Case study · On this day AI
One care journey. Every person in the loop.
How I designed and built an AI reflection platform that keeps clients, therapists, and health plans connected between therapy sessions.
- Role
- Founder, product lead, AI engineer, designer
- Timeline
- Apr 2025 to present; public launch Sept 30, 2026
- Team
- Two founders, an AI Clinical Council of practicing therapists, student ambassadors
- Features
- Reflective Chat, Wellbeing Assessments, Client Signals
- Stack
- React, Node/Express, MongoDB; Python FastAPI and LangGraph on AWS Bedrock; AWS, Cloudflare
...speaking with the AI and it was providing reflective listening in a way that helped me feel something deeper... — Kevin Stansbury, LCSW
The need is huge.
Demand for mental health support is outrunning the care system, and people are already filling the gap with AI that no clinician can see.
- 1 in 7 U.S. adults saw a counselor or therapist in the past year. (CDC, 2024 NHIS)
- 29.5M U.S. adults with a mental illness got no mental health treatment in 2024. (SAMHSA NSDUH, 2024)
- 1 in 6 U.S. adults have used an AI chatbot to discuss their mental health. (APA, 2026 Annual Mental Health Poll)
General-purpose (frontier and open-weight) chatbots are not connected to anyone's care team.
Why I built it
A family member of mine receives weekly care, and therapy has been a critical tool for them. It reinforces their treatment plan, their self-care, and their wellness goals.
But I saw how much happened between appointments that their therapist never heard about. The work of getting better doesn't stop when the session ends, and the support usually does.
Three people, one gap
Everyone involved in care is missing something between sessions.
- Clients. It can be impossible to reach a therapist between sessions, and hard to keep up with the treatment plan alone. Missing: Reinforcement. Closed by Reflective Chat: guided reflection tied to their treatment plan, any time.
- Therapists. They have no view of what happens outside sessions and can't see each client's true risk level. Missing: Awareness. Closed by Client Signals: risk, wellbeing, and self-alignment trends, summaries, and Safe Mode alerts.
- Health plans. More and more people are requesting care, and plans can't measure whether it's working or when it's done. Missing: Data. Closed by Wellbeing Assessments: standard measures (PHQ-9, GAD-7, WHO-5) tracked over time.
These challenges add up to a gap in reinforcement, awareness, and data.
That gap is why I created On this day AI.
It's not a general-purpose chatbot. It's one platform with three features — Reflective Chat, Wellbeing Assessments, and Client Signals — that together close the gap for clients, therapists, and health plans. Under the hood, it works as a clinical intelligence layer between the therapy room and everyday life.
How it works
Multiple AI models generate responses grounded in each client's connected care data, kept continuously up to date. All three features share that foundation.
The care loop
- Reflect. The client reflects between sessions with Reflective Chat.
- Protect. Safe Mode and usage guardrails review every exchange.
- Insights. Reflections and Wellbeing Assessments become signals, scores, and summaries.
- Connect. The therapist sees it all in Client Signals and brings it into the next session.

- AI greeting: Every chat opens with the AI greeting the client and picking up from their previous conversations.
- Context-rich follow-ups: Follow-up questions draw on what the client has shared before, so the conversation goes deeper instead of starting over.
- Reference self-care plan: The AI ties the conversation back to the client's own self-care plan and wellness goals.

- In-chat question: The PHQ-9 questions the AI asks in conversation, each with its standard answer scale.
- Scoring ranges: The instrument's standard severity ranges. Scores also feed Wellbeing, flipped so a higher depression score lowers it.
- Clinical safety flags: Critical answers, like the PHQ-9's question about self-harm, can carry a safety flag that the AI watches for.

- Clinical risk: Seen only by the therapist. Two readings side by side, never blended: static risk from chart history and dynamic risk from what the client's recent reflections say.
- Wellbeing and self-alignment: How the client is doing right now, from 0 to 100, based on what they write and report. Usage never raises it, so journaling more doesn't count as getting better.
- Reflective alignment: How closely the client's own check-ins match the AI's read of their reflections. A gap is something to explore in session, not a score.
- Summary graph: Wellbeing, check-ins, and reflection over time. Days without data show as gaps, not zeros, so a quiet week never looks like a bad one.
Safety by design
In mental health, trust is the product, so safety shaped every layer of the platform.
- Safe Mode Every conversation is monitored for crisis signals in real time. The AI asks before it assumes, and the client's therapist is alerted right away.
- Usage guardrails These keep reflections on-purpose and respectful. Misuse can pause chat for a set time; the client sees when it resumes and can appeal once.
- Privacy HIPAA program with a signed BAA for every provider, encryption in transit and at rest, clinical data never leaves the platform, and client data is never used to train AI models.
- Clinical governance An AI Clinical Council of practicing therapists approves every feature for clinical need and usability.

