The gap between AI pilots and production AI is, more often than not, a data governance gap. At Black Hat USA 2026, executives from HPE, BigID, and Fortanix outlined the three-layer stack that enterprises now need to move AI from experimentation to deployment: discovery of what data exists and where (including shadow AI environments<\/a> built outside IT oversight), classification of sensitive assets, and encryption that protects data even inside GPU memory during active inference workloads.<\/p>\n The question most enterprises are getting wrong isn’t “can we build an AI model” but “do we actually know what data that model trained on, and can we prove it to a regulator.” If your organization has spent the past two years running AI pilots without a parallel data inventory effort, you’re not behind on AI, you’re behind on the prerequisite. The CISO is the one who absorbs that liability when production deployment exposes a compliance gap nobody mapped.<\/p>\n The shadow AI finding is the sharpest signal here. BigID’s scanning is surfacing rogue sandbox models built outside compliance oversight, and that pattern is almost certainly more common than most security teams want to admit. A developer spins up a model against a data export that was never classified, trains on it locally, and the resulting artifact carries data lineage nobody can reconstruct. The fix isn’t a policy memo. It’s continuous scanning coverage that reaches endpoints and SaaS environments, not just the data warehouse.<\/p>\n The on-premises rehoming trend Fortanix flagged deserves attention separately. Banks, healthcare systems, and government agencies are pulling AI workloads back from shared cloud environments because the data legally can’t move. That creates a specific architecture problem: how do you apply consistent governance policy across on-prem GPU clusters and cloud infrastructure simultaneously? HPE’s pitch with GreenLake is a unified control plane answer to that question, and the incentive to sell hybrid infrastructure naturally colors how they frame the complexity. But the underlying problem is real regardless of vendor. Organizations that haven’t decided yet whether their AI workloads belong on-prem or in the cloud are about to have that decision made for them by their compliance team.<\/p>\n The vendor to watch in your next renewal cycle isn’t the model provider. It’s whoever holds your data discovery and classification contract. If that tool doesn’t cover SaaS endpoints and can’t flag unregistered model training environments, it was scoped for a pre-AI threat surface. I’d revise this view if enterprises with mature data catalogs turn out to have meaningfully fewer shadow AI incidents, but so far the evidence runs the other direction.<\/p>\nWhat this means for your business<\/h2>\n