught, every clean one kept. Reproducible in a way an LLM judge is not, which is what makes it auditable.
- And this does not age out as models get better. A stronger base model still cannot prove to FINRA or a compliance team which current source backed which claim. Provenance, an audit receipt, and a policy gate are durable. Raw drafting quality is not.
- Genuine question for RevOps and compliance folks: would you let an AI SDR auto-send a claim sourced from a filing over a year old if the number is technically still accurate? Where exactly is your line on source age? #RevOps #AICompliance
- We built this as a demo, not a deployed pipeline (the integrations are stubbed). If your team is weighing how to put AI outreach in front of regulated buyers, we would like to hear how you draw these lines. Walkthrough: https://veriprajna.com/demos/ai-sales-intelligence
Published on X · July 17, 2026
On social media
See this post on its original platform
In our archive