AI System Card
Last updated · June 2026A plain-language description of the AI in Atlas: what it does, the models behind it, the data it uses, and its limits. Published in the spirit of the EU AI Act transparency duties.
01What Atlas AI does
Atlas runs an AI assistant named Alexia, plus specialized division agents, that help brand operators monitor and optimise their e-commerce across marketing, growth, operations, finance, product, sales, creative, and intelligence. The agents read your connected data and workspace context to surface insights, draft work, and propose actions.
02Models used
Anthropic Claude (Opus, Sonnet, and Haiku class) as the primary models, OpenAI GPT as a fallback, and DeepSeek for low-cost drafts. Creative images and video are generated with image and video models run through fal. We may change models to improve quality, cost, or reliability.
03Data used
Your own e-commerce data (for example Shopify, Klaviyo, and ad platforms) and your own workspace context (campaigns, briefs, decisions). We use this only to serve your requests. We never use your data to train any model.
04What Atlas AI does not do
It does not make decisions with legal or similarly significant effect on your customers by automated means. It does not execute a send, publish, or spend without a human on your team approving it. It does not profile individual end-customers for discriminatory purposes, and it does not generate content impersonating a real person.
05Risks and mitigations
AI can be wrong (hallucination), so a human approves every action that has an external effect. Third-party content could attempt prompt injection, so inputs are sanitised and tools are allowlisted. Cost can run away, so per-tenant budgets and per-run ceilings bound it. AI-generated creative is labelled as such.
06Auditability and updates
Every agent invocation is logged immutably and every model call is traced. We communicate material changes to how the AI works on this page. Questions: atlas@arthea.ai.