A research and product design engagement for a mental health startup: what happens to therapy patients in the days between sessions, and how a product can support that gap without stepping into the therapist’s role.
01 · the problem
02 · research strategy
I ran two research tracks in parallel: a patient survey (20 respondents) and semi-structured interviews with 5 practicing therapists. Both were scoped across the full patient journey, from symptom onset to ongoing recovery, not just the features we already had in mind.
Why a survey. A survey let me see patterns across a larger group fast: what already helps patients cope, and how they actually feel about AI involvement. That gave the interviews a foundation in real behavior, not assumption.
How the therapist interviews were structured. Interviews opened broad, before any product concept was shared, and were built around three questions: how therapists piece together a patient’s week, what helps most across their caseload, and where a patient’s account in session diverges from what actually happens day to day.
Segmenting by recovery. Patients don’t share one definition of recovery. I split respondents into three groups by presenting reason and defined recovery separately for each:
Managing an active clinical symptom. Recovery means symptom reduction.
Struggling but not clinically symptomatic. Recovery means regained stability.
In therapy by choice, not crisis. Recovery means growth, not symptom change.
A finding about recovery for someone managing panic attacks doesn’t transfer to someone working through a breakup.
03 · key findings
The coping strategies patients relied on most were not the ones they rated as most effective. Talking with another person and exercise were rated the most helpful; generic meditation-app content and passive waiting were rated among the least helpful.
AI attitudes were a genuine three-way split, not a soft majority. Some patients described real value; others found responses inconsistent or subtly steering; a smaller group rejected the premise outright on trust grounds. That split is substantive, and designing around a single average would have misrepresented how patients actually engage with AI.
Therapists provided the most consequential insight in the study: recovery is distinct from task compliance. One patient’s meaningful change was not reflected in a higher completion rate; it showed up in how quickly they recovered from a missed day, without the earlier pattern of self-criticism and withdrawal.
The primary barrier is not insight but action, what one therapist described as “bodily inertia” rather than a lack of motivation. This shaped a core product requirement: reduce the effort required to start, rather than introduce another habit that demands sustained willpower.
04 · product decisions
Every idea generated in research was weighed against the strength and convergence of the evidence behind it. Three features cleared that bar for the beta; everything else waited.
A near-zero-effort capture surface, including voice.
Evidence: patients described decision paralysis in the moment, and therapists independently named bodily inertia as the primary obstacle to any self-directed effort (Finding 4). The feature answers both directly.
A present-focused reflection space, not an advice-giving one.
Evidence: patients said what they wanted most in a hard moment was to feel heard, not advised, a role that maps directly onto forms of journaling therapists already prescribe, giving the feature a clinically grounded template rather than a speculative one.
A patient-reviewed recap of the week, shared only with approval.
Evidence: therapists independently named the same blind spot, a recency-biased, reconstructed recap of the week. Because the AI-trust split (Finding 2) showed patients don’t uniformly want their reflections shared, control over what reaches the therapist stays with the patient.
Reminders and streak tracking were cut. Patients already struggle with habit tools, and therapists were clear that interpreting patterns is clinical work, not an early AI’s job.
05 · evaluation plan
A staged plan, usability first, then a live beta, so a good demo couldn’t be mistaken for a good product.
Usability · 8 participants
Five core tasks, with go/no-go rules set in advance: 80%+ task completion, 100% ending completion, and a scripted safety suite requiring zero critical violations to pass.
Live beta · 2 weeks
Behavior over stated intent: activation, repeat use at 7 and 14 days, and an unprompted-use rate to check the product earns engagement rather than running on reminders.
06 · scope
Both samples were small and deliberately targeted rather than randomly sampled. I recruited patients already in therapy, since understanding this gap required people with direct, current experience of it, not a general population that could only speculate. At this stage, the survey’s job was to establish direction, whether this kind of support was worth building at all, and what shape it should take, not to produce a statistically representative study.
A few of the sharpest insights, the recovery-versus-compliance distinction and the bodily-inertia framing, came from a small number of therapist accounts. Validating them across a larger caseload is the natural next step.
20-respondent patient survey; interviews with 5 practicing therapists. Insist Health, 2026.
All projects