This section of the latest All-In episode stayed with me.
The panel was talking about AI being used as a lawyer, doctor, therapist, and research partner. Then they asked why AI conversations can receive weaker protection than email, even when the subject matter is far more personal.
That feels like one of those questions that sounds obvious after somebody says it.
We have put a lot of private thinking into a box that looks like a conversation. The box makes it feel one-to-one. The infrastructure underneath may not be.
Your name is not the only valuable thing in the prompt
A lot of privacy language stops at personal information. Remove the name, email address, and company identifier, and the data is treated as safe enough.
But what if the valuable part is the idea?
A research team can ask an AI model to work through a novel scientific problem. The company name can be removed. The names of the researchers can be removed. The underlying method can still be useful.
The product direction can still be useful. The commercial strategy can still be useful. The line of reasoning can still be useful.
De-identified is not the same as deleted.
Organisational IP is often not a document labelled “confidential”. It is the accumulated thinking inside the documents, prompts, decisions, and conversations.
That is the part privacy conversations often miss.
Local is useful, but it is not the whole answer
The obvious response is to run an open model locally. That can be the right answer for some work.
But a local model does not automatically give a team shared knowledge, retrieval, memory, collaboration, or a workflow people will actually use. A machine under your desk is not the same thing as a private knowledge system.
The real requirement is simpler to say and harder to build:
Use powerful AI with your private context without giving ownership of that context to the platform running the model.
Privacy needs to start before the prompt
This is where policy alone starts to feel thin.
A company can promise not to train on your data. That may be useful. But the promise can change. The company can change direction. It can enter your market. It can decide that aggregated insight is strategically valuable.
The stronger question is: what can the platform see in the first place?
ZDrive starts there.
Your files are encrypted before they leave your device. The vault stores encrypted data. Private inference runs inside a protected computing environment so the model can work with your context without the platform needing ordinary access to the underlying files.
The technology is not the story. It is the boundary that makes the story possible.
The point is not to stop using AI
AI is already part of serious work. Telling people to avoid it is not a strategy.
The better goal is to use it without turning your private context into somebody else’s training advantage, product advantage, or market intelligence.
Your files contain more than information. They contain judgement, unfinished ideas, decisions, and the context that makes your organisation different.
That is the asset worth protecting.
The question is no longer whether AI will become part of your most sensitive work. It already is.
The question is whether your privacy model is ready for that reality.
Private AI is not a policy checkbox. It is architecture.
Watch the source excerpt from 1:10:47, then try ZDrive.