How the AI works
What Salisan's AI features do, which providers and data they use, and how a practice limits them.
This page follows the privacy policy, the Terms of Service and the current code. Provider names and regions come from the configuration of the EU region and can change. The privacy policy and Terms prevail.
What the AI features do
- Agent. Answers questions about records the person may read, drafts text, proposes record changes, and prepares emails, forms, appointments and tasks. See Agent and you.
- Document reading. Extracts text from uploaded patient documents and images.
- Transcription. Turns recorded visits and chat dictation into text.
- Research. Searches the web and published guidelines for general questions.
- Translation. Translates template text and letter subject lines.
- Support assistant. Answers from public documentation and Salisan's support knowledge. It has no access to clinic data.
Providers and what they receive
Salisan sends Agent, document, transcription and research requests through Requesty, which the privacy policy lists as the AI routing provider. Salisan uses Requesty's EU endpoint. Models in the current configuration:
| Feature | Model provider | What it receives |
|---|---|---|
| Agent and document reading | Azure OpenAI, Sweden Central | Your request, relevant conversation history and what the Agent's tools retrieved. This can include patient names and clinical content. Document reading receives the original file. |
| Transcription | Mistral Voxtral | The original audio. |
| Research | Google Gemini, EU managed model | A general question, checked for names and other identifiers first. |
| Translation | Amazon Translate | Template text and letter subject lines. |
Salisan has not published an approved subprocessor register that names these model providers. See GDPR.
How data is handled
- Salisan does not use customer content, including patient data, to train AI models. Its agreements with AI providers do not allow it. Providers may keep requests for up to 30 days to detect abuse.
- The Agent reads only what the signed-in person may read. It treats retrieved records and web pages as data, not instructions.
- Patient tools give age or birth year instead of an exact birth date, and contact availability instead of contact details.
- Local text filters redact supported patterns for government and insurance identifiers, payment details and credentials before a text request leaves Salisan. They do not make content anonymous and do not catch every case. They do not redact original documents, images or audio.
- Do not give AI features personal data about anyone under 18, government or health-insurance identifiers, payment-card or bank details, passwords or access credentials. This covers chat, pasted text, attachments, documents and recordings.
- Salisan keeps records of AI usage for billing, not their content.
Excluding a patient
Choose Exclude from AI assistance in the patient's details. The Agent cannot read or act on the record, uploaded documents are not scanned, and visits cannot be recorded or transcribed.
Patients under 18 are excluded automatically from their recorded date of birth. The exclusion stays after they turn 18 until a member removes it. Setting an exclusion does not change existing Agent conversations.
Chats and recordings
The Agent deletes chats nobody has used for 90 days by default. An owner or admin can choose 30, 60, 90, 180 or 365 days under Agent settings → Chats. Anything saved to a record stays.
The same page sets what happens to visit recordings: keep with the patient record (default), delete after 30 days, or delete once the transcript is final.
Before a recording starts, the app shows Before you record. It says informing patients and getting consent where the law requires it is the practice's responsibility, and that Salisan does not ask patients for you. Support assistant chats are kept for 90 days.
See Agent settings and Transcription for the steps.