Jimini Health Built Mental Health AI Like a Clinic, Not a Chatbot

When we look at a healthcare AI company, the interesting question is not whether it can demo well. A lot of AI products demo well. The harder question is whether the company is built around the actual constraints of care delivery: clinical accountability, patient trust, safety boundaries, and workflow fit inside a real treatment process.

I recently sat down with Jimini Health CEO and co-founder Luis Voloch to talk about the company’s clinician-supervised generative AI product for mental health. Our discussion of nearly 30 minutes is recorded for you to re-watch here, meanwhile, I’ve put together key insights on what you can learn from it, especially those who are interested in healthcare and any high-stake AI.

Is Chatbot The Answer to Mental Health Treatment?

Both Luis Voloch and I agree that it is both yes and no.

Yes, because people are already using general-purpose AI like ChatGPT for mental health support at massive scale. They are turning to it because standard care is not quick and easy enough to get and even harder to maintain over time. A therapist might be available once a week. A psychiatrist might be available once a month. That leaves a wide stretch of time where people are on their own. If they are struggling on a Tuesday night and their next appointment is five days away, the tool that answers immediately becomes the tool they use. That is how general-purpose AI becomes part of mental healthcare by default, even if nobody intended it to.

No, because ChatGPT was not built as a mental health product. It does not know the patient’s diagnosis, treatment plan, medication context, or the exact thing a clinician is trying to help them work through. Yet people are already using it for one of the most sensitive tasks possible.

Behavioral health covers a wide range of acuity, from lower-level support needs to serious illness, and that range matters. An AI system cannot safely improvise across that spectrum without clinical context, a treatment plan, and a human who remains responsible for care.

So, Jimini’s response was not to launch a direct-to-consumer bot and hope safety could be patched in later. The team decided to build mental health AI that works inside clinical care.

Jimini’s Bet: Extend the Therapist Between Sessions

Jimini’s answer is not to replace the therapist. It is to extend the therapist into the hours between sessions. Its product, Sage, is introduced by the clinician during normal care. The patient does not download a mental health chatbot and start from zero. The therapist or psychiatrist brings Sage into the relationship as an additional support tool, with clear boundaries around what it is for. 

Sage knows the care plan and what the clinician wants the patient to work on before the next session. The human professional still owns diagnosis, treatment planning, and clinical judgment. That design choice matters because psychotherapy is not just a sequence of appointments. The appointments are the anchors, but much of the actual difficulty happens in between. If support exists only inside the weekly session, most of life is still happening outside the frame.

This also helps with one of the biggest failure modes in mental health AI: sycophancy. General models are often tuned to be agreeable, which can become dangerous in emotionally sensitive contexts. Jimini’s answer is to anchor the interaction to clinician-defined goals rather than the patient’s moment-to-moment desire for validation. The model is not there to win the conversation. It is there to support the care plan.

The 25-Model System Under the Hood

Luis described Sage as a coordinated architecture of roughly 25 specialized models, each handling a different function. Some focus on escalation, screening for patterns that need to be surfaced back to the clinician. Some handle memory, which in mental health is not just a list of stored facts but a structured record of what matters emotionally, clinically, and over time. Others help keep responses aligned with the clinician’s strategy for that patient.

That architecture sounds more credible than the standard “one model plus a wrapper” story because mental health care is not one task. It is a sequence of judgments, reminders, boundaries, and pattern recognition steps that have to stay connected to a treatment context. A system built from multiple specialized components makes more sense here than pretending one generic model can do everything reliably.

The key point is not the number 25 by itself. It is what the number implies: Jimini appears to have spent serious effort decomposing the work into narrower functions, with supervision and coordination built in. That is usually what mature healthcare software looks like when a company is solving for reliability instead of novelty.

The Real Moat Is Owning the Clinical Operation

The most important part of the conversation was not how smart the AI was. It was the operating model behind it. Jimini runs its own clinic with full-time licensed clinicians treating real patients, and every model update is tested there before it goes out to partner organizations.

That creates a much deeper moat than a polished interface or a clever prompt stack. The company gets a live clinical environment where product, workflow, and safety assumptions are exposed quickly. Synthetic testing can catch some issues. But it cannot fully reproduce what happens when a patient is distressed, when a clinician has limited time, or when trust in the tool is uncertain. Real care settings generate the edge cases that matter.

This is where Jimini starts to look differentiated as a healthcare company, not just an AI application company. Owning the clinic means owning the feedback loop. The team can observe how model behavior lands in real workflows, how clinicians actually introduce the product, what patients respond to, and where failures show up before they become scale problems. That kind of loop is slow to build and hard to fake.

For SCB 10X, that is the part worth paying attention to. The company does not sound like it is chasing engagement with a consumer chatbot and retrofitting a healthcare story afterward. It sounds like it is building a supervised care system where AI becomes useful because the clinical operation around it is real. In a category full of thin wrappers, that is the kind of rigor that is defensible.

Watch the full conversation: https://www.youtube.com/watch?v=cnvk1HibUOM