Confidentiality by Design, Not by Prompt: Introducing Projects

Ask an AI system a question and it searches everything it knows to find the best answer. That's normally the whole point. But when what it knows includes documents from three different clients, "best answer" and "right answer" can quietly become two different things.

That's the problem Projects in Discreetly AI is built to solve.

One Shared Pool Is a Confidentiality Problem

Most document AI tools work the same way underneath: everything uploaded gets indexed into one searchable pool, and every chat searches that whole pool for the closest match. For a single-client use case, that's fine. For a law firm running a wrongful dismissal file and a commercial lease dispute at the same time, or a clinic with more than one patient's records on hand, it's a real risk.

A chat scoped to "the lease dispute" has no inherent reason to ignore a paragraph from the dismissal file, if that paragraph happens to be the closest match to the question asked. Nothing stops it. The documents are all sitting in the same pool.

How Projects Work

A project is a named container for one matter, one client, whatever boundary actually matters. Creating one takes a name and a click.

From there, documents get tagged to a project the moment they're uploaded.

This is the part worth being precise about, since it changes what you can actually promise a client:

What It Looks Like in a Chat

Start a new chat, select a project, and every answer from that point on comes only from what's tagged there, with sources cited so the answer can be checked against the original document.

Here's a real example, using a fictional statement of claim tagged to a "John Doe vs Jane Doe" project. The question asked was straightforward: what's the basis of claim for this matter. Discreetly AI answered using only that one document, with the sources shown pointing straight back to it.

Quick recap of what actually changes with Projects turned on:

• Documents tagged to a project never surface in a chat outside that project, by exclusion, not by ranking.
• Project selection locks once a chat starts, so scope can't drift partway through a conversation.
• Every answer still shows its sources, project-scoped or not.
• Untagged, general documents stay separate from every project too.

For firms juggling multiple client matters, or clinics with multiple patients, or manufacturers running multiple RFQs at once, that separation is the actual product, not a feature bullet point.

It's also the difference between confidentiality as a policy people are asked to follow, and confidentiality as something the system enforces on its own, without anyone having to remember to set it.