EU AI Act · Art. 50
AI Transparency Statement
In one minute
- Every Kosmly agent is an AI system. We are its provider; you, the merchant, are its deployer.
- All our agents are limited or minimal risk. We do not build agents for hiring, credit scoring or other high-risk uses.
- Everything an agent creates carries a machine-readable AI label, and a person can always review, change or undo it. A visible AI label for your visitors is optional and off by default; when you turn it on, it always states truthfully what the AI did.
Kosmly is based in the United Kingdom and offers its AI systems to businesses in the European Union, so we follow the EU AI Act (Regulation 2024/1689) as a provider, together with UK data protection law.
1. Our agents and their risk level
| Agent | What it does | AI Act risk | Transparency measure |
|---|---|---|---|
| Vega · Blog | Writes articles and product stories | Minimal | AI-generated label in page metadata; "reviewed by a human" flag when you approve it |
| Sirius · SEO | Meta titles, descriptions, keywords | Minimal | Changes logged in Mission Control |
| Lyra · Translations | Translates products, pages and checkout | Minimal | AI-translation label in metadata |
| Orion · Pricing | Suggests prices within the limits you set | Minimal | Never personalises prices per person; discounts always show the lowest price of the last 30 days |
| Nova · Images (beta) | Writes alt texts for product images | Minimal | AI marker in metadata; C2PA content credentials once Nova starts editing images |
| Stella · Reviews (beta) | Asks for and answers real reviews | Minimal | Never writes or publishes fake reviews |
| Pulsar · Social (beta) | Writes social posts from your own articles and products | Minimal | Publishes only what you approve, from your own account; no contests or fake urgency |
| Comet · Email (beta) | Writes newsletters for your own lists | Minimal | Sent by your email service with its consent and unsubscribe; Kosmly never sees contacts |
| Halley · Abandoned carts (beta) | Counts abandoned carts and writes reminder emails | Minimal | Totals only; at most 3 reminders, a discount only from your own code, capped at 20% |
| Luna · Customer chat (beta) | Answers shop visitors from your site's information | Limited (transparency, Art. 50) | Always says it is an AI; hands refunds, complaints and orders to a person; chats deleted after 30 days |
| Rigel · Ads (beta) | Reads ad results, pauses losing ads, suggests ad texts | Minimal | Never raises budgets or creates campaigns; every pause can be undone |
Every agent follows the same rules. Atlas (stock) and Polaris (weekly report) are in beta too: they only read data and write reports.
2. Human oversight
Each agent works on approval by default: it prepares, you approve. You can switch any agent to Automatic: low and medium risk work then runs on its own within the hard limits you set (such as a minimum margin), while high-risk changes still wait for your ok, and an automatic brake stops the agent if something looks wrong. Every action is recorded in Mission Control and can be undone in one click.
3. Models we use
Kosmly uses general-purpose AI models. Anthropic Claude writes: served through Amazon Bedrock in Frankfurt or Google Vertex AI in an EU region. A model from a different company then checks each text against the facts of your store: Google Gemini on Vertex AI in the EU, Mistral Large in France or an open model on Scaleway in Paris. A GPT model from OpenAI on Microsoft Azure in Paris edits finished texts so they read naturally. Pictures come from Google Gemini and Imagen models on Vertex AI in the EU. Calls may pass through Requesty's EU router in Frankfurt. All of them run in EU data centres under contracts that forbid training on your data and limit retention. Names, emails and phone numbers are masked before text reaches a model whenever the task allows it. We use the same top model for every plan and every task; a backup model only steps in if the primary is unavailable, and Mission Control shows when it did.
4. What we will not do
- Use manipulative techniques or dark patterns, or exploit people's vulnerabilities (AI Act Art. 5).
- Profile individual shoppers to charge them different prices.
- Create fake reviews, fake scarcity or fake urgency.
- Use your data or your customers' data to train AI models.
5. Traceability and testing
Every proposal records which model and which prompt version produced it, so we can explain and trace each change. Before a new prompt or model goes live, it must pass our automatic quality and safety checks (no invented links or facts, prices and numbers kept, rules of your brand respected).
6. Tell us if something is wrong
If an agent produced something wrong, biased or harmful, write to ai@kosmly.app. We review every report and publish a summary of what we changed.
7. Our web readers
Kosmly reads web pages only for the shops that use it, and says who it is. KosmlyBot reads the public price of competitor product pages a merchant chose, at most once a day per page, and only if that site's robots.txt allows it. KosmlyAudit checks a merchant's own site for search and AI visibility. KosmlyPreview reads one page when someone asks for a free preview. None of them logs in, fills a cart or collects personal data. To keep KosmlyBot out, add User-agent: KosmlyBot and Disallow: / to your robots.txt; questions to ai@kosmly.app.