AI transparency

The assistant and AI-written drafts, scores and summaries are identified as AI in the product. A person using the assistant is always talking to an AI system, never to a human pretending to be one. Nothing AI-written is published to the public or sent to a customer without a person reading and sending it, so there is no AI-generated content passed off as human on our side. If you publish AI-generated text to the public yourself, you are responsible for any disclosure the law requires.

• Approval first. The AI never sends email, changes who owns a record, deletes or exports data on its own. A person approves each of those, and the approval is checked against the same permissions as every other action.
• Permissions apply. The AI only sees what the person asking may see, including their reporting line, hidden fields and role limits. Authorisation is decided by our server, never by what the model says or asks for.
• Switches. An administrator can turn AI off for the whole organization or for a person, and cap how much each person can use.
• Safeguards. We screen text that tries to manipulate the AI (prompt injection), treat web pages and emails as untrusted data, and strip anything from an answer that could send data somewhere it should not go.
• Feedback. People can rate answers and say why; the ratings show an administrator how well it is doing and refine its instructions within your organization only.

The AI scores deals and leads, not employees, applicants or consumers, and its output is advice. We do not build or sell Kairoq for decisions that have legal or similarly significant effects on a person, such as hiring, firing, pay, credit, insurance or access to essential services, and our Terms and Acceptable Use Policy forbid using it that way. For that reason we do not classify the features above as high-risk AI systems under the EU AI Act. The People (HR) module does not use AI.

• A request sends the AI provider only the content it needs. Providers are listed on the Sub-processors page: by default Anthropic for the models, Voyage AI for search embeddings and Deepgram for transcription, or the model provider your deployment configures.
• No training. We and our providers do not use your data to train or improve models. Providers may keep a request for a short time to detect abuse, under the terms of our agreement with them.
• We record how much AI is used (tokens, model, cost) per organization and person for limits and billing. Answers stay in the person's own chat history until they delete them.

• AI can state things that are false, miss context, or be out of date. Check anything that matters against the record before acting on it.
• Transcripts can contain errors, especially with accents, background noise, and mixed languages such as Hindi and English; speaker labels can be wrong. Don't treat a transcript or a call score as a verbatim or objective record.
• Scores and forecasts reflect the data in your workspace. If that data is thin or skewed, so are they; they show their reasons so you can judge them.
• Drafts can sound confident and still be wrong about a price, a date or a promise. Read before you send.

Organizations that deploy AI are expected to make sure the people using it understand it (EU AI Act, Article 4). The Security page and the in-product tutorial explain how the AI works inside the rules; this page explains its limits. Administrators should tell their teams which features are on, that AI output needs review, and that people can ask for a human to look again at any score about them.

To ask how a feature works, report harmful or wrong output, or ask a human to review something, write to us through the Support page. For your personal data rights see the Privacy Policy.