Whether AI builds you a useful business model depends far less on the prompt you type than on the content you give it to read. Think of the glossary that settles what counts as a customer, the requirements from the last workshop, the vendor data dictionary nobody has opened since go-live.

Content is step two of the five-step loop this how we use AI series follows: set the context, bring in the content, describe the outcome, review the proposal, then refine and approve. The last session, Business context is not an AI prompt, set the context in your own words. This one fills the library the AI reads while it works, because no clever instruction later makes up for thin material here. Skip this step and the AI decides what counts as a customer from whatever fit in the chat box. You spend the review arguing with a definition nobody on your team wrote.

Watch this in action:

Your AI knowledge base quotes your documents verbatim

Getting content in takes three steps:

  1. Open Knowledge and click Upload.
  2. Select a document type and set the scope, organization-wide or a single project.
  3. Attach the file and click Upload.

The document goes up exactly as you wrote it, unchanged.

The upload dialog offers three document types plus a Let AI decide card, a scope of either the whole organization or one project, the file, and an optional description

The type and the scope are the only two decisions. Pick Let AI decide and DeltaVault picks the type for you, proposing a split for your review when the file is mixed.

So what happens after the upload? DeltaVault earns its place here. It reads the document against the conventions you defined in step one, pulls out the business entities and the attributes they carry, and quotes them verbatim rather than paraphrasing. Then it lists them on the document’s page for you to accept or reject. When your requirements keep circling back to a customer, DeltaVault notices. Everything it finds carries forward into the modeling workshop as a compact digest. Every upload gets this pass, and you never have to ask for it.

What happens when three documents define the same thing?

Upload enough documents and sooner or later three of them will describe the same concepts in three different ways. Select them, click Consolidate, and pick the output type, or let the AI decide for you. Two switches change how deep the run goes.

Optimize for business categories drafts a second document alongside the merge. That document is a business glossary grouped under your own category lanes, together with proposed business entities you can apply straight to your business model.

Extended consolidation sends the same merge to several AI vendors and puts a judge over their proposals. So a definition only one vendor thought worth keeping survives instead of quietly being dropped.

The review is the actual work

Every run ends on the same screen, and it’s the most important one in the product. The review lays out what the document consolidation merged, why, and which content went in, with the full proposed body underneath it. Read it properly. From there you can accept the document, reject it, or refine it. Refining means describing the change you want in plain language and checking the rewrite before it lands. Merged sources are archived rather than deleted, so nothing you started with is lost. When someone asks where a definition came from, the original is still there, and that’s what data governance asks of a merge. The AI assembles; you decide.

Research drafts what you can’t upload, so review it like an upload

What if the material isn’t yours to upload at all? That route carries a real caveat. You’re handing the whole first draft to the AI. Point Research at a public repository or a documentation site, and several AI vendors research it independently, each with live web search, then consolidate their findings into drafts. Every draft arrives with the web sources that vendor actually cited attached as links. We pointed it at the Willibald sample dataset from the German Data Vault User Group, where the source requirements are written in German. The drafts came back in English.

Research isn’t instant. Our run took close to five minutes. Weigh that against transcribing the material yourself, then review the output with the care you’d give any upload. Open the cited sources. Read a vendor’s unmerged document when you want to know what it actually said. Then accept, edit, or reject each draft one at a time. Accepted drafts land in the knowledge base scoped to your project, beside everything you uploaded.

A model built from your own documents arrives speaking your language. A model built from whatever fit in a chat box speaks somebody else’s. Open Knowledge, upload the three documents your team already argues over, and consolidate them. Next session is AI Data Modeling: Enough Talk. Let’s See It in Action, where we put the whole library to work in a business modeling workshop.