When the AI Assistant suggests an intervention, surfaces a pattern, or drafts a clinical document, you need enough context to evaluate it. This article explains how the Assistant arrives at its responses, what shapes them, and how we test it. For everything else about what the Assistant can do, see Upheal AI Assistant.
The Upheal AI Assistant is currently in beta. We'll keep this article up to date as the feature evolves.
How the Assistant produces a response
The Assistant routes each question to the right kind of information, rather than searching everything each time. Depending on what you ask, it draws on different sources.
Questions about a client's clinical history, session content, or patterns: The Assistant searches that client's session transcripts, Progress Notes, Treatment Plans, and profile. It retrieves the most relevant material, and synthesizes a response with sources you can click to verify.
Practice management questions (scheduling, calendar, billing, invoices, forms, messaging): For these types of questions, the Assistant uses only the data that is in Upheal. Numbers and statuses come from your account.
Clinical or research questions: The Assistant can pull from a curated body of published clinical research (diagnostic criteria, clinical guidelines, peer-reviewed literature) to support suggestions with evidence.
Questions about Upheal itself: The Assistant searches the support center and our public roadmap, and shares the relevant article or feature post.
A single response often combines more than one of these. For example, "What's Alex's outstanding balance and the last thing we worked on?" pulls from both your billing data and that client's session notes.
For any record-based or clinical claim, sources appear in the response with links back to the underlying note, transcript, or research citation. Clicking through is the most reliable way to verify a suggestion before acting on it.
Clinical input is always framed as a suggestion with a brief rationale, never as a directive or a final clinical decision. The Assistant is built to give you a faster, better-informed starting point. Always use your professional judgement to evaluate any suggestions it makes.
What the Assistant is, and isn't, trained on
The Assistant operates on two levels:
The underlying language model: The Assistant is a HIPAA-compliant tool built on a foundation language model from a third-party AI provider, trained on broad, general-purpose text (publicly available web content, books, licensed datasets, and similar material). It is not trained on Upheal customer data, on PHI, or on any clinical records.
What it draws on at the moment you ask: Depending on the question, the Assistant pulls from your own clinical records, your Upheal practice data (calendar, billing, forms, and so on), a curated body of published clinical research, or our support center.
The Assistant’s responses have the context of previous messages within the same conversation, as well as information from past conversations. The Assistant’s Memory feature automatically saves details from conversations, like workflow preferences and recurring topics, to give more personalized support over time. This feature is enabled by default, but you can manually edit memories, or turn it off entirely.
More here on the Assistant’s Memory feature.
For how Upheal uses de-identified session data to improve other AI features, and how to opt out, see Opting out of de-identified data use for AI improvements.
How we validate it
The Assistant is in beta, and we validate it on an ongoing basis. Our process includes:
Pre-release evaluation: Before updates are released, the Assistant's responses are evaluated against representative clinical and practice-management scenarios. We look at accuracy and faithfulness to the source records.
Clinical review: Mental health clinicians on our team also review how the Assistant behaves, including in sensitive scenarios (risk language, diagnostic suggestions, treatment recommendations), and refine the guardrails in accordance with advice from legal experts.
Sources on every response: Record-based and clinical claims include a source link back to the underlying note, transcript, or research citation. Clicking references is a good habit. It lets you quickly check the clinical context behind a suggestion before acting on it.
Human-in-the-loop by design: The Assistant always asks for your explicit confirmation before creating, editing, or sending anything (an invoice, a message, a calendar event, a note). Nothing leaves your practice without you approving it first.
Ongoing monitoring. We track thumbs-up and thumbs-down feedback on responses and review issues regularly, so we can keep improving quality.
If a response ever looks wrong, use the thumbs-down button on that response and include a quick note about what could be improved. It's the fastest way to flag it for the team, and context from you helps us improve the Assistant.