Mindbloom case study
Designing AI support before and after therapy
I designed AI-assisted intention setting and voice journaling for at-home ketamine therapy. Intention quality rose 23%. Our three-person team shipped both flows in eight weeks.
At a glance
About the company At-home ketamine therapy in 35 states and DC, with more than 200,000 sessions delivered.
What my work did
- intention quality after release, beating the team’s target
- +23%
- from brief to two shipped flows
- 8 weeks
In this project
Two moments the app wasn’t helping with
Mindbloom’s clinical team had seen two patterns. Clients who set an intention before a session rated the session higher. Clients who journaled afterward were more likely to continue their program.
The app helped with neither. Intention setting was loosely guided, and journaling happened in paper notebooks mailed to new clients. The brief was to bring both into the app and use AI to help.
Three questions, then three suggestions
Most clients did not know what a good intention looked like. Too open, and they froze. Too prescriptive, and it felt impersonal.
We prototyped in Figma: one long chat or one question at a time, whether clients could ask questions back, and how many questions before it felt like a chore. Three was right. Clients answer by tapping, typing, or speaking. The app then suggests three intentions based on their answers. They choose one or write their own, so the final words are always theirs.
- Personalization
- Short direct question
- Validation & affirmation
- Go deeper – understand
- Validation & affirmation
- Suggest a theme, but don’t be too prescriptive
- Yes / No will fork the experience. Let’s design for the easy path here.
Design for the person after a session
Research showed memories were most vivid right after a session. Clients who closed the app to journal later rarely came back. So journaling starts the moment the session ends, on the same background, as a continuation rather than a new task.
After a session, some clients have blurred vision or reduced motor control. I doubled the type and button sizes, made the flow voice-only, and asked one question per screen.
Before a session
After a session
Results, and a lesson about AI
Intention quality improved 23% on the team’s quality measure, beating the target. But a vocal group of clients did not want AI involved in something so personal.
The help itself was not the problem. The label was. We stopped leading with “AI” in the interface and described the feature by what it did: suggestions based on the client’s own answers. Completion rose right away. People wanted the help; they did not want a technology pitch in the middle of therapy.
Journaling missed its first goal. There was no baseline, since it replaced paper notebooks. The team began testing a shorter flow and accepted that some clients would rather journal later.
What I’d do differently
“Journal while it’s fresh” made clinical sense, but it pulled against “don’t jolt someone out of a therapeutic state.” We never fully resolved that. With more time, I would have tested the timing as carefully as the interface.
