How Care Tales uses AI models, what is sent to them, what is kept and what the technology cannot do.
Care Tales uses AI in the chat assistant and in the production of module content. This document sets out which models we use, what reaches them, what we keep, what rules the server enforces and what a customer should tell its patients.
We use Google Gemini models and nothing else. The default chat model is gemini-2.5-flash. Google Gemini is also used for embeddings, topic classification, module generation, translation, speech, image generation and video generation. We also use Google Cloud Text-to-Speech and Google Cloud Translate. Care Tales does not use OpenAI, Anthropic or any other model provider anywhere in the product and does not host its own models.
Patient chat always runs on Care Tales' own Google Cloud account, through Gemini on Vertex AI in us-central1 in the United States. An organization may supply its own model key; it is used only to index the documents that organization uploads to its knowledge base, never for patient messages.
For a chat answer, the server sends:
The scraped host page text is used to answer the question and is not stored. We do not send a patient's name, email address, phone number or account profile to the model as part of a chat request.
Module generation and translation send module content and source material, not patient data.
Stored, indefinitely: organization chatbot conversations on caretales.com and in the embeddable widget. We keep the verbatim patient message and the verbatim assistant response, along with session metadata such as page URL and title, per visitor counters and daily and lifetime rollups including an estimated model cost. For signed in users, chat history is stored under the account.
Stored with more detail in the widget: the embeddable WordPress widget also stores the full untruncated IP address, country, city, region, coordinates, browser, operating system, device type, language, timezone, screen size and referrer.
Not stored: the in-module "Ask Cari" assistant. Only counts of how often it was opened and how many messages were sent are recorded. The message text is not retained. Scraped host page text is not retained. On the main site chat path the IP address is used for the geolocation lookup and is not stored.
There is no retention schedule for stored transcripts. They persist until the account or the organization is deleted. Deleting an organization recursively deletes its chatbot transcripts.
Customer and patient data is not used to train models. Care Tales does not train models. Patient chat, the embedding of patient questions and topic classification run on Gemini through Google Cloud Vertex AI, under Google Cloud's terms, which do not permit Google to use that data to train its models. Module generation, translation, narration, images and video handle module content and no patient data; they use the Gemini API on Care Tales' paid account, under which Google does not use requests to improve its models. We do not opt any customer into a model improvement program.
If an organization supplies its own model key, indexing its uploaded documents is billed to that organization's own Google account and governed by that account's terms. An organization on Google's free tier should know that the free tier permits Google to use the content to improve its products. Patient messages never go out on an organization's key.
The server appends a set of rules to every organization system prompt. An organization can write its own prompt and load its own knowledge base, but it cannot remove or override these:
Raw patient messages never leave the server for reporting. Aggregate insight terms are shown to an organization only once a term has appeared in at least three separate conversations.
Module content is not published to patients straight from a model. Every module moves through Draft, then Submitted, then Verified, and verification is a review by Care Tales clinical and editorial staff before the module is available. Modules are produced at three reading levels and in six languages: English, Spanish, French, Russian, Portuguese and Haitian Creole.
The chat assistant's individual answers are generated at request time and are not reviewed by a human before the patient sees them. The knowledge base those answers are drawn from is controlled by the organization.
Every organization has a switch that turns AI features off for that organization. When it is off, the chat assistant stops answering for that organization's patients and no further requests are sent to a model on that organization's behalf. A customer can ask us to set it at any time, and we will set it for the customer on request.
We recommend the customer tells patients, in its own privacy notice or at the point of use, that:
Care Tales cannot identify a patient who has no account, so any patient request to see or delete chat history has to come through the organization.
We update this disclosure when the models, the data flows, the safety rules or the retention behavior change. Each revision carries a new effective date and is notified by email to the administrative contact for each customer organization.