Most AI assistants at work follow the same path. You type a question, or a meeting gets recorded, and the text or audio goes to a commercial AI provider's API. The answer comes back. Your data has left the company.
For many teams that is fine. For a founder or CTO who has to answer "where does our meeting audio go?", it is not.
What self-hosted AI means
Self-hosted means the models run on hardware you or your vendor control, not on a third party's API. Two things have to be true:
- Speech-to-text runs on your own machines, so meeting audio never leaves them.
- The language model runs there too, so prompts, messages and transcripts never reach an outside API.
If either piece calls out to a commercial provider, the data has left.
How AntFabric does it
AntFabric records and transcribes meetings with Whisper, an open speech-to-text model, running on hardware we run. The assistant answers with an open-weights language model on the same hardware. No commercial model provider is configured, so no transcript, message or prompt reaches a third-party API.
This is a fact about how AntFabric is built, not a promise about a specific deployment. On the demo we show you exactly where the audio goes and which model answers.
What it costs you
Open models are behind the largest commercial ones on some tasks. We think that trade is worth it for work data, and we would rather say so than pretend otherwise. An assistant that drafts a reply from your own chat, calendar and mail does not need the biggest model in the world. It needs your context and a clear boundary.
See it on your own work. Book a demo.