Your Team Is Already Pasting Contracts Into ChatGPT - Here’s What That Actually Costs You

If you think your organization doesn’t have a shadow AI problem, you probably just haven’t looked. A 2025 survey from UpGuard found that more than 80% of workers, including nearly 90% of security professionals, use AI tools their employer never approved. This isn’t a junior-staff issue either - the same pattern shows up at the top of the org chart, where leaders who write the AI policy are often the ones quietly ignoring it.

The pattern is consistent across most of these studies, so let’s skip the debate about whether this is happening at your company and talk about what it means.

---

This isn’t hypothetical

In 2023, Samsung engineers pasted proprietary source code and internal meeting notes into ChatGPT to get help with debugging and meeting minutes. The data ended up stored on a third party’s servers, and Samsung responded by banning generative AI tools company-wide. No malicious intent, no hack, no phishing email. Just employees trying to get their work done faster.

A more recent survey of 1,000 employees found something similar at a larger scale: three out of four people using unapproved AI tools admitted to sharing potentially sensitive information through them, most often customer data, employee data, or internal documents - and 93% of executives and senior managers in the same survey said they personally use shadow AI at work too. Source: Journal of Accountancy

If your policy says “don’t use unapproved AI tools” and your VP of Sales is pasting a client contract into ChatGPT to summarize the renewal terms, the policy isn’t working. It just isn’t being enforced, or noticed.

---

Paying for the tool doesn’t solve this

A common assumption is that upgrading to a paid plan fixes the privacy problem. It doesn’t, not fully. Free and personal Plus accounts can still be used to train the underlying model unless someone manually opts out. Business and Enterprise tiers are better: OpenAI’s own enterprise privacy documentation confirms that business customer content is excluded from model training by default and conversations are encrypted. That’s a real improvement.

But notice what it doesn’t change: the document still leaves your infrastructure and lands on a vendor’s servers. You’re now trusting a contractual promise rather than controlling the boundary yourself. For most internal communications that’s a fine trade-off. For a signed contract, a patient file, or a financial statement, it’s worth pausing on.

---

The -Québec wrinkle most companies miss

If your organization handles personal information tied to -Québec residents, there’s a compliance layer on top of the security one. Law 25 requires a privacy impact assessment before personal information is communicated outside -Québec, and that obligation applies the moment a vendor’s servers (or support staff) are located somewhere else, even if the vendor is otherwise Canadian. BLG’s overview of the cross-border transfer requirements lays out what that assessment needs to cover, including how a destination jurisdiction’s surveillance laws factor into whether the transfer is even permitted. An employee pasting a client file into a US-hosted AI tool isn’t just a security question at that point - it’s a documented compliance gap.

---

What’s actually worth doing about it

A blanket ban rarely works - employees just move the behaviour somewhere less visible. A few things are more useful:

• Find out what’s already happening before deciding what to do about it. Most leaders are working from assumption, not data.
• Decide where the line is by document sensitivity, not by department. A marketing brainstorm and a client contract don’t belong in the same policy bucket.
• When you evaluate any AI tool for sensitive document work, ask directly where the data lives, who can access it, and whether it ever leaves infrastructure your organization controls. Vendors should be able to answer this in one sentence, not a sales deck.
• Treat this as an infrastructure decision, not just a training problem. Telling people not to do something they’ve already found useful rarely outlasts the first deadline crunch.

This is the exact gap we built Discreetly to close: document intelligence that runs on infrastructure your organization controls, with Canadian data residency built in rather than bolted on. But regardless of what you end up choosing, the first useful step is the same one: find out what’s actually happening in your organization before someone else finds out for you.