Shadow AI: 63% of Business Software Sends Your Data to AI Models You Never Approved

What “Shadow AI” Actually Means

You’ve heard of shadow IT: employees using unauthorized apps at work. Shadow AI is worse. It’s not rogue employees. It’s the vendors themselves.

Here’s how it works. A company buys an AI-powered customer service tool. The vendor’s marketing page says “AI-powered.” The contract says the vendor processes your data. What it doesn’t say: your data is also being sent to a third-party AI model (maybe OpenAI, maybe Anthropic, maybe some smaller outfit) for processing, training, or inference. That third-party is a “subprocessor,” and under most privacy laws, you need to know about it [1].

DataGrail tracked 2,400 leading business systems for their 2026 report. The finding that should keep every privacy officer up at night: 63.6% of vendors that prominently advertise AI capabilities do not disclose a third-party AI subprocessor in their legal documentation [1].

Not “some.” Not “a few bad actors.” Nearly two-thirds.

“If there’s one word that sums up data privacy in 2026, it’s ‘more’: more regulation, more risk, more pressure,” said Daniel Barber, CEO of DataGrail. “The only thing there isn’t more of is privacy professionals to handle it” [2].

The Numbers Are Ugly

DataGrail didn’t just look at vendor disclosures. They audited the entire privacy operations pipeline across hundreds of enterprise customers and 5,000 websites. The picture isn’t pretty [1][2][3]:

• 63.6% of AI vendors don’t disclose third-party AI subprocessors
• 32.8% of AI systems participate in at least one high-risk activity: sensitive data processing or automated decision-making
• 145 AI-related laws enacted by state legislatures in 2025, with 1,000+ additional bills introduced
• 567% surge in data deletion requests since 2021 (fifth consecutive year of all-time highs)
• 87% of all data subject requests are deletion requests (“delete my data”)
• 398% increase in deletion requests to data brokers from 2024 to 2025
• 63% of websites fail to honor Global Privacy Control and universal opt-out mechanisms
• $1.5 million per year: estimated cost for a mid-sized company to manually process data subject requests
• $4.3 million in public CCPA consent settlements in California in 2025
• 1,400+ class action lawsuits filed over tracking pixels and session replay

And the kicker: 42% of companies abandoned AI projects in 2025 specifically because of data privacy concerns [3]. Companies know this is a problem. They’re just not fixing it fast enough.

One in Three AI Systems Touches High-Risk Data

That 32.8% number deserves its own spotlight. Nearly one in three AI systems flagged by DataGrail participates in high-risk activities: processing sensitive personal data, making automated decisions about people, or both [1].

Think about what that means in practice. The AI screening your job application might be sending your resume to a third-party model that your prospective employer never vetted. The AI analyzing your health insurance claims might be routing your medical data through subprocessors that aren’t listed in any privacy policy you ever clicked “agree” on.

Under GDPR, that’s potentially illegal. Under CCPA, it’s a violation. Under the 145 new AI laws passed in 2025, it’s a minefield. But enforcement takes years, and the data has already been processed.

People Are Fed Up: The Deletion Surge

The 567% increase in deletion requests since 2021 tells its own story. People are done asking nicely. They want their data gone [1].

Data brokers got hit hardest: a 398% spike in deletion requests between 2024 and 2025 alone. The average company now handles over 2,000 monthly deletion requests and 900+ “do not sell/share” requests [3].

But here’s the problem. When you submit a deletion request to Company A, they might delete your data from their systems. What about the AI subprocessor they never told you about? The one that already ingested your data for model training? The one that isn’t listed in their privacy documentation?

You can’t delete data from a system you don’t know exists.

And the cost of handling all these requests manually? $1.5 million per year for a mid-sized company with 5 million annual visitors [1]. Privacy teams are already shrinking (headcount cuts of up to 33%) while the workload is exploding [3].

Your Opt-Out Button Doesn’t Work

DataGrail audited 5,000 popular websites for consent compliance. The result: 63% don’t honor Global Privacy Control (GPC) signals, the browser setting that says “don’t sell my data” [1].

GPC is legally binding in over 10 states. California’s CCPA explicitly requires companies to treat it as a valid opt-out. And yet nearly two-thirds of websites just ignore it.

Only 15% of users actively opt out of non-essential tracking [3]. That’s not because 85% of people love being tracked. It’s because the opt-out mechanisms are buried, broken, or simply ignored.

California hit companies with $4.3 million in CCPA consent settlements in 2025 [1]. That sounds like a lot until you realize it’s pocket change for companies processing millions of users’ data. The math still favors non-compliance.

It’s About to Get Worse

Gartner predicts 40% of enterprise applications will feature task-specific AI agents by the end of 2026, up from under 5% in 2025 [1]. That’s not a gradual increase. That’s an 8x jump in one year.

AI agents don’t just process data. They make decisions, take actions, and interact with other systems autonomously. If shadow AI is the vendor quietly sending your data to an undisclosed model, AI agents are the next evolution: automated systems making decisions about you using data from sources you never authorized.

California’s CCPA now mandates privacy risk assessments with annual audits beginning April 2028, personally attested by company executives under penalty of perjury [3]. That’s the regulators trying to catch up. But April 2028 is two years away, and the AI agent explosion is happening now.

What You Can Do

• Enable Global Privacy Control: Install a browser that supports GPC (Firefox, Brave, DuckDuckGo) or add a GPC extension. Yes, 63% of sites ignore it. But the ones that don’t are legally bound to honor it, and it gives you standing to complain to your state AG when they violate it.
• Submit deletion requests: Use services like DataGrail, YourDigitalRights, or Privacy Bee to send mass deletion requests to data brokers. The 567% surge shows these are working.
• Ask your employer about AI vendors: If your company uses AI-powered software, ask the IT or legal team: “Which third-party AI subprocessors does this vendor use?” If they can’t answer, that’s a problem. If the vendor can’t answer, that’s a bigger problem.
• Check vendor subprocessor lists: Reputable vendors publish subprocessor lists. If your vendor’s privacy documentation doesn’t mention any AI subprocessors but their marketing says “AI-powered,” you’ve found a shadow AI problem.
• File complaints: If a website ignores your GPC signal or fails to process a deletion request, file a complaint with your state attorney general. California’s $4.3 million in settlements didn’t happen on their own: complaints drive enforcement.
• Read the privacy policy (yes, really): Specifically, search for “subprocessor,” “third party,” and “AI” in the privacy policy of any service that handles sensitive data. What you don’t find is often more important than what you do.

Sources

1. DataGrail: “Privacy and AI Trends Report 2026: Shadow AI Emerges as a Growing Threat” (June 1, 2026)
2. VentureBeat: “DataGrail Report Finds Your Vendor May Be Sending Data to AI Models You Never Approved” (June 2026)
3. Help Net Security: “145 AI Laws Passed in 2025 and Privacy Teams Aren’t Catching a Break” (June 1, 2026)

Related Coverage

• The SECURE Data Act: June 3 Hearing Could Kill 20 State Privacy Laws
• Philly Cops Are Tracking People Who Criticize AI Data Centers Online