Lifelike AI should be clear about what it is and careful about what it does.
Athira deployments are designed around explicit identity, approved knowledge, defined actions, uncertainty behavior, and human handoff - so a natural interface does not become an unbounded one.
The deployment boundary
Identity
Never pretending to be human.
Every deployment states that it is AI in its first message, in plain language.
Asked whether it is a person, it answers honestly and offers a human instead.
Tone and expression stay warm, but the interface never implies a human is typing.
Boundaries a customer cannot switch on.
These limits hold across all Receptionist Mode deployments, with no configuration path around them.
When the answer is not known, Athira says so, offers follow-up, and keeps helping with the rest.
The open question and its context are recorded so a teammate can pick it up.
Human control, before and during.
Knowledge, behavior, permitted actions, and handoff points are signed off before launch.
Authors read, edit, arrange, and preview AI-generated structure in Build View before publishing.
Signals inform teaching decisions - not medical diagnosis, and not employment decisions.