Data & AI

The first thing we ask is who will do something differently once the answer exists. Sometimes a report with the right definitions is enough, and no AI is needed. That is a good outcome, not a failed AI project.

What do you need to be able to decide?

Three common cases, three different answers.

• You know which figure you want to follow, but you can’t trust it. Then we start with reporting on a modelled source: one definition per figure, written down and approved by the people who will use it, and a refresh that happens by itself. A separate data warehouse is needed when analysis slows the live system or several sources have to be combined, but that is a choice that needs a reason.
• The answer already exists, in an instruction or an old case, but nobody can find it. Then it is search. The database you already have may manage more than you think, and that should be tested with real searches.
• You want to ask a question and get a combined answer from your documents. Then it is an assistant with RAG, where the answer is built on documents fetched first. It stands on the search, so that has to work before the language model goes on top.

How fresh the data has to be is a question per flow. Real time is a cost to justify, not a goal.

Where do we start?

Bring a real question, the material you have today and an example of an answer you would accept. That gives us more to work with than a wish for AI. We won’t promise any accuracy up front, but we can see which way the question points and what needs a closer look. Get in touch with the question, and we’ll start there.