Case study · Evoke
Evoke is a Polish AI companion for difficult emotions: conversations, short exercises and a marketplace of psychologists. It earns on subscriptions and booked sessions, and for many people it is the step before therapy. The client came to us with a live MVP; I led the next iteration, v2, now in development.
Lift 30-day retention from 3.1% to 5% within three months of launch; the category median is 3.3%. Retention feeds both revenue streams: people subscribe and book only after weeks of use. No daily-use target on purpose: people open Evoke when something happens.
Tracked alongside it
First messages: screens before the first message in the flows, the share who send one after launch
Follow-ups: opened, and returns within two days, the early read on retention
Upgrades: explaining the paid plan in testing, upgrades after launch
Bookings: reaching the booking screen from a chat in testing, bookings after launch
Usability: task success, SEQ and SUS on the MVP's five tasks
v2 isn't live yet: the live MVP against the v2 prototype, same five tasks, 8 participants each. Screens are counted in the flows.
Screens before the first message:
5to1
Explained correctly what the paid plan adds:
3 of 8to7 of 8
Task success, reaching the booking screen from a chat:
2 of 8to6 of 8
Task success, finding an exercise for how you feel:
4 of 8to7 of 8
Average task ease, SEQ 1–7:
4.1to5.8
SUS score:
58to76 (+18 points)
Only 3.1% of MVP users still opened Evoke 30 days after install, and both revenue streams went with them: 0.8% upgraded, 0.4% booked a specialist. The MVP asked for an account before the first word, cut conversations at a message limit, graded people's emotions and hid specialists in a tab few opened.
10
in-depth interviews
with MVP users, 4 of whom had left
8
baseline test participants
5 tasks on the MVP, and SUS
4
licensed psychologists
from the marketplace, interