Dialogis
Legal Home

AI Transparency Notice

How Dialogis AI Minds work, how they disclose that they are artificial intelligence, which models power them, what they cannot do, and how humans stay in control. Published to meet the transparency obligations of the EU Artificial Intelligence Act (Regulation (EU) 2024/1689) and to help our customers meet theirs.

Version 2026-10-10 Effective 10 October 2026 Replaces version of 2 September 2026
  1. 1. What an AI Mind is
  2. 2. Who is responsible
  3. 3. How we tell you that you are talking to an AI
  4. 4. What models and services are used
  5. 5. Capabilities and limitations
  6. 6. Human oversight
  7. 7. Prohibited uses
  8. 8. Logging and records
  9. 9. AI literacy
  10. 10. Reporting problems and incidents
  11. 11. Changes

1. What an AI Mind is

An AI Mind is a conversational AI assistant built on the Dialogis platform. A customer (an expert or a business, the "Operator") uploads its own material (documents, recordings, talks, web pages, posts) and configures a persona. The Mind then answers End Users' questions in text and, where enabled, voice, using that material. It is offered on the Operator's website as a chat widget, on a hosted page, and on Telegram, WhatsApp and Instagram.

A Mind is an AI system intended to interact directly with natural persons. It is not a general-purpose chatbot: it is designed to answer from the Operator's own material and to say when it does not know. Its intended purpose is to inform. It is not designed, tested or offered for any of the high-risk uses listed in Annex III of the AI Act, and the Acceptable Use Policy prohibits them.

2. Who is responsible

  • Dialogis.ai s.r.o. develops and provides the platform and is the provider of the AI Mind system within the meaning of the AI Act.
  • The Operator configures and publishes a specific Mind and is its deployer. The Operator is also the controller of End Users' personal data under the GDPR.

Both roles carry obligations, which are set out in the Terms of Use (Sections 6 to 8), the Acceptable Use Policy and the Data Processing Agreement.

3. How we tell you that you are talking to an AI

Article 50(1) of the AI Act requires that people are informed that they are interacting with an AI system. We build that disclosure into every Mind so that Operators cannot omit it:

  • On the web (widget and hosted pages), a screen before the first message states that you are interacting with an AI assistant whose answers may be inaccurate, explains that messages are stored and processed, and links to the End-User Terms and the Privacy Policy. You must agree before chatting.
  • On Telegram, WhatsApp and Instagram, the Mind sends a notice at the start of a conversation stating that you are chatting with the expert's AI assistant and linking the same documents. The notice is sent once, at the start of a conversation on that channel; it links the documents, whose current version and effective date are always shown at the top of each document.
  • Voice answers are generated synthetically and are disclosed as AI-generated in the same notices. Where a Mind uses a voice modelled on the expert (a Voice Clone), the expert has consented, and the voice is still disclosed as synthetic.
  • Voice calls (real-time conversations, currently started from Telegram or the web) open with a spoken statement that you are talking to an AI assistant; on the web, a notice before the call connects says so as well. Calls are limited to 10 minutes and end with a transcript you can read.
  • The Mind is instructed to confirm that it is an AI whenever asked, and Operators are contractually prohibited from configuring a Mind to claim to be human, to deny being an AI, or to hide these disclosures.
  • If a human joins a conversation (human handoff), the chat says so.

Marking of AI-generated output (Article 50(2)). In every interface we control, AI-generated messages and voice are visibly labelled as generated by an AI assistant. In machine-readable form, every answer the platform produces is marked as generated by the assistant in the structured data of our chat interface and API (each message carries the role of its author, so software reading a conversation can tell AI output from human input; the chat widget's message and streaming endpoints return an X-AI-Generated header, and every message bubble in the widget, on hosted pages and in the Telegram voice mini-app carries a data-ai-generated attribute), and transcripts of voice calls mark the AI's turns the same way. On messaging channels (Telegram, WhatsApp, Instagram) there is no metadata channel we control, so the disclosure is the text notice at the start of the conversation and the Mind's own statement that it is an AI. The Commission's Code of Practice on transparency of AI-generated content (final version of 10 June 2026, assessed as adequate by the Commission and the AI Board on 8 and 9 July 2026) and its Guidelines on Article 50 (20 July 2026) describe the state of the art as an imperceptible watermark for free-form text and, for audio, digitally signed metadata combined with a watermark, each with a detection mechanism. Our text answers and our synthetic voice do not yet carry such watermarks: the language and voice models we use come from third-party providers, and interoperable watermarking of their outputs is not yet available to us. Until it is, the disclosure for voice is the spoken statement at the start of every call and the labelling of voice answers in the chat. The Platform was placed on the market before 2 August 2026, so under Article 111(4) of the AI Act, as amended by Regulation (EU) 2026/1744, our marking must meet Article 50(2) by 2 December 2026. We are implementing watermarking and detection for text and voice in line with the Code's measures by that date, and we will update this notice when each step is in place.

