No training on your data

AtelyaOS doesn't train models. Anthropic's commercial API terms state customer prompts and outputs are not used to train Anthropic's foundation models.

The 30-second version

AtelyaOS does not run model training of any kind. Your briefs and proposals are sent to Anthropic via the Claude API for inference only. Anthropic's published commercial API terms commit that customer-submitted prompts and outputs are not used to train Anthropic's foundation models. Your data stays in your workspace; the LLM provider does not retain it for training.

Why this matters

For agencies handling unannounced launches, sealed M&A messaging, or politically sensitive client work in the EU and Middle East, "is my data being used to train an AI?" is the first question. The honest answer needs three parts: what AtelyaOS does with your data, what the LLM provider does with your data, and what's contractually enforced.

How it works

What AtelyaOS does

AtelyaOS is a workroom platform, not a model lab. We do not:

• Train, fine-tune, or distil any foundation model.
• Build a corpus of customer briefs or proposals for any internal use beyond running your workrooms.
• Aggregate customer content for analytics in a way that exposes prompts or outputs.

Concretely, your workroom inputs and outputs are stored in your workspace database row(s) and used only to render the workroom in your browser, run your revisions, and produce your exports. Sensitive fields are encrypted at rest (see encryption).

What Anthropic does

When an agent runs, AtelyaOS calls the Anthropic Claude API with the relevant prompt content (your goal, paste blocks, agent system prompt, Style Memory). Anthropic's published commercial API terms commit that:

• Customer prompts and outputs submitted via the Claude API are not used to train Anthropic's foundation models.
• Customer content is retained per Anthropic's published retention policy (zero-retention options are available for qualifying customers).

These commitments are between AtelyaOS and Anthropic under a paid commercial agreement. They are stronger than the terms that apply to consumer products (e.g. claude.ai free chat).

For the canonical statement, see Anthropic's usage policies and commercial terms.

What is sent to the LLM provider

When the Composer runs, the API call typically includes the agent's system prompt and Style Memory rules, the workroom goal and deliverables, pasted inputs, and prior task outputs in the workroom. That is the minimum context needed to produce a sensible draft.

What is not sent

Other workspaces' data, other workrooms in your workspace, billing details, member lists, integration credentials, and activity logs are never included in LLM calls.

Where data stays inside AtelyaOS

Your data is bounded by your workspace. We don't:

• Move data between workspaces.
• Aggregate workroom content across customers for any product feature.
• Surface one customer's content in another customer's workroom.

Workspace boundaries are enforced at every API endpoint via session and workspace-scoped queries.

Logs and observability

Production logs go to Sentry for error reporting. Sentry filters strip PII (cookies, emails, message content, brief content) before any error report is recorded. Our scrubber redacts cookies, auth headers, request bodies, extra context, contexts, breadcrumbs, and user records (only id is retained in errors). See the Sentry config in lib/sentry/filters.ts if you're auditing.

What this is not

This page describes the policy and contractual posture today. It is not a claim that AtelyaOS sets explicit per-request opt-out HTTP headers on every Anthropic call (it relies on the contractual commercial-terms-level commitment). If you have a procurement requirement for explicit per-request opt-out signalling, raise it with sales.

Common pitfalls

• Confusing API and consumer terms. "Anthropic uses my data to train Claude" can be true for free claude.ai chats but is not true for commercial API customers (which is what AtelyaOS is). Don't conflate them.
• Pasting credentials into a workroom. Goals and deliverables are stored as plaintext fields in the database (see encryption). Don't paste API keys or passwords into workroom inputs.
• Assuming your custom-agent system prompt is private from the LLM. It's not — it's literally the LLM's system instruction. Treat it as something the LLM provider sees, because it does.
• Treating this page as legal advice. It's product documentation. For a contractual data processing agreement, request the DPA from sales.

What's next

• Data handling — what's stored where.
• Encryption — at-rest encryption details.
• Compliance — GDPR / SOC 2 / ISO posture.