The 2017 Contract That Already Knew What to Do With Your AI

Yuri was about to start a new job, consulting for a large company. Before signing anything, he did what he always does: he read the code of conduct and the access agreement, with an AI assistant helping him work through the denser parts. The documents were dated 2017 and 2018. Naturally, there was no clause anywhere about an "AI assistant," because in 2017 that phrase meant nothing to whoever drafted the contract.

And yet the clauses had it covered. One of them said, roughly, that data cannot be shared with third parties without prior approval. Read literally, that line applies perfectly to pasting a chunk of company code into a personal AI chat. The third party doesn't have to be a person. It can be a server belonging to another company, on the other side of the world, holding onto that chunk indefinitely.

That's this week's lesson: anyone working with AI needs to reread old paperwork with new eyes. The rule was written before the tool existed. Even so, it still governs the tool. Nobody updated the contract to mention AI, because nobody needed to. The word "third party" already covered it, the whole time.

The uncomfortable part is that the right question isn't "can I use AI on this job." It's a narrower, more annoying one to answer: which tool, under which account, approved by whom, in writing. Without those four answers, using AI for anything work-related is a bet, not a decision.

And this isn't just excessive caution. Large companies have already run into this publicly. An electronics maker restricted employee use of generative AI tools after discovering that sensitive company data was being typed into these services, which belong to outside companies and hold on to what they receive. A major bank restricted use of this kind of tool among staff. A large retailer warned employees not to put confidential information, including code, into these assistants.

The underlying problem, according to contract specialists who've written about this, is that confidentiality agreements from before generative AI already prohibit, unknowingly, feeding confidential information into open AI systems. Doing so effectively transfers the information to whoever provides the tool, on a non-confidential basis, and from there to who knows where. Most confidentiality agreements carry a line like "confidential information may not be disclosed to third parties without prior written consent." That line never mentions AI, because it doesn't need to. The trouble is nobody, at signing time, thought of it as being about AI. It only comes up after something has already gone wrong, when someone has to decide whether that assistant counted as a third party. By then it's too late.

There's a name now for what Yuri nearly did without meaning to: shadow AI. It's the use of AI tools by employees without approval or oversight from whoever handles information security. The name is new. The behavior isn't. Between 2023 and 2024, enterprise employee adoption of generative AI tools jumped from 74 percent to 96 percent. More than a third of employees admit to having shared sensitive work information with an AI without permission. Nearly half say they've adopted some AI tool without company approval, often the free version, which happens to be the one that guarantees the least about what happens to whatever you type into it.

The more telling number isn't about employees. It's about bosses. Most presidents and senior executives say they're comfortable with this unauthorized use, favoring speed over caution in the race to adopt AI. That changes how you should read Yuri's story. It isn't about one careless employee taking a risk. It's about an entire organization, top to bottom, collectively deciding it's easier to use the tool now and sort out the contract later, if it ever comes to that.

Yuri did the opposite. He read the contract before touching any tool, not after. It's extra work that most people skip, because no manager asks for it on day one and no onboarding session mentions clauses signed years before AI existed. But the contract never asked whether he'd read the company's AI policy. It asked whether he was about to share data with a third party. And without an approved account, in writing, the answer was yes.

Maybe the most useful advice here isn't technical, it's a habit: before opening a new tool on a new job, it's worth asking who already wrote a rule about this, even if the rule was written for something else entirely. Old paper, read carefully, still governs new tools. The temptation is to assume a clause from 2017 can't possibly speak to something that only exists today. It can. It speaks to the action, not the tool. The action is still exactly the same one: taking data that isn't yours and sending it somewhere else.