Version: 2026-08-16.1 (provisional). How Atly uses AI with your data — product transparency page (ADR-015 / ADR-016).
1. Purpose
This page explains, in plain language, how Atly uses large language models (LLMs) and related AI pipelines with Customer data. Binding processor terms are in /legal/dpa (AI annex). Service terms: /legal/terms.
Provider: AtlyTech s.r.o., IČO 29944627, DIČ CZ29944627, registered office Dandova 2619/13, Horní Počernice, 193 00 Praha 9, registered at Městský soud v Praze, oddíl C, vložka 454920.
2. What AI is used for
Depending on settings and plan, Atly may use AI to:
- Draft replies and summarise threads.
- Parse inbox / WhatsApp / Messenger / (later) missed-call transcript text for operational suggestions.
- Polish morning briefing wording (hard counts stay exact and are taken from your data).
- Extract proposed Company DNA (after local PII anonymisation on that path).
- Later: transcribe missed work calls (STT) and produce short follow-up summaries.
- Optional: chat and job-note dictation (the account holder must switch it on) — microphone audio is transcribed to fill the chat box or a job note, then discarded; usage counting applies.
Outputs can be wrong. Always verify before sending or relying on a draft.
3. When customer content reaches an LLM
Customer content is sent to an LLM only when:
- A provider access key is configured for the deployment, and
- The company account holder enables LLM customer content, and
- The autonomy emergency stop is off.
Otherwise draft/parse/briefing/chat fall back to local sample paths where implemented.
4. Providers and no training on your data
- Primary inference: Together AI. Atly selects models for your plan.
- Optional: OpenAI only when you connect your own OpenAI API key in product settings (where available). Those calls use store=false where the API supports it.
- Atly does not use Customer content to train Atly-owned foundation models.
- Tool outputs passed back into models are redacted/truncated where implemented.
5. Human approval and disclosure
- Outbound email, WhatsApp, Messenger, and SMS require human approval; they do not auto-execute. SMS is only to a known contact or a missed-call caller, and only where the company number can send SMS.
- Assistant suggestions in chat are labelled as AI-generated.
- Outbound mail includes an AI disclosure footer and X-Atly-AI-* metadata where implemented.
- The approving user remains responsible for the final customer-facing send.
6. Channels beyond email
WhatsApp is the next channel after email/calendar (Twilio or Telnyx company number; human approval before send). Messenger is a sibling Meta channel on a company Facebook Page (Twilio only; 24-hour window; not a personal profile). SMS on that number is only where local rules and the plan allow it, and only to a known contact or a missed-call caller — not a general SMS inbox, not a campaign, and not the owner's personal phone. Missed work-call capture can play a greeting and, if you turn recording on, transcribe a left message; a short approved template or SMS can follow where the number can send SMS. Voice audio, when used, is ephemeral for transcription and is deleted after success; usage counting and FUP may apply. Optional chat dictation (product settings → Automation & AI) sends short microphone clips for speech-to-text only; audio is not stored; the transcript stays in the chat compose box or a job note until you send or save. Team/Trial remain email-oriented until channels are enabled.
7. Your controls
- Product settings: switching LLM use on or off, switching chat and job-note dictation on or off, emergency stop, autonomy level, disconnect integrations, DNA delete.
- Privacy panel: DPA/Terms acceptance and re-acceptance when versions change.
- Request broader deletion / account closure via the account holder's support path.
8. Related
/legal/dpa · /legal/privacy · /legal/terms · product Security documentation