Norwegian firm says companies waste tokens when querying their own documents

Norwegian company Rock Sovereign says many businesses spend more than they need to on AI because they send too much text to external language models. When a company lets an outside model answer questions about its own documents, the relevant pieces of text must be included in every single request, and the more text that is sent, the higher the bill.

Rock Sovereign is newly founded. Its solution splits content into small, natural paragraphs at the point where the data is indexed. That way only the most relevant pieces are passed on to the model with each query.

Morten Krogh-Moe, managing director of Tellusr, the company behind the software, describes the difference this way: instead of sending large chunks of text to large models, which use more tokens, they send smaller and more precise ones.

Paal Hatledal, managing director of Rock Sovereign, argues that cutting AI costs by 40 to 50 percent is easy to do. The text available here does not explain how that figure was calculated or tested.

The point matters because token use directly drives the cost of running AI on internal documents. Reducing how much text goes to the model is presented as a straightforward way for companies to lower that expense.