You passed SOC 2 and the deal still stalled. That's because security reviews now ask AI-specific questions that sit entirely outside a SOC 2 report — how your model uses customer inputs, who your LLM subprocessors are, what's retained. Founders lose enterprise deals here every week. Below are the eight questions buyers actually send. Mark the ones you can answer cleanly today and see where you're exposed.
Every "not yet" is a question a buyer's security team can stall your deal on. The fix is rarely a new audit — it's documenting the controls you already have and putting an answer on the record a GC will accept. That's what a signed attestation does: it turns these eight questions into one artifact the buyer can rely on.
See how the attestation works → or tell me about a stalled deal →
SOC 2 attests to controls around security, availability, and confidentiality at the organization level. It was never designed to answer questions about how a large language model treats the specific data your customer sends through it. So a Series A vendor can hold a clean SOC 2 report and still get frozen in procurement when a buyer's security team asks "does our data train your model?" and the answer is a shrug.
The questions above are the recurring gap. None of them require a new framework or a six-month audit. They require knowing your own data flow — which provider sees what, what's retained, what's contractually off-limits — and being able to put that on the record in a form a reviewer will accept. Most AI-native companies already have defensible answers; they just haven't documented them, so the deal sits while a sales engineer and a security analyst trade emails.