How Usva uses AI
Usva is a drafting tool built on language models. This page says which ones, what they see, and where they let you down.
Last updated 16 November 2026.
Which models
The default is gpt-oss-120b, an open-weights model, deployed by BSUM in Azure AI Foundry in the EU. For work where the writing matters more than the cost — an investor document, a funding application — you can choose a stronger model per project or per question. Those are priced separately and shown before you run.
Usva keeps a larger margin on the smaller models than on the larger ones, so the cheaper choice is the one it nudges you toward. The price of each option is on the page when you choose.
Where they run
In BSUM's own deployment, in the European Union. Your content is not sent to a model provider outside that deployment, and not to a consumer AI product.
What they are given
The files you selected for the project, the public sources Usva found while researching, and the instructions for the document type. A project can have a confidential folder that the search agents cannot read, which is used only when a document is being written.
What they are not used for
Your content is not used to train models. The deployments are configured not to retain it for training. BSUM does not use your projects as examples, case studies or demonstration material without asking you first.
Where it is unreliable
- Numbers. A market size or a growth rate in a draft is a starting point. Open the source and check it.
- Citations. Every claim carries a source, but a source can support a claim loosely. The link is there so you can tell.
- Recency. Public sources are found by search, and search misses things. A patent filed last month may not appear.
- Absence. Usva can tell you what the documents say. It is weaker at telling you what nobody has written down.
Nothing leaves on its own
Usva does not send email, file applications, or submit anything anywhere. It produces documents, and you decide what happens to them.