Your Valuation Is Your IP, and It Can Leave One Paste at a Time
A senior engineer at a Series B biotech is three weeks from a diligence deadline. To move faster, they paste a chunk of the assay pipeline — proprietary parameters, a novel scoring function, comments describing the method — into a public AI tool and ask it to refactor the code. The refactor is good. The problem is that the thing that made the company worth what the term sheet says it's worth just left the building, and there is no way to call it back.
For an R&D-intensive company, this isn't a data-hygiene footnote. Your valuation is your IP. The multiple an investor or acquirer puts on you is, in large part, a bet on inventions that are still defensible — patentable, secret, or both. Each of those properties is fragile in a specific, legal way, and a single prompt can break it.
Why the crown jewels move through the prompt box
The uncomfortable structural fact is that the people closest to the most valuable IP are exactly the people most likely to paste it into an AI tool. Your best engineers and scientists are not careless — they are productive. They reach for the fastest tool that helps them ship, and the fastest tool is often a general-purpose model in a browser tab.
The scale here is not marginal. LayerX found in 2025 that 77% of AI users paste data directly into prompts, and 82% of that pasted content comes from unmanaged personal accounts. Cyberhaven measured the sensitive share of corporate data flowing into AI rising to roughly 35%, up from about 11% two years earlier. The trend line points one direction: more of what matters, moving through a channel most organizations can't see.
And it moves in high-value chunks. A support rep summarizing a ticket leaks a name. An engineer debugging leaks the method that the name is attached to. When the pasted material is the invention — a training pipeline, a molecule's synthesis route, a draft patent claim — the exposure isn't a record, it's the asset.
Disclosure and irreversibility: two doors that don't reopen
Pasting IP into a public tool damages value through two mechanisms, and neither is recoverable after the fact.
The first is legal exposure. Trade-secret protection depends on the information actually being secret — on the holder taking reasonable steps to keep it that way. In Trinidad v. OpenAI (N.D. Cal., Jan 2026), a trade-secret claim was dismissed because developing the alleged secrets through ChatGPT was treated as voluntary disclosure. Secrecy, once surrendered, doesn't come back. The same logic threatens patentability: a public disclosure of an invention before filing can compromise novelty. For a company whose valuation assumes a granted patent, that is a direct hit to the number.
The second is irreversibility. Once content is submitted to a public AI tool, it can't be recalled. It may be retained, processed by sub-processors elsewhere, or used to train the provider's models — and pasted content becomes subject to that provider's terms of use, which can grant broad rights to retain and use it. The Samsung case in 2023 remains the cleanest illustration: within roughly twenty days of allowing ChatGPT, engineers had pasted in source code, a defect-detection algorithm, and an internal meeting transcript. The data couldn't be recalled. The response was a company-wide ban.
That's the trap. After submission there is nothing left to control. The window to protect the asset closes the instant the prompt is sent.
Why the ban doesn't hold
Samsung's instinct — shut it off — is the one most leadership teams reach for, and it's the one that fails. Bans don't remove AI from the workflow; they push it onto personal accounts and phones where the company has no visibility at all. Gartner's 2026 survey found that 88% of employees with enterprise AI access also use personal AI tools for work, and that 69% of organizations suspect or have evidence of prohibited public GenAI use. A prohibition you can't observe isn't a control. It's a blind spot with a policy attached.
The deeper reason bans fail in deep-tech is cultural. You hired for velocity. Telling your researchers they can't use the tools that make them faster either gets ignored or slows the very output your valuation depends on. The goal isn't to stop people from using AI. It's to make sure the invention doesn't leave when they do.
That means moving the control to the only place it can still work: before the prompt reaches the tool. If sensitive content — source code, draft claims, synthesis parameters, deal data — is caught and redacted before submission, the engineer still gets the refactor, and the crown jewels stay inside the organization's control. Just as important, the moment of near-miss becomes a moment of learning: a plain-language explanation of what was flagged and why teaches the team to recognize the boundary next time, without a meeting or a memo. This is the principle Sanitized AI is built on — act at the prompt, teach in the moment, keep the velocity.
The question to ask this quarter is narrow and answerable: if one of your engineers pasted the core of your technology into a public model tomorrow, would anyone know, and would anything have stopped it before it left? If the honest answer is no, that gap is sitting directly on top of your valuation. If you want to see what closing it looks like without slowing your team down, request a demo.