AI disclosures

• Three AI features use Anthropic Claude Haiku (model family routed via Vercel AI Gateway): the support chatbot on the marketing site, the Pro-tier "Style with AI" widget customizer, and the optional site-language translator.
• Risk category under the EU AI Act: limited-risk (Article 50). All three features are identified as AI in the UI.
• Zero retention: messages aren't stored by us, by Vercel AI Gateway, or by Anthropic under our gateway configuration.
• Not used to train any model.
• No AI is used in the embed widget itself, the JSON API that serves it to customer sites, or the moderation queue. Approvals, rejections, ordering, and rendering are deterministic, non-AI code.

The chat widget at the bottom-right of the marketing site is powered by an AI language model. Conversation flow:

• You open the widget and type a question.
• Your message + a static product-information system prompt are sent to the Vercel AI Gateway endpoint.
• The Gateway routes the request to Anthropic and streams the response back through us to your browser.
• The conversation is held in your browser session only; we don't write transcripts to our database.
• Page refresh = new conversation. We never see the history again after it leaves your browser.

Provider: Anthropic, via Vercel AI Gateway. Model: Claude Haiku family. System prompt:static text describing Plauditly features, pricing, and routing rules — viewable in the repository (lib/ai/system-prompt.ts) for full transparency.

Under Article 50 of EU Regulation 2024/1689 (AI Act), users interacting with an AI system must be informed that they're interacting with one. This applies to us. So:

• The chat widget is labeled "Plauditly AI" or similar throughout the conversation — no attempt is made to imply a human is responding.
• The first message you see in the widget identifies itself as AI and names the provider (Anthropic).
• Replies may be inaccurate, incomplete, or out of date. Don't rely on the chatbot for legally, financially, medically, or contractually consequential decisions.
• You can always escalate to a human by emailing hello@plauditly.app — we read every message.

Per Article 50(2), if we ever generate AI text or media that could plausibly be mistaken for human-authored output (e.g. AI- drafted testimonial responses on customer behalf), we will mark the output in a machine-readable form. This is not currently a feature.

What goes to the AI provider: your chat message text and the static system prompt. That's it. No account email, no IP address linkage, no cookies, no inferred attributes.

Retention:

• Vercel AI Gateway operates on zero data retention — messages pass through and are not stored.
• Anthropic API has zero retention by default for the gateway-routed configuration we use.
• We store nothing on our side either — no chat logs, no transcripts.

Training: messages sent through our Vercel AI Gateway are not used to train Anthropic models, per gateway policy.

Cross-border transfer: Anthropic processes requests in US regions. Transfers from the EEA, UK, or Switzerland rely on the EU-US Data Privacy Framework (Anthropic is DPF-certified) with SCCs as a fallback.

Pro subscribers see a Style with AIbutton in the widget editor. The feature takes a short plain-English description (“more like Apple, less border”) and generates CSS the founder can apply to their embedded widget.

• Trigger: explicit click on the "Style with AI" button. Never invoked automatically.
• Inputs sent to Anthropic: the founder's description prompt, the widget's current type and accent color, and a static system prompt teaching the model the safe CSS surface it can target. No testimonial content, no author information, no account identifiers.
• Output: a small CSS block (validated server-side, dangerous selectors stripped) returned to the editor for preview. The founder must explicitly Save before the CSS is persisted to their widget.
• Provider: Anthropic, via Vercel AI Gateway. Model: Claude Haiku family.
• Retention: same as the chatbot — zero retention at the Gateway and at Anthropic; we do not log prompts or completions on our side.
• Training: not used to train any model.

In-product the section is labeled as AI-powered with a sparkle icon and the "Pro" tag, satisfying Article 50(1) disclosure where users must be told they are interacting with an AI system.

The language selector in the top navigation lets visitors view the marketing site in their preferred language. Translations are produced on demand by an AI model and cached in the visitor's browser.

• Trigger: explicit selection of a non-English language from the language picker.
• Inputs sent to Anthropic: short snippets of the marketing-page UI strings being translated (no personal data, no account information, no testimonial content).
• Output: translated strings, applied to the page in-browser and cached in localStorage so repeat visits don't re-call the model.
• Provider: Anthropic, via Vercel AI Gateway. Model: Claude Haiku family.
• Retention: zero retention at the Gateway and at Anthropic; we do not store translations server-side.
• Training: not used to train any model.
• Limitations: legal documents (Terms, Privacy, this AI disclosure, etc.) are NEVER machine-translated — the authoritative version is always the English text on this site.

Article 50(2) requires AI-generated text that could pass for human-authored to be machine-readably marked. Translations of marketing copy are explicitly framed in the UI as machine translations from English; reverting to English is one click away.

The dashboard's moderation queue (approve / hide / delete), embed.js, the iframe fallback, the public widget JSON API, the billing flows, and the free SEO tools (templates, schema generator) are deterministic, non-AI software. No AI is involved in decisions about which testimonials are shown, in what order, to which visitor.

If we ever introduce a new AI feature (for example AI-suggested replies to testimonials, AI tone analysis, AI fraud detection on submissions), we will:

• Update this page to describe the feature, the provider, and the data flow before the feature ships.
• Add an in-product disclosure where the feature is used.
• Make the feature opt-in for any processing of testimonial author personal data.
• Notify active subscribers by email 30 days before launch where the change is material.

Per our Acceptable Use Policy, customers may NOT submit AI-fabricated testimonials and pass them off as real customer feedback. This is independently required by the FTC's Consumer Reviews and Testimonials Rule (16 CFR Part 465, $51,744 / violation) in the US, and by the EU Omnibus Directive's amendments to UCPD across the EEA.

If we detect AI-generated testimonials being submitted at scale, we will remove them and may terminate the account responsible.

• You can choose not to use any AI feature — the chatbot is optional, the language translator is opt-in, and the Pro "Style with AI" button is never auto-invoked.
• You can request that we delete any logs we have (we don't keep any for the AI features described here, but the right applies to anything we do hold; see /legal/privacy §4).
• You can object to processing based on legitimate interest (all three AI features are processed under this basis) at any time. We'll honor the objection for your future sessions.

About this disclosure: privacy@plauditly.app. About a specific AI feature's output (a wrong chatbot answer, an unexpected translation, a Style-with-AI suggestion that doesn't fit): hello@plauditly.app.