Artificial intelligence systems notice

This is an English rendering for convenience. The binding version is the Polish one, available at aivatary.com/pl/informacja-o-ai/. In case of discrepancy, the Polish text prevails.

1. You are speaking to an AI system

The AIvatary agent is an artificial intelligence system. We say so in three places at once, so that nobody has to work it out: in the first message of the conversation, in the description of the conversation window, and in this document. The conversation does not pretend to be contact with a human, and the agent does not present itself as a person.

2. What we use artificial intelligence for

3. What happens to content sent to the model

1. The model receives the question, the current conversation history and knowledge base passages found for that specific question. We do not send the whole knowledge base or any other organisation's data.
2. Retrieval is limited to a single organisation at the level of the query itself. This is not a filter over results but a condition without which the query does not exist. A passage belonging to another organisation has no way into the model's context.
3. Content sent to a model provider is covered by their processing agreement. Providers are named in the subprocessor register.
4. We do not use our customers' conversation content to train our own models.

4. What the model may not do

Three limits sit in code in front of the model rather than behind it, so they do not depend on how the model behaves.

1. An answer must be grounded in sources. The check happens before the customer sees the first character. An ungrounded answer is rejected whole and the conversation goes to a human. There is no send now, correct later mode.
2. The model does not change data. An operation that would change any state becomes a confirmation card approved by an operator with a separate click, leaving a trace in the audit log.
3. The model does not choose its own tools. The catalogue of available operations is closed and configured by our customer, with two independent kill switches.

5. Records

Every conversation turn leaves a record in the audit log: which knowledge passages were retrieved, what was decided and what was sent. The record holds no conversation content: it carries identifiers, codes and metrics, and any national identification numbers, bank account numbers, phone numbers, email addresses, card numbers and tokens that would pass through the decision trace are replaced with markers before the event reaches the log. A person's data appears in the log only where it is the subject of the entry. The log cannot be overwritten.

6. Human oversight

1. A conversation goes to a human when the customer asks, when coverage in the knowledge base is missing, when a tool fails, or when the topic falls outside the scope our customer defined. The catalogue of reasons is closed and every handoff is recorded.
2. The knowledge base runs through an editorial cycle managed by people. The agent sees only published versions, so editors, not the model, decide what may be said.
3. Questions without coverage go to the gap register, a list for the editors. That closes the loop: a human supplies the knowledge rather than the model inventing it.

7. What this technology cannot do

A language model operates on probability, not understanding. That is why the whole construction rests on keeping its freedom at the level of word choice, never reaching substance or consequences. An agent's answer is not legal, medical or financial advice, unless our customer published such content themselves and takes responsibility for it as its author.

8. The Ada test drive

On the "Try Ada on your own website" page a visitor can check, without an account, how Ada would answer customers of an indicated company. The test drive differs from a deployment at our customer in three places, and we describe them plainly:

1. Knowledge without human editing. The content of up to 25 public pages of the indicated website, without pages about people, is published automatically, skipping the editorial cycle described in section 6. Ada answers only from that content, so she may repeat a mistake or outdated information from the website. This is the only exception to the rule that a person decides what the agent may say.
2. No tools and no handover to a person. In the test drive Ada books nothing, records no requests and does not pass the conversation to a team, because there is no team in the test drive: when an answer has no coverage in the website content, Ada says she did not find it there. She also says what she would do in the full version.
3. Outbound call. When the visitor orders a call and confirms the number with an SMS code, Ada calls that number once. The call opens with a fixed sentence played by the system, not composed by the model: that the caller is an assistant based on artificial intelligence, that this is a demonstration of AIvatary using the indicated company's website as an example rather than a call on behalf of that company, and that the call is transcribed as text. Should the synthesis of that sentence fail, the model delivers the same introduction following a fixed instruction. Answers are phrased by OpenAI's voice model on the basis of the website content and turned into sound by ElevenLabs. The call lasts about five minutes and never more than six, and when the caller wants to finish, Ada says goodbye in one sentence. We delete the call transcript after the call ends or, if the caller gave an email address, after sending it there.

Test drive data is described in Part III of the privacy policy, and the rules of the service in the test drive terms.

9. Contact

Questions about how the AI systems work go to contact@aivatary.com. Data processing is described in the privacy policy.