Record less. Keep what matters.
The whole idea in one sentence: today's tools record everything and clean up later; Cull decides first, and only writes down what everyone agreed to. Not a notetaker, not a dashboard, but the layer that stands between what a machine perceives and what it is permitted to keep, say, and do.
- Everything is saved the moment it is captured
- Every voice is transcribed, consenting or not
- Voiceprints are kept indefinitely
- Deletion happens afterward, if ever
- The original recording still exists somewhere
- When something goes wrong, the system keeps more
- Sound and video live only in a seconds-long buffer
- Each person's consent is checked before anything is saved
- Non-consenting speech is discarded, never transcribed
- No voiceprint without separate written permission
- The unapproved original is never written at all
- When something goes wrong, the system keeps less, by design
How it works, in five steps:
Announce: everyone hears that capture may begin, including late joiners and dial-ins. Ask: one click or a spoken yes, and silence never counts as yes. Decide: each speaker's words kept or discarded, person by person. Obey: anyone can say “Cull.ai, stop,” even a guest with no account. Forget: voice data destroyed when the meeting ends. Prove: a one-click report of exactly what happened.
The compliance report: who was told, who agreed, what was kept, what was destroyed, all verifiable, none of it editable after the fact. A general counsel's first question is “how do I prove this?” Our answer is a document they can hand to a regulator, not a promise on a policy page.
Kill the process mid-meeting: zero recoverable audio or transcript of any non-consented speaker. That is the patent choke point, running in front of you. Watch it →
Their business runs on holding data; our promise is that the vendor holds nothing, inside the customer's walls, even offline, even on a robot. They optimize the cloud API; our system works with the network cable pulled. No large lab wants recording-law liability in fifty states. And every improvement in their models makes our layer cheaper and sharper: foundation models are our input, not our rival.