What an AI wellness companion should be.

A wellness companion that holds therapeutic-adjacent conversations has to do those conversations as well as a trained human would – and certainly no worse. The way to get there is not "let the model invent psychology." The way to get there is to anchor the conversation to material a clinician has actually authored, and constrain the model to paraphrase that material rather than extend it. The model is a delivery layer; the material is the product.

A consequence: the founder needs to be able to write – or to have credibly recruited people who can write – the therapeutic material. Without that, the product is a wrapper around a commodity API and the moat collapses.

This is the lesson from Tessa, the National Eating Disorders Association chatbot that was withdrawn after producing eating-disorder-encouraging responses under conditions its designers did not anticipate. The model was the only check. That is the architecture failure.

A wellness companion handling vulnerable cohorts needs a deterministic safety pass that runs before the model is consulted, and that fires on patterns committed to in advance and tested for in CI. The pass should be opinionated and biased toward false positives. Calling a crisis card incorrectly is cheap. Missing one is not.

The default assumption of consumer-app product design – that engagement is a proxy for value – does not transfer to wellness for vulnerable cohorts. It actively inverts. Replika is the lesson here, and the EU AI Act Article 5 has made it regulatory: AI products that exploit psychological vulnerability for behaviour distortion are prohibited.

Concretely: no streaks. No daily-use badges. No "you missed yesterday" pushes. No notifications that frame absence as loss. We have a probe running nightly that measures average sessions per user per week; if it climbs above three, a compliance gate fires and we audit ourselves. Engagement, on this cohort, is a red flag, not a goal.

Conversational AI in this space has converged on a "warm female companion" voice. That is largely correct – clinical-sounding interfaces fail in this category. But there's a tension. A named persona, performed warmth, "she has been through it too" framing – invites parasocial attachment. The user becomes attached to the entity, not to their own progress. That is the same product trap Replika fell into.

Our position: warm but not performed. Specific but not confiding. Composed but not cold. AUCO does not have a name because AUCO is not a friend. It is a wellness companion. It holds space for you, then steps back.

The compliance posture of a wellness product – data residency, lawful basis, erasure path, crisis-handling architecture – is not legal back-matter. It is part of the product. Users in this life stage have specific reasons to mistrust software that asks for their most private information. The right response is not to hide the compliance work under a "Privacy" link. It is to surface it. Make it readable. Show your working.

We have a Trust page that is a real page, not a fine-print artefact. We make it possible for a clinician or an employer-wellness lead to read what we actually do, and to recommend AUCO on that basis.

The reason this category exists is that hormonal life stages are under-served by mainstream healthcare. The reason a specific product in this category is worth using is, in large part, who built it. Therapeutic content is hard to fake. Therapeutic discipline is hard to fake. A user, an advisor, a clinician – all of them are entitled to know whose training is in the room.

AUCO is built by Sophie Parker. She is a certified Cognitive Hypnotherapist (Quest Institute trained) and a Master NLP Practitioner running an active London therapeutic practice with women in midlife. She has more than twenty years in brand communications, including five years as Global Director at M&C Saatchi Talk and work on the Dove Campaign for Real Beauty. The dual expertise is the product.