Your SaaS Stack Is Leaking Data to AI Models You Never Approved

The number that should stop every enterprise CIO

There’s a number that should stop every enterprise CIO in their tracks: 63.6%.

That’s the share of SaaS vendors, according to a DataGrail analysis of 2,400 providers, that are actively sending your company’s data to AI models you never reviewed, never approved, and may not even be able to name.

Most procurement teams believe they’ve handled this. They negotiated an AI addendum. They signed a Data Processing Agreement. They checked the compliance box and moved on.

But here’s the problem: your vendor signed the addendum. Their subprocessor didn’t. And neither did the model provider sitting two layers down the supply chain.

The shadow AI supply chain

Enterprise software has always had a vendor ecosystem - a web of integrations, subprocessors, and third-party dependencies that power the tools your teams use every day.

What’s changed is that AI has been quietly woven into every layer of that ecosystem, often without explicit disclosure.

Consider what’s happening right now inside your own stack:

• Your CRM is shipping customer interaction prompts to a large language model to generate summaries and suggested responses.
• Your support platform is fine-tuning on your support tickets - your customers’ words, your internal processes, your product knowledge - to improve its AI features.
• Your analytics tool is sending usage telemetry to a model you can’t identify, to power predictive features you may not even use.

Each of these is what DataGrail calls a “workflow-intelligence leak” - a transfer of sensitive business data to an AI system that operates entirely outside your governance perimeter.

You didn’t consent to it. You may not even know it’s happening.

Why your AI addendum isn’t enough

The instinct to solve this with contracts is understandable. Legal teams have been drafting AI addenda and updated DPAs at a furious pace over the past two years.

But contractual governance has a fundamental limitation: it only binds the party that signs it.

Your vendor may have agreed to your AI use restrictions. But their infrastructure provider, their model API vendor, their fine-tuning partner - none of them are party to your agreement. The data flows anyway.

This is the core insight that DataGrail’s research surfaces:

Enterprise AI governance isn’t primarily about the LLM choices you make. It’s about the LLM choices your vendors make.

And those choices are largely invisible to you.

The problem compounds as AI capabilities become table stakes for SaaS products. Vendors that don’t ship AI features lose deals. So they integrate whatever model APIs are fastest to deploy, often prioritizing speed over transparency.

The result is a shadow AI supply chain that grows faster than any procurement team can audit.

The case for model ownership

There is one structural solution that addresses this problem at its root: owning your own models.

When your organization runs its own models - whether fine-tuned open-source models deployed on your infrastructure, or models hosted in a private cloud environment you control - the data flow question has a clear answer.

Your data stays within your perimeter. There is no third-party model provider receiving your prompts. There is no subprocessor fine-tuning on your tickets.

This isn’t a new argument, but the DataGrail findings reframe it.

Model ownership has often been positioned as a preference for organizations with unusually high security requirements - defense contractors, regulated financial institutions, healthcare systems.

The new reality is different:

If 63.6% of your SaaS vendors are routing your data through AI models you haven’t approved, then model ownership isn’t a preference for the security-conscious. It’s the only reliable way for any enterprise to answer the question:

Whose model is reading my data?

What this means for enterprise AI strategy

The implications extend beyond security teams. For CIOs and CTOs building enterprise AI strategies, the DataGrail findings suggest a few concrete shifts:

1. Vendor AI disclosure should be a procurement requirement. Before signing any SaaS contract, require explicit disclosure of every AI model the vendor uses, including subprocessors, and the data categories each model processes.
2. AI addenda need to travel down the supply chain. Contractual AI restrictions are only meaningful if they bind every party in the data flow - not just your direct vendor.
3. Model ownership deserves a place in your AI roadmap. For workloads involving sensitive data, the calculus on building or deploying your own models has shifted. The cost of ownership is increasingly competitive with the governance risk of relying on vendor-embedded AI.