Sayli runs in rooms where the conversation is the work. We built it the way we would want a tool used in our own calls. This page is precise on purpose. It says what we do and what we do not, with no weasel words.
What we record
We record the transcript of your call and the brief we write from it. Both are scoped to your organization and to your visibility settings. That is what makes Sayli useful, and it is all we keep to do the job.
What we never do
We never send a bot into your meeting. Sayli captures on your own machine, so the host sees no extra participant and no recording banner. That holds on every platform.
We do not claim more than that. Keeping the Sayli window out of a screen capture you start yourself is the operating system's decision, and it answers differently on Windows, macOS, and Linux. What each platform delivers states each case exactly.
We never train our own models on your meetings, and we never sell your data.
The speech-to-text and language-model work runs on third-party processors. We will not warrant here what every one of their default terms allows, because that is their document and not ours. What each one's terms permit, and which model-improvement settings we have taken on each, goes to customers and prospects under a mutual NDA alongside the named list. Ask at privacy@sayli.ai.
Redaction at ingest
On transcript text, before it is stored, shown, or read by the language model, Sayli masks the sensitive parts. The model sees the conversation, not the card number somebody read out loud in it.
It runs on the transcript, and only on the transcript. A document you upload to the Knowledge Base, and a file you attach to a meeting, are stored and sent to the model as you provided them, with nothing masked. So do not put a spreadsheet of customer card numbers into the Knowledge Base and expect Sayli to strip it. The call audio itself also reaches the speech-to-text processor before any masking happens, because masking is what we do to the text that comes back.
Where your data lives
Your data runs on cloud infrastructure in the United States. Sayli relies on a small set of processors: speech-to-text, language-model inference, embeddings, vector search, web search, cloud infrastructure, email delivery, billing, analytics, and error reporting. Each one sees only what its job requires.
Web search is worth calling out, because it is the one that can carry words from your call off to a third party. When an answer needs something current, Sayli sends a search query, and on the live path that query is built from what was just said. It is not the transcript, but it is derived from it. There is no switch to turn it off for one workspace today. If that matters for your rollout, write to privacy@sayli.ai before you start.
The named list, with what each one sees, goes to customers under a mutual NDA. Ask at privacy@sayli.ai, or as part of a signed DPA. The full security posture, including what we do not have yet, is at Security overview.
Recordings
Sayli can keep a recording so you can hear any moment again. The recording is tied to your account. It is never used to train models, and you can delete it.
Retention and deletion
You can delete a meeting. The delete cascades, so the transcript, the brief, and everything derived from them go with it. When data is gone, it is gone.
Nothing expires unless you choose it. There is no history window on any plan, free included, so a call you took a year ago is still readable and searchable unless you delete it. An Owner or Admin can set a transcript retention window for the whole workspace in Settings, under Workspace, and Sayli then removes transcripts older than that.
You can take it all with you at any time. Settings, then Privacy, then Download, and Sayli builds one JSON file with your profile, your workspaces, your meetings with their briefs, transcripts and action items, and your questions. Audio recordings are not part of that file.
Tenant isolation
Every query is scoped to your organization by the data model itself, not by convention. There is no code path for one organization to read another's data.
Where HIPAA stands
Sayli is not HIPAA-compliant today, and we will not pretend otherwise. No business associate agreement exists with us or with any of our vendors, and there is no HIPAA mode you can switch on. The code has a per-organization flag reserved for one, but it points at a compliant inference endpoint that no environment configures, so it is scaffolding for work that has not been done rather than a setting that protects anything. A dedicated compliant inference plane is on the enterprise roadmap. If you handle PHI, talk to us before using Sayli: enterprise@sayli.ai.
Sign-in and SSO
Sign in with Google, with Apple, or with a six-digit code emailed to you. There is no password on a Sayli account, so we store no credential that could be leaked or reused. SAML single sign-on and SCIM provisioning are on the enterprise roadmap.