AI privacy is usually presented as a compliance problem: whether a provider trains on your data, retains your prompts, or shares information with third parties.
Those questions matter. But they are not the whole risk.
In a recent interview, Palantir co-founder and CEO Alex Karp described a concern he says is making customers livid: the “theft of their alpha.” His point was simple. A company puts its operating knowledge, data, and hard-won methods into an AI system, then discovers that a competitor may benefit from the same intelligence.
That is a more useful way to think about AI privacy. The thing at risk is not only personal information. It is the advantage inside the information.
Your advantage is rarely labelled confidential
A company’s most valuable knowledge is often distributed across ordinary material:
- product plans
- customer patterns
- pricing logic
- operational playbooks
- research notes
- financial models
- unfinished ideas
- the reasoning behind important decisions
Remove a company name and the material can still be commercially valuable. Anonymise a prompt and the method inside it may still reveal how a business works.
De-identified is not the same as deleted. A privacy model that protects names but exposes the underlying strategy is protecting the label, not the asset.
The closed-model trade-off
The convenience of a hosted model is obvious: upload context, ask a question, receive an answer.
The trade-off is less visible. The provider’s infrastructure sits between your organisation and the model. Even where a provider makes strong commitments about retention or training, those commitments are still policies. Policies can change, and they do not eliminate the need to trust the operator.
The real question is not only, “Will this provider train on my data?”
What can the provider see in the first place?
That is the difference between privacy as a promise and privacy as architecture.
Private AI should protect the context, not just the account
The answer is not to stop using AI. AI is already becoming part of serious work.
The answer is to build a boundary around the context that makes your work valuable. Files should be encrypted before upload. Storage should not expose plaintext. Inference should happen in a protected environment, with access limited to the computation that needs it.
That is the model behind ZDrive: your files are encrypted before they leave your device, stored as encrypted data, and used for private AI inference without giving the platform ordinary access to your underlying content.
Your files contain more than information. They contain the decisions, context, and unfinished thinking that make your organisation different.
Your AI. Your data. Nobody else’s.