1. Principles
- Transparency — users should understand they are talking to an automated assistant.
- Human in the loop — significant decisions are made by a human, not a model.
- Customer data stays the Customer's — it is not used to train third-party models.
- Bounded scope — the bot answers within a defined context and does not replace a professional.
- Verifiability — answers can be traced back to context sources and logs.
2. How answers are generated
2.1. An answer is produced by a language model from the user's message, the Customer's knowledge base and the instructions configured for the bot.
2.2. Models can make mistakes, produce plausible but inaccurate statements and lose context in long conversations. This is a property of the technology, not a defect of a particular setup.
2.3. For critical scenarios we configure guardrails: an allow-list of topics, mandatory handover to an operator, or templated answers instead of generation.
3. Transparency for end users
3.1. The Customer must inform end clients that they are interacting with an automated assistant at the start of the conversation or in the channel description.
3.2. A path to a human operator must be available in the conversation.
3.3. Creating the false impression that the user is talking to a specific individual is prohibited.
4. Human involvement and prohibited scenarios
4.1. The Service is not intended to take standalone decisions with legal or similarly significant effects for individuals — in particular denial of service, creditworthiness assessment, HR decisions or medical diagnosis.
4.2. Bot answers are not medical, legal, tax or investment advice.
4.3. For sensitive topics, mandatory escalation to an operator must be configured.
5. Data and model training
5.1. Customer and end-client data is not shared for training third-party models. Agreements with model providers exclude such use.
5.2. Service improvement is based on anonymised and aggregated data, or on Customer data only with the Customer's separate written consent.
5.3. Model providers, processing regions and context retention periods are recorded in the DPA.
5.4. The Customer is responsible for the content of the knowledge base and for keeping excessive personal data out of it.
6. Quality and control
6.1. Before launch, the bot is tested against a set of the Customer's typical enquiries.
6.2. A dialogue log is kept so that a disputed answer can be reviewed and the configuration adjusted.
6.3. The Customer can report an incorrect answer through support; quality reports are handled within the timeframes set by the SLA.
6.4. We do not guarantee the factual accuracy of every generated answer. Ultimate responsibility for communication with end clients remains with the Customer.
7. Customer obligations
- comply with the Acceptable Use Policy;
- keep the knowledge base lawful and up to date;
- assign staff responsible for conversation escalation;
- inform end clients about automated processing;
- not use the Service in the scenarios listed in section 4.
8. Feedback
8.1. Questions about AI use, requests to disclose the models used and reports of incorrect answers can be sent to alibek.abdekov2110@gmail.com.
8.2. This policy is reviewed at least annually and whenever the Service's technology stack changes materially.