Artificial intelligence transparency notice

Contents

• 1. Why this notice
• 2. What artificial intelligence does in KISMO
• 3. Classification under the AI Act
• 4. Transparency: knowing when an AI is involved
• 5. Models and providers used
• 6. Known limits
• 7. Human oversight and the controls available to you
• 8. Data and confidentiality
• 9. Regulated and prohibited uses
• 10. Artificial intelligence governance at KISMO
• 11. Contact and reporting

1. Why this notice

1.1. KISMO is an AI conversation assistant: the KISMO service (the “Service”) uses artificial intelligence systems to transcribe conversations, produce structured memories from them, answer your questions from your memories and organize your knowledge. This notice explains, in plain language, what artificial intelligence does in KISMO, which models are used, what their limits are and which safeguards KISMO puts in place.

1.2. It meets the transparency obligations of Regulation (EU) 2024/1689 of 13 June 2024 laying down harmonised rules on artificial intelligence, as amended by Regulation (EU) 2026/1744 (the “AI Act”), in particular its articles 13 and 50, as well as the information obligations of articles 13 and 14 of the GDPR. It supplements KISMO’s Terms of sale and subscription, its Terms of use and its Privacy policy.

1.3. KISMO, a French simplified joint-stock company with its registered office at 21 chemin du Moulin, 84530 Villelaure, France (Avignon Trade Register 100 952 829), is the provider of the Service within the meaning of the AI Act. The Client, which uses the Service in the course of its business, is its deployer. Contact: hello@kismo.ai; data protection: dpo@kismo.ai.

2. What artificial intelligence does in KISMO

2.1. Capture. You create a memory by calling your KISMO Number, recording from the application, inviting the Meeting bot to a video meeting, uploading an audio or video file or a document, pasting a text or giving the address of a web page, by connecting your third-party accounts, which the Service collects, or your Plaud account, whose recordings you import, or by sending it on WhatsApp to the Assistant, which turns it into a memory at your request. No dedicated hardware is needed.

2.2. Transcription. A speech recognition model turns the audio into text, punctuates it, cleans it up and separates the speakers. The language is detected automatically (French, Canadian French, English and other supported languages) and can be corrected by you.

2.3. Analysis and writing. Language models analyze the transcript or the document, determine the type of memory (meeting, interview, call, complaint, memo, list, dictation, document, summary, etc.), write a title, a summary, the structure of the memory (topics, decisions, actions) and a mind map, identify the people who take part or are mentioned, and send you the memory by email.

2.4. Assistant. The Assistant answers your questions from your memories and, if you ask, from your transcripts and the web, citing its sources; it can create and edit memories at your request; it can also consult the third-party accounts you have connected and act on them at your request, and you can write to it on WhatsApp, from the WhatsApp you have linked to your account. The voice assistant, when it is turned on for your Space, answers you by voice during a call.

2.5. Brain. The Brain distills your memories into concepts and maps linked to one another. Nothing goes in without your validation, unless you turn on automatic distillation; an invalidated concept is kept with its reason and is no longer presented as true.

2.6. Address book. The Service matches the people identified in your memories with contact cards (identity, contact details, role, relationships), suggests merging duplicates and lets you validate, correct or delete. Only natural persons get a card; no voiceprint is created.

2.7. AI assistant connector. You can authorize a third-party assistant of your choice to read your memories or create them, with detailed permissions that you can revoke.

3. Classification under the AI Act

3.1. The Service is an artificial intelligence system that integrates general-purpose AI models provided by third parties. KISMO does not train its own foundation models; it designs the instructions, processing chains, controls and interfaces that turn these models into a service for conversation memories and knowledge.

3.2. The Service is not intended for a high-risk use within the meaning of Annex III of the AI Act. In particular, it is not designed to recruit, select, evaluate, score, rank or monitor people, or to take decisions about them. When it is used to write up an interview, it performs a preparatory task (transcribing and structuring what was said) without evaluating the person or building a profile: the decision and the judgment remain human. KISMO has documented this assessment in accordance with article 6 of the AI Act, reviews it at each significant change to the Service and carries out the registration provided for in article 49 in the EU database, whose deadline is set at 2 December 2027 for Annex III systems that their provider considers not to be high-risk. KISMO voluntarily applies to the functions of the Service requirements inspired by those for high-risk systems (risk management, logging, human oversight, technical documentation, post-market monitoring), so that it can fully comply with them if a high-risk use were ever offered.

3.3. The Service uses none of the practices prohibited by article 5 of the AI Act: no emotion recognition from biometric data, no social scoring, no biometric identification, no manipulation. The “perceived sentiment” function of complaint memories is an analysis of the transcribed text intended for relations with your own customers; it must not be applied to employees or candidates and does not rely on any biometric data.

4. Transparency: knowing when an AI is involved

4.1. Interactions. The written Assistant introduces itself in the interface under the name KISMO, and reminds you, below the area where you write to it, that it can make mistakes and that important information should be checked. The voice assistant and the telephone greeting of the KISMO Number introduce themselves: the greeting says “This is Kismo. Go ahead, I’m listening.” in the Member’s language, and the voice assistant, which speaks only French, says “Ici Kismo. Allez-y, je vous écoute.” The Meeting bot appears under the name KISMO, followed by the first name of the Member who invited it through a meeting link, and displays, while recording, an image with the KISMO logo marked “REC”; the Website page dedicated to the Meeting bot (https://kismo.ai/en/meeting-bot) explains how it works.

