What Enterprise AI May Know

A confidential conversation with the management team. A quick voice note on the way to the next meeting. A discussion in a project channel. Knowledge that can be valuable for working with AI is created everywhere.

With every additional surface, however, one challenge grows: the AI should be able to recognize connections and pick up conversations where they left off. At the same time, confidential information has to stay confidential, personal thoughts have to stay protected, and outdated assumptions have to lose their influence.

That is why, in the Validation Lab, we are working on a question that goes beyond the choice of a language model: how do we design a memory for AI that works reliably in everyday corporate life?

Knowledge Beyond the Boundaries of a Chat

Our starting point was collaboration via Mattermost. There, channels and threads give the exchange a clear structure. With additional points of access, voice interaction for instance, that structure has to work across individual surfaces as well.

If you first discuss an idea with the AI and later develop it further in the project, you don’t want to explain the whole background again. That is what we are building cross-channel knowledge management for.

The decisive point: a shared knowledge base does not mean the AI should know or use everything for every question. For the task at hand it needs a fitting excerpt - relevant, current, and cleared for the specific situation.

The Place of the Conversation Is Part of the Permission

Imagine the management team discussing confidential budget information with the AI. Later, the same person asks about the project status in an open company channel.

The identity of the person asking is not enough on its own to determine the right answer. What also matters is where the conversation is taking place and who can read along.

That is why, for us, the specific channel is part of every request. Our Dissemination Control System takes this context into account and limits access before protected content reaches the model’s context window.

Confidentiality has to take effect while the knowledge is being selected. When someone switches between surfaces, a previously permissible conversation context must not simply carry over unchecked.

Personal Thoughts Need a Space of Their Own

Not every conversation with an AI is immediately corporate knowledge.

People need room for unfinished ideas, personal notes, and first assessments. At the same time, they should be able to assign such content to a project later, or share it deliberately with others.

In doing so, we distinguish between subject-matter assignment and clearance. A note can belong to a project and still remain personal.

Operating it should stay simple: the agent supports the classification and asks when something is unclear. In the conversation, you can decide where a piece of information belongs and who may use it.

Memories Have to Be Correctable, Too

Another challenge is easily underestimated: stored knowledge does not automatically stay correct.

Hypotheses are disproved. Decisions change. An earlier assessment can become useless after a new experiment. In a Validation Lab in particular, that kind of development is part of daily work.

So you have to be able to tell the AI: “This assumption no longer holds” or “For this project the previous rule still applies”.

The original insight remains traceable but takes on a changed status. That makes it possible to understand later why a decision seemed sensible at the time, without using the same assumption as a valid basis today.

The AI can point out contradictions. Where assessments are contested, the human decision remains authoritative.

Trust Extends into the Infrastructure

Part of confidentiality is the question of who can access stored content behind the surfaces.

That is why, in addition to access control, we are looking at the encryption of sensitive information and the separation of data storage from key access. This protection concept has to cover summaries and search functions as well.

These questions define the next stage of our Validation Lab. Our ambition is an AI that makes existing knowledge usable at the right moment, respects its limits, and can be corrected.

Which of your business processes would benefit if the AI kept the context - and you kept control over what is used from it?