AI and data

AI can be wrong

Replies, summaries and reminders written by your AI assistant can be inaccurate, incomplete or out of date. Check anything important before you act on it. Do not let an assistant make decisions with legal, medical, money or similar serious effects without a person reviewing them.

Which model is used

Relay itself uses no AI model. Each workspace connects its own assistant, for example Grok Bot, Claude, Cursor, or another assistant that supports MCP (the Model Context Protocol). That assistant, and the model behind it, is chosen and paid for by your workspace, not by us.

How your assistant works with Relay

1. Relay stores messages from channels it is invited to and direct messages sent to it.
2. When someone mentions Relay, sends it a direct message, or uses a wake word your workspace set up, Relay sends your assistant a short notice. The notice has only IDs and labels: workspace, channel, message ID, time, thread, why Relay woke the assistant (a mention, a direct message or a wake word) and, for direct messages, that it came from a direct message. It does not contain the message text.
3. Once a day, Relay sends your assistant a notice about messages that did not wake it, again with IDs only.
4. Your assistant then reads the messages it needs and posts replies or reminders through Relay's MCP tools. Its access token only works for your workspace.

What your assistant can read

• Stored messages from channels Relay was invited to, and direct messages sent to Relay, for as long as Relay keeps them.
• The list of channels Relay is in, and the daily summary.
• When it sets a reminder, a person's name and time zone from Slack, looked up at that moment and not stored by Relay.

Your assistant cannot see channels Relay was not invited to, or direct messages between other people. Anything your assistant reads is handled by its provider under that provider's own terms and privacy policy.

What Relay does itself, without AI

• Relay answers some direct messages with fixed text: help when someone asks for it, and a note when no assistant is connected, Relay could not reach it, or it has not answered yet.
• When your assistant answers a message, Relay checks where the reply is going. It blocks a post that would move content from a direct message, a private channel or a Slack Connect channel to people who should not see it, and records the block in its history. This check only works when your assistant says which message it is answering, and it cannot catch text your assistant copied from somewhere else. When your assistant does not say, the post goes out, and Relay records which assistant posted where (no message text).
• The Home tab in Slack shows each person the channels they share with Relay and this notice: AI can be wrong.

How long data is kept, and how it is used

Relay keeps messages, replies and daily summaries for 1, 3 or 7 days (7 by default), then deletes them (a daily cleanup job does this, so an item can last up to one extra day). A workspace admin or owner can make the period shorter on the Relay settings page. When Relay is removed from Slack, they are deleted right away, and everything else Relay kept for the workspace is deleted within 14 days (see the privacy policy). Relay uses them only so your assistant can read and reply in your Slack. Relay never uses Slack data to train an AI model, and does not let anyone else do so. How long your own assistant keeps what it reads depends on its provider.

Data tenancy

Relay keeps each workspace's data separate. In the database, every request runs as a role that can only see rows for the one workspace it is working on. Each workspace's messages are encrypted with that workspace's own key (AES-256-GCM). An assistant's token is tied to one workspace, so it cannot read any other workspace's messages. Relay has no shared model and no shared AI memory between workspaces.

Data residency

Relay's app and database both run in the United States: Vercel iad1 and Neon iad1, both in Washington, D.C. Relay offers no other region today. Where your assistant processes data depends on the assistant and provider your workspace chose.

The demo assistant (for testing)

For people trying Relay without an assistant of their own, we plan to offer a demo assistant run by us. [PENDING: details from Temper once the demo assistant is built: whether it uses any AI model, what it reads, what it keeps]

More

Full details are in the privacy policy. Questions? See support.