Why Enterprise AI Deals Die in the Security Review: What CISOs Actually Block and Why

At TechCrunch Disrupt 2026, Databricks co-founder Ali Ghodsi told the audience what every enterprise AI vendor already knows but rarely admits out loud: most enterprise AI deals do not die because the technology fails. They die in the security review. The Chief Information Security Officer, not the buyer, holds the final veto, and a new wave of risk data shows exactly which signals trigger that veto first.

The pattern is consistent across Forrester, IDC, and ISACA research from Q1 2026: pilots get green-lit, technical proof-of-concept succeeds, then the deal stalls for months while CISOs work through four checkpoints that vendors routinely fail. Compounding the problem, a 2026 Cyberhaven report shows that AI risk inside the enterprise is not evenly distributed. It is concentrated in a small group of power users, which is exactly the profile that makes CISOs nervous.

The Four Checkpoints That Kill Deals

Every enterprise security review walks through the same four checkpoints, and a failure on any one of them is enough to halt the deal. Data residency is the question of where training data, prompt data, and inference data physically live. A CISO needs to confirm whether prompts touch a US region, an EU region, or both, and whether logs are retained on-prem or in the vendor cloud. Model provenance asks whether the model is a generic foundation model, a fine-tuned variant trained on customer data, or a hybrid. If customer data was used to train the base model, the legal team gets involved and timelines extend. Access scope evaluates what the AI agent is allowed to do once deployed. Read-only access is approved quickly. Write access to production systems, CRMs, or financial records triggers a full risk assessment. Audit trails are the last gate: can the CISO, six months from now, reconstruct exactly which prompt the AI saw, which model version responded, and which downstream system it touched? Without immutable logs, the deal usually dies.

The Power User Problem: Why 8% of Your Team Is Your Biggest Risk

The Cyberhaven 2026 AI Risk Report found that AI risk inside the enterprise is wildly concentrated. Roughly 8% of employees - the so-called AI power users - generate over 55% of all AI-related data exposure incidents. These are the engineers, analysts, and executive assistants who paste source code into ChatGPT, drop customer contracts into Claude for summarization, and chain unapproved models together into shadow workflows that no IT team has reviewed. CISOs flag this profile during vendor security reviews because the vendor cannot guarantee end-user behavior. Even a perfectly compliant AI platform becomes a liability when 8% of the workforce treats it as a personal sandbox. During review, the CISO will ask: how do you detect prompt injection from internal users, how do you prevent sensitive data from being pasted into prompts, and how do you enforce model allow-lists? Vendors who cannot answer these three questions in the first meeting almost always slide into the 4.7-month delay window IDC measured.

What Separates the Deals That Close

The pattern is consistent. Vendors who arrive at the first security meeting with a pre-built security package close deals 2 to 3 times faster than vendors who treat security review as a series of ad-hoc requests. The package is specific. It includes a Data Processing Agreement (DPA) drafted to GDPR and CCPA standards, a Business Associate Agreement (BAA) ready for any HIPAA-adjacent deal, model cards aligned to NIST AI Risk Management Framework 1.0, a SOC 2 Type II report less than 12 months old, a recent penetration test summary, and an access log API that lets the customer pull immutable audit trails into their own SIEM. The vendors who close also bring a one-page data flow diagram that maps every place customer data travels, from prompt to inference to log retention to deletion. CISOs sign off when they can hand that diagram to their legal and compliance teams without needing follow-up questions.

How to Survive the Security Review

1. 1 Map data flows before the CISO meeting. Know exactly where inference calls go, where logs are stored, and where deletion happens. Document it on one page.
2. 2 Produce a model card aligned with NIST AI RMF 1.0 before the first security touchpoint. Include intended use, limitations, training data provenance, and evaluation metrics.
3. 3 Identify your power user cohort internally. Present a usage governance policy to the CISO proactively rather than waiting for them to surface the risk.
4. 4 Negotiate data residency options upfront. EU, US, and on-prem deployment should all be on the table from the first conversation, not introduced as concessions later.
5. 5 Request a formal security package from the vendor: SOC 2 Type II report, penetration test summary, and an access log API that exports to your SIEM.