A startup called Onix launched this week claiming to be “a Substack for chatbots”: subscribe to an AI replica of a named human expert — a therapist, a paediatrician, a nutritionist — and get on-demand consultations at a fraction of the in-person cost. Co-founder and CEO David Bennahum (former Wired contributor, company headquartered in Canada) pitches it as “Personal Intelligence”: user data stored encrypted on-device, no scraped IP because experts train their own bots, hallucinations suppressed by domain guardrails. Seventeen vetted experts in the beta. Subscription pricing not yet published but positioned below typical hourly rates.

That’s the signal. Now map the system.

←TODAY: Onix enters beta with 17 health-and-wellness experts; guardrails demonstrably failed Wired’s testing in at least two documented hallucination incidents.
→3012: Every senior AEC specialist carries a queryable knowledge twin — the value isn’t the bot, it’s the corpus curation discipline required to build one honestly.
Fulcrum: The expert-bot model only holds if the domain guardrails actually hold; in high-stakes technical fields, that bar is categorically higher than NBA playoff trivia.

Onix’s architecture has three load-bearing claims. One: on-device encryption means the company cannot hand over conversation data to a government beyond a user’s email address — a meaningful privacy posture, though Canadian PIPEDA, not GDPR, is the governing framework here. EU and Swiss users interacting with the platform sit in a regulatory grey zone: no data adequacy decision specifically covers this on-device model, and Swissmedic oversight of AI-delivered medical guidance remains underspecified. Two: because experts supply their own training content, the IP compensation problem that has driven the NYT lawsuit and Authors Guild actions since 2023 is structurally avoided. Three: domain guardrails keep the model from wandering. As Wired’s Steven Levy reported, that third claim broke in beta — one bot followed an off-topic NBA thread and hallucinated conference finals details; a second, diverted from ketamine therapy into a conversation about the indie band the Mendoza Line, reframed the band’s breakup as “a powerful expression of their neurobiology in distress.” Still in beta, yes — but those failures were on the live product during review.

The comparable precedent is instructive. Manhattan psychologist Becky Kennedy runs a parenting-advice business featuring a chatbot named Gigi trained on her methodology; that operation pulled in $34 million last year, per Wired’s reporting. Onix isn’t inventing the category — it’s trying to industrialise it with a platform model and better privacy plumbing.

Here is where it gets directly relevant to your desk. The expert-bot model maps cleanly onto AEC specialisms: structural engineering rule-of-thumb consultants, Minergie certification advisors, SIA norm interpreters, BIM execution plan specialists. A senior building physicist with 20 years of hygrothermal envelope knowledge is exactly the profile Onix’s white paper describes when it writes that “the expert’s knowledge base becomes a capital asset that generates revenue independent of their time.” The PAZ cohort includes practitioners who already teach their workflows as products — the logical extension is a queryable version of that expertise.

Atelier: Before you model this for your own practice, stress-test the guardrail assumption against your actual domain. An NBA hallucination is embarrassing; a hallucinated fire-resistance rating or a fabricated SIA 380/1 U-value threshold is a liability event. The Onix model requires not just training content but an adversarial testing protocol — someone systematically trying to break the bot before a client does.

Two structural risks the platform doesn’t resolve. First, the disclaimer problem: Onix displays a notice that engagement with medical bots is guidance, not treatment. As Wired notes, in a world where people already treat Claude and ChatGPT as therapists and where real healthcare is expensive, that warning “seems destined to be widely ignored.” The same dynamic applies in AEC: a practitioner who offers a “building code guidance” bot will find users treating its outputs as binding interpretation. Second, the revenue-split terms between Onix and its expert providers are not yet public — the platform’s long-term incentive structure is opaque at launch.

The move for a PAZ reader is not to rush onto the waitlist. It’s to watch Delphi.ai and Maven AGI — direct competitors with longer track records — alongside Onix, and to begin the corpus curation discipline now regardless of platform choice: document your decision logic, your standard responses to recurring client questions, your domain exceptions. That corpus is the real asset. The bot is just the delivery layer.

Source: Wired