Trust & data practices.

A named participant, never a hidden recorder

Harlyn joins as a visible participant with a name and a tile in the grid, like any colleague. Everyone in the meeting can see it from the moment it arrives. When it wants the floor, it raises a hand and waits to be recognized. You should still follow your local rules on recording consent — Harlyn's visibility makes that conversation easy, not unnecessary.

It tells the room it's an AI

Visibility goes beyond a tile in the grid. Nobody should realize only afterwards that they were talking to software, so Harlyn identifies itself as an AI on three surfaces, in both modes, each surface reaching a kind of participant the others can't:

• The roster name. It joins as "Harlyn (AI delegate for Dana)", or "AI note-taker" when it is only there for notes, so what it is and who it represents are on screen from the first second, including for latecomers who missed every announcement.
• The meeting chat. A one-time introduction when the conversation starts: who sent it, and what it is there to do.
• Its first spoken words. The first time it takes the floor, it leads with a one-line self-introduction, which reaches dial-in participants who never see a screen. It never speaks in a meeting where it has not identified itself.

Most meeting AI stops at the display name. Harlyn treats the introduction as part of its etiquette: the room learns what it is before it speaks, the same way it asks before it acts.

Each introduction also lands in the meeting's activity record. Transparency rules for AI systems, such as Article 50 of the EU AI Act in application since August 2026, expect people to be told when they are interacting with an AI. Harlyn is designed for transparency: disclosure is built-in behavior, not a setting.

What we capture, and where it lives

• Transcript — a speaker-labeled transcript of the meeting, stored in our database so recaps, search, and cross-meeting questions work.
• AI-generated notes — the summary, action items, commitments, and messages-for-you that Harlyn produces from that transcript.
• Recordings — kept by our meeting-bot infrastructure provider (Recall.ai), not copied to our storage. When you press play, we fetch a short-lived link on demand; those links expire within hours.
• What we don't keep — no raw audio in our database, and the public site uses no cookies and no visitor tracking scripts.

What it will and won't say

As your delegate, Harlyn represents only two things: what you briefed it on and facts from the tools you connected. Everything else — opinions, judgment calls, commitments you didn't authorize — it declines and brings back to you. Your briefing separates what may be spoken from internal context Harlyn uses but never states aloud. When in doubt, it defers. Off-brief questions reach you mid-meeting; your reply is spoken back into the room, attributed to you, word for word.

Your CRM answers to your people

When outside parties are in the room, your connected tools respond to your side of the table. If an outside participant asks for something from your CRM, Harlyn asks your side for permission first — one internal "yes" unlocks it for that meeting, an internal "no" keeps it closed. An outsider's request can never unlock your data by itself.

Nothing writes without your yes

Harlyn never modifies your CRM or files a task on its own. In the room, it reads the plan back and your spoken yes runs it. After the meeting, action items wait in your dashboard until you click Approve & run. The approval is the mechanism, not a setting you can forget to enable.

Every approval leaves a trace

Each meeting keeps an activity record of the consequential moments: the plan Harlyn read back, who approved it and where — a spoken yes in the room or a click in the dashboard — what actually ran and its result, plus data-sharing permissions granted in mixed company and replies you sent into the room live. It's on the meeting page as "Activity — who approved what," so anyone reviewing later sees the full chain, not just the outcome.

Two companies, one meeting

When both sides run Harlyn, one presence serves the room but each company's data stays on its own side: your briefing, your CRM, and your recap are yours; theirs are theirs. Each side's answers come only from its own tools, and the shared record becomes trustworthy because both sides can sign off on it.

Your credentials

OAuth tokens and API keys for the tools you connect are encrypted at rest and used only to act as you, for you. Disconnecting a tool in the dashboard removes our access; you can also revoke Harlyn from the tool's own settings at any time, effective immediately.

Google user data — Limited Use disclosure

Harlyn's use of information received from Google APIs adheres to the Google API Services User Data Policy, including the Limited Use requirements. Calendar data is read only to know which meetings to join and who is attending; it is never sold, never used for advertising, and never used to train AI models.

Who processes your data

Harlyn is built on a small set of infrastructure providers, each doing one job:

Provider	What it does
Recall.ai	Puts the bot in the meeting; handles audio, video, transcription, and recordings
Anthropic	AI reasoning — answers, summaries, action-item extraction (API terms; not used to train models)
xAI	Streaming conversational voice
Deepgram	Text-to-speech fallback voice
Supabase	Database and sign-in (Postgres, hosted in the US)
Railway	Application hosting
Resend	Email delivery (recaps and digests you opt into)

Tools you choose to connect — your CRM, Slack, calendar, task trackers — receive data only when you or your briefing direct Harlyn to use them.

Deletion & questions

Ask and we delete — your meetings, transcripts, notes, and account, fully and promptly. Write to harlyn@harlyn.ai for deletion requests, security questions, or anything this page didn't answer. The full legal versions live at Privacy and Terms.