4.2. Generated content. Memories, summaries, maps and answers are generated by artificial intelligence. Sharing pages carry the name and link of KISMO. The Service does not generate images, videos or voices imitating real people.

4.3. Results and sources. The Assistant’s answers cite the memories and passages they rely on; a partial or uncertain result is flagged as such. Identified people carry a reliability score.

5. Models and providers used

KISMO uses the following providers. The authoritative list is Annex 2 of the Terms of Sale; any change to it is announced to Clients under the conditions of section 23.3 of the Terms of Sale.

Your content is not used to train these models: KISMO ensures this by contract and by configuration (improvement programs turned off, European regions where they exist, zero retention or retention limited to abuse monitoring). The behavior of the Service (style, length, accuracy) may vary depending on the model used, especially during a failover.

6. Known limits

The Service produces useful but imperfect results. The following limits are known and should be kept in mind:

• Transcription: quality depends on the audio (noise, distance from the microphone, overlapping speech, accents, technical vocabulary, proper nouns); language detection can be wrong with a strong accent; you can correct the language and enrich your personal vocabulary.
• Attribution of statements: the Service may attribute a statement to the wrong speaker or confuse two people with similar names; you can replace, swap or add people in the memory.
• Summaries and structures: a summary may leave out a nuance, generalize, or present as a decision what was only a hypothesis; a memory may come out longer or shorter than the material justifies.
• Assistant answers: like any language model, the Assistant may produce a plausible but inaccurate answer, especially outside your memories; check the sources cited before acting.
• Address book: matching people and extracting roles or relationships are suggestions; spelling variants from the transcription can create duplicates.
• Perceived sentiment: the tone analysis of a complaint is a rough estimate based on the text, with no value as an assessment of a person.
• Bias: the models used may reflect biases present in their training data (languages, registers, genders); KISMO monitors the quality of the results and corrects the instructions given to the models when a bias is found.
• Languages: the Service is optimized for French, Canadian French and English; other languages are supported with variable quality.

7. Human oversight and the controls available to you

7.1. The Service is designed so that a person stays in control: every memory can be reread, edited and corrected; the Assistant cites its sources; you validate what goes into the Brain by starting its distillation, with the Distill button or by asking the Assistant or a connected AI, and you correct it there at any time; merges of contact cards are suggested, never imposed, and the cards KISMO creates can be corrected or deleted; recordings can be stopped and the Meeting bot removed; automatic distillation is an option you turn on, and the voice assistant an option KISMO turns on for your Space.

7.2. You can report any problematic result (serious error, inappropriate content, unexpected behavior) with the “Send feedback” button in the Service or at hello@kismo.ai. KISMO analyzes the reports, fixes the defects and informs the Clients concerned of significant incidents.

7.3. As a professional deployer, you must use the Service in accordance with this notice, ensure human review of results used for decisions, inform the people concerned (participants, employees, candidates) and make sure your Members have sufficient command of the tool, in accordance with article 4 of the AI Act.

8. Data and confidentiality

The content you submit is processed to provide you with the Service, in the European Union for the components operated by KISMO, and sent to the providers listed above only to the extent needed for each function, without any identifier of your account. It is never used to train models. Recordings are deleted at the telephony and video conferencing providers as soon as KISMO retrieves them. The details (retention periods, rights, transfers) are in the Privacy policy and in section 17 of the Terms of Sale.

9. Regulated and prohibited uses

Authorized uses, regulated uses (recording conversations, recruitment and evaluation interviews, complaints, people mentioned, imported content, sensitive data) and prohibited uses (monitoring, scoring, automated decisions affecting people, unlawful content, misuse of the models) are described in section 3 of the Terms of Use and sections 12 and 15 of the Terms of Sale; model providers also publish their own usage restrictions, which the Client undertakes to respect (Terms of Sale, section 15.5).

10. Artificial intelligence governance at KISMO

10.1. KISMO keeps technical documentation of the Service (description, architecture, models, classification assessment, risk register, logging, human oversight measures, incident procedures), updates it at each significant change and makes it available to the competent authorities and, under confidentiality, to Clients who need it for their own obligations.

10.2. KISMO monitors the operation of the Service after it is placed on the market (quality indicators, reports, incidents), logs the processing carried out by the models (function, model used, date, cost, errors) for as long as the Client’s account exists, and applies an incident management procedure that includes informing Clients and, where required by law, the authorities.

10.3. KISMO supports the development of the skills of its staff and providers in artificial intelligence (how models work, limits, data protection, regulatory obligations), in accordance with article 4 of the AI Act.

10.4. KISMO follows developments in the AI Act (implementing acts, harmonised standards, guidelines, codes of practice) and adapts the Service and this notice accordingly.

11. Contact and reporting

For any question about artificial intelligence in KISMO, to report an incident or to request the documentation intended for deployers: hello@kismo.ai. Data protection: dpo@kismo.ai. Address: KISMO, 21 chemin du Moulin, 84530 Villelaure, France.