Legal

AI Data Use Addendum

Version 2026-04-18.1 · Effective April 18, 2026

DRAFT — under legal review. This addendum describes how Saltare's AI features interact with your data. It supplements the Privacy Policy and is binding on users of AI features.

Saltare — a product of Saltare Machina Corporation (1449 S Michigan Ave, STE 13683, Chicago, IL 60605, United States) — offers AI features (agent responses, hybrid search, AI-assisted autofill, summarisation, and transcription) that represent the most data-sensitive surface of the Service. This addendum describes, plainly, what data is sent where, why, and under what retention terms.

1. What leaves the tenant boundary

The following data categories are transmitted to AI subprocessors when you use the corresponding feature:

Feature Data sent Provider
Agent response to @-mention Mentioned message + tiered context (agent memories, relevant channel/document excerpts) within a bounded token budget Anthropic
Agent tool calls (fetch_url, search, GitHub, Slack, etc.) Arguments the agent supplied for the tool call + the tool's response Third-party API for the tool
Multi-agent collaboration Back-channel messages between agents in a collaboration thread Anthropic
Hybrid search Your query string; document/message/memory bodies are embedded at write time, not per query OpenAI (embeddings)
AI autofill columns in Databases Row context + column prompt Anthropic
Voice call transcription (paid plans) Call audio LiveKit + configured transcription provider

2. What we never send

  • Passwords, MFA secrets, WebAuthn credentials, API key plaintext (only digests persist)
  • Stripe card numbers (stored by Stripe, not us)
  • Content from workspaces other than the one making the request
  • Deleted content once the deletion has completed

3. Training

We do not use Customer Content to train AI models. We require the same of our upstream AI subprocessors via zero-retention / no-training contractual terms. If an upstream provider materially changes its training-data stance, we will notify affected customers in advance.

4. Retention at the provider

  • Anthropic: Zero-retention API mode. Prompt and response logs are not retained by Anthropic beyond transient processing.
  • OpenAI: No-training enterprise tier. Embeddings are generated and returned; the raw input text is not retained for model improvement.
  • LiveKit: Call audio is transmitted in real time and is not retained by LiveKit after the call ends. Transcripts we generate are stored in your workspace.

5. Moderation and safety

All messages sent to AI agents are screened by our content-moderation layer before being forwarded upstream. Content flagged as violating the AUP is blocked and may be logged for review by our team. A hashed, non-identifying user identifier accompanies each upstream call to allow abuse aggregation without exposing your account.

6. Your controls

Workspace admins may:

  • Disable AI features entirely per agent
  • Restrict which tools an agent may call
  • Revoke connector authorisations at any time
  • Export and delete agent memories
  • Cap overage spend to prevent runaway credit usage

7. Changes

We will notify workspace admins via email of any material change to the categories of data transmitted to AI subprocessors or to the providers engaged, at least 30 days before the change takes effect. The current state of this document is authoritative.

8. Contact

Questions about this addendum: [email protected] or Saltare Machina Corporation, 1449 S Michigan Ave, STE 13683, Chicago, IL 60605, United States · +1 (888) 804-8426.