4. What models and services are used

A Mind combines several components. As of the date of this notice:

  • Retrieval: the Operator's material is transcribed, read, OCR-processed, summarised and converted into vector embeddings (Google Vertex AI) stored in our EU-hosted database. For each question, the most relevant passages are retrieved.
  • Answer review: where the Operator has turned it on, a Gemini model on Google Vertex AI classifies whether a question touches a sensitive topic. If it does, the Mind's answer is held for the Operator to approve, edit or discard (Section 6). The check fails closed: if the classifier cannot answer, the reply is held.
  • Bookings: where the Operator has connected a calendar service (Cal.com), the Mind can offer meeting times and create a booking with the details the End User gives.
  • Answer generation: large language models from OpenAI (GPT family) and, where configured, Google (Gemini, via Vertex AI or the Gemini API) generate the answer from the retrieved passages, the persona instructions and the recent conversation.
  • Speech: ElevenLabs provides speech-to-text, text-to-speech, real-time voice calls and Voice Clones.
  • Document processing: Mistral AI provides OCR; Apify and Jina AI fetch web pages and public posts on the Operator's instruction.

These are general-purpose AI models developed by third parties; Dialogis does not train its own foundation models. The current list of providers and their locations is in the Sub-processor List. We may change providers or models where the replacement offers comparable capability and data protection.

Neither we nor our providers use Operators' material or End Users' conversations to train general-purpose models.

5. Capabilities and limitations

A Mind can answer questions, explain, summarise and point to the Operator's material, hold a conversation across several turns, talk in a voice call, collect a contact request when the End User agrees, offer and book a meeting where the Operator has connected a calendar, and hand the conversation to a human.

A Mind cannot guarantee correct answers. Known limitations include:

  • Errors and invention. Language models can produce confident answers that are wrong, incomplete or invented, even when relevant material exists. Retrieval reduces but does not eliminate this.
  • Out-of-date knowledge. A Mind knows only what the Operator uploaded, as of the time it was uploaded.
  • Not the person. A Mind is not the expert and cannot know the expert's current views, availability or personal circumstances. Its answers are not statements by the expert at the time of the conversation.
  • No professional advice. Answers are general information, not legal, medical, financial or other professional advice.
  • Bias. Models may reflect biases present in their training data or in the Operator's material.
  • Manipulation. Users may try to make a Mind behave outside its instructions ("prompt injection"). We apply safeguards, but they are not perfect.
  • Language and voice quality vary by language; transcription of accents, names and noisy audio may be inaccurate.

6. Human oversight

  • Operators must review a Mind before publishing it and monitor its conversations afterwards. The admin platform shows every conversation and lets Operators correct the Mind by adjusting its material and settings.
  • Every Mind can hand a conversation to a human, and Operators must give End Users a way to reach a person.
  • Operators can switch on answer review: questions on sensitive topics are detected automatically and the Mind's answer is held, for up to 24 hours, until the Operator approves, edits or discards it. The End User is told that the answer is being checked. Held answers that are not reviewed in time are discarded, not sent.
  • A Mind provides information and assistance only. It must not be used to make or materially influence decisions with legal or similarly significant effects on people (for example about credit, employment, education, housing, healthcare or public benefits) without meaningful human review, and deployment in the high-risk contexts of Annex III of the AI Act is prohibited by the Acceptable Use Policy; an Operator who does it anyway becomes the provider of a high-risk AI system under Article 25(1)(c) of the AI Act.
  • Dialogis staff read conversations only as the Operator's processor (when the Operator asks for help improving its Mind, or for support), or to investigate abuse, fraud and security incidents, always under confidentiality. Platform-wide improvement uses aggregated or anonymised data, never the content of conversations.

7. Prohibited uses

The practices prohibited by Article 5 of the AI Act (manipulative or exploitative techniques, social scoring, predictive policing based on profiling, untargeted facial scraping, emotion recognition at work or in education, sensitive biometric categorisation) are never permitted on the platform. See Section 3 of the Acceptable Use Policy.

8. Logging and records

Conversations, voice-call transcripts and booking records are stored so that Operators can review them, meet their record-keeping duties and respond to complaints. Technical logs are kept for 30 days. Consent records (which version of the terms an End User accepted, and any consent to be contacted) are stored with the conversation. Retention periods are in Section 9 of the Privacy Policy.

9. AI literacy

Article 4 of the AI Act, as amended by Regulation (EU) 2026/1744, requires providers and deployers to take measures that support the development of AI literacy among their staff and the people who operate AI systems on their behalf. Dialogis trains its own staff. For Operators we provide this notice, the product documentation, guidance in the admin platform on testing a Mind before publishing it, and the Acceptable Use Policy. Operators must ensure that the people who configure and supervise their Minds understand these materials.

10. Reporting problems and incidents

If a Mind produces harmful, illegal or seriously misleading output, or if you suspect misuse, please report it to the Operator or to hello@dialogislabs.com. We record AI-related incidents, investigate them, and, where the law requires it, report serious incidents to the competent authorities. Illegal content can be reported as described in Section 19 of the Terms of Use. You also have the right, under Article 85 of the AI Act, to lodge a complaint with a market surveillance authority; in the Czech Republic the Government has designated the Czech Telecommunication Office (Český telekomunikační úřad, www.ctu.gov.cz) as the main market surveillance authority for AI systems, pending the Czech act that implements the AI Act.

11. Changes

We update this notice when models, providers, disclosure mechanisms or the law change. The version and effective date are shown at the top of the page.

Dialogis.ai s.r.o. · ID No. 23985348 · Sokolská 1883/8, 120 00 Praha 2, Czech Republic · hello@dialogislabs.com