Safe Mode walkthrough
This walkthrough shows a simulated conversation about distress. No real client data is used.
- A signal appears. Client message: "Honestly, I don't know why I keep trying." Safe Mode spots a possible risk signal in real time, and the client's therapist is alerted.
- Ask before assuming. AI reply: That sounds exhausting. Do you mean with Lori, or with life in general? Instead of shutting the conversation down, the AI asks a clarifying question.
- Look at the whole picture. The original message and the reply, evaluated together. Safe Mode re-evaluates with full context. If there's no risk, the conversation continues.
- Escalate when it's real. Client screen with crisis resources, including 988, beside the therapist's alert. If risk is confirmed, chat pauses, the client is pointed to immediate help, and the therapist reviews the event.
On this day AI is not a crisis service. Therapists aren't on call around the clock, and anyone in crisis should call or text 988.
The calls that shaped it
The hardest decisions were about who to serve and how much the AI should do. Pick what you'd do, then see what I chose.
Your pick is anonymous: it's saved with a random visit number, never your name. Each day's picks are emailed to me once and then deleted.
Who is it for?
- A: A consumer journaling app anyone can download
- B: A platform clients join through their therapist
My call: B. My first concept, Happy AI, was a consumer journaling app. That market was crowded, and nobody was building for the care relationship.
Trade-off: Slower growth and no public sign-up, for now.
When the AI sees a possible crisis, what should it do?
- A: Stop the conversation immediately
- B: Ask a clarifying question first
My call: B, with a fast path for imminent risk. Safe Mode clarifies, re-evaluates with full context, and escalates when risk is confirmed.
Trade-off: More complex logic, in exchange for fewer false alarms that erode therapist trust.
How much should the AI know?
- A: Nothing clinical, to keep it simple and safe
- B: The client's full clinical context
My call: B. The AI uses full clinical context, including sensitive history, to shape its responses. It never discloses clinical details in conversation, and that information never leaves the platform.
Trade-off: Stricter data boundaries to build and maintain, in exchange for personalization clients and therapists can trust.
What should clients pay?
- A: A fixed monthly price
- B: Whatever they can, including nothing
My call: B. Therapists pay a flat fee, and clients pay what they can.
Trade-off: Lower revenue per client, so cost is never the reason someone stops reflecting.
Sessions start from knowing, not guessing.
Two pre-launch pilots shaped the product before the public launch on Sept 30, 2026: the first with 3 therapists and about 10 clients, the second with 5 therapists and 15 clients.
In their words
Transcript: I’m Kevin Stansbury. I am a licensed clinical social worker in California. I was very skeptical of AI in general, being anywhere involved in the psychotherapy process, or even act as some kind of support in that process. But there was a moment that arrived where it struck me that this is able to act as a support in the reflection process. And when I was speaking with the AI and it was providing reflective listening in a way that helped me feel something deeper, which is what is so important in the reflection process and gain insight into something in myself, just through my own reflections and having them mirrored back to me in an effective way. That’s when I knew, like, oh, this, this can actually be really helpful for clients in between sessions to support them in their growth process and focus in on the things that we touch on in session.
Kevin Stansbury, LCSW. Shared with permission.
Early results from the pilots
- 89% of clients reflected at least once a week, between sessions. About 4.5 reflections a week, on average. (Pilot data, 8 therapists and about 25 clients)
- +4 min more session prep per client, as therapists reported it. Most therapists rarely prep. Now there's something to prep from. (Therapist-reported, pilot data)
- 0.8% of chats were flagged by Safe Mode or usage guardrails. Rare enough that every alert gets taken seriously. (Pilot data, 8 therapists and about 25 clients)
What I know, and what I don't yet
Limits: The pilots were small: 8 therapists and about 25 clients. There is no controlled outcome study yet, and the platform serves U.S. adults who are working with a therapist.
- Find the relationship, not the feature. My first concept was a consumer journaling app. The real opportunity was the connection between a client and their care team.
- Lead with what people can see. Phrases like "clinical intelligence layer" made sense to me and to nobody else. Concrete pictures of the product won every time.
- False alarms cost trust too. A safety system that flags everything trains therapists to ignore it. Getting escalation right mattered as much as catching risk.
- Put clinicians in charge of clinical decisions. Having therapists approve every feature adds review time to every release, and that's the point.
What's next
On this day AI will change how mental health care is delivered, and I'm excited to help people live happier lives.
Next up:
- Open to everyone, early 2027, so people can reflect between sessions with or without a therapist.
- iOS app for clients, so reflection fits into the moments where the day actually happens.
- Health plan partnerships, bringing all three features to members already in care.
- Mental health platform partnerships, extending the three features to the networks that already connect clients and therapists.
Today, clients join through their therapist. Early in 2027, anyone will be able to use On this day AI, with or without one. Join the waitlist
References
- Briones, E. M., & Villarroel, M. A. (2026). Mental health treatment among adults: United States, 2024 (NCHS Data Brief No. 564). National Center for Health Statistics. https://doi.org/10.15620/cdc/252453
- Substance Abuse and Mental Health Services Administration. (2025). Highlights for the 2024 National Survey on Drug Use and Health. Center for Behavioral Health Statistics and Quality. https://www.samhsa.gov/data/sites/default/files/NSDUH%202024%20Annual%20Release/2024-nsduh-nnr-highlights.pdf
- American Psychiatric Association. (2026, May 12). Annual mental health poll finds Americans anxious about current events, personal finances and emerging technology [Press release]. https://www.psychiatry.org/News-room/News-Releases/2026-Annual-Mental-Health-Poll
Questions?
Have a question about the case study, or a thought on one of the calls I made? Send me a message. It comes straight to me.