Mental-health support creates risks that ordinary consumer AI products may not face. InTune therefore needs explicit boundaries, written down and designed for.
AI should not make clinical decisions. Every consequential judgement passes through a person.
Supporters need clearly defined responsibilities and limitations. They should not represent themselves as licensed professionals when they are not.
Some situations should move beyond the Supporter role. The system should provide a clear path toward qualified professional assistance.
Designing as though the model is reliable is itself a safety failure. The architecture assumes AI can:
AI output should never automatically become a consequential action.
These are explicit engineering and product requirements, not policy language:
Collect and retain only what the support relationship genuinely requires.
Where transcripts live, in what form, and for how long.
Who can read a session, and under what circumstances.
Defined lifetimes, and a real path to erasure.
In transit and at rest, across every component of the pipeline.
Explicit consent for recording and analysis, and clarity on how AI providers handle the data.