# Your working knowledge deserves a private space

2026-10-20 · Bubbling AI team

Canonical: https://bubblingai.com/blog/your-working-knowledge-deserves-privacy

The value in your context is a reason to protect it. How we think about provider privacy, automated safety checks, and a personal place to work.

*Our direction, while Bubbling AI is in development. This explains the privacy standard we are building toward; sign-up is not open yet.*

The useful part of a workspace is often more than the final file. It is the reason behind a decision, the approach that failed, the detail a customer cares about, or the experience that helps you recognize a good result.

That working knowledge has value. We want you to be able to bring it to an AI partner without treating it as material for someone else to collect and study.

This is why privacy matters to Bubbling AI. It is part of respecting the work and the person behind it.

## The assistant works with your context for you

An assistant needs information to help. It may read your selected notes, search your workspace, or send relevant context to a model. Those are parts of doing the work you asked for.

Our commitment is that your content is not a staff browsing feed, a training dataset for us, or conversation text copied into product analytics. Choosing to share feedback or authorize support is a separate, deliberate action with its own scope.

Company workspaces also need a visible distinction between personal work memory and company-held or shared material. A private discussion and a report you publish to your team should not be treated as the same thing. The workspace policy must explain that boundary before you rely on it.

## Privacy includes the model provider

For Bubbling AI-managed model calls, our policy requires approved **zero-data-retention and no-training routes**. The purpose is to obtain the answer without turning the prompt and response into content retained for model training.

That policy must cover the actual model, provider, supported feature, and fallback path. A label on a provider is not enough. We check routing eligibility against approved information and need to keep it current as models and terms change.

Zero data retention is a specific processing policy. It is separate from the conversations and records saved in your own workspace, and it does not mean that no operational or legally required records can exist. Tools you connect through your own accounts also keep their own terms; our managed-model policy does not automatically transfer to them.

We want some of the confidence and control people value in a local AI setup to be available in an easier hosted experience. This is still cloud processing: approved providers receive the context needed to answer or check a request. It is not the same as running a model entirely on your device.

## Automated checks are not a person reading over your shoulder

Useful AI also needs boundaries against harmful requests and instructions hidden inside material it reads. A web page or uploaded document should not be able to quietly redirect the assistant or obtain access it was never given.

Our design uses automated checks before relevant material enters an AI workflow and before certain actions can proceed. Some checks use an approved cloud classifier, so the relevant content is processed by that service under the selected route's data rules. A result can allow, withhold, or pause forwarding; an incomplete check should not pretend the content passed.

Uploaded files have another layer: server-side scanning for malicious bytes. That is separate from evaluating the text used by an assistant. Neither layer can promise to detect every threat.

These automated paths do not make ordinary staff review part of every conversation. They do produce bounded security records. Some refused outgoing content may also be kept encrypted in the person's account so the person can inspect what was stopped. If someone requests an appeal or support review, specific content can become available through that separately authorized process.

Service rules still apply. Repeated or serious violations can affect access, and providers can apply their own policies. Screening is not a promise that a request can never lead to an account restriction.

## Clear boundaries are part of trust

We do not need to say another company is unsafe to explain our choices. We need to say what we process, why we process it, what is saved, who can access it, and what changes when you connect another tool.

The goal is a workspace you can understand well enough to trust: useful memory, careful sharing, and a clear distinction between work done for you and content collected for another purpose.

Read [our privacy explanation](/privacy-and-security), [why people remain central](/blog/ai-should-make-room-for-people), and [how we approach context](/blog/context-you-can-work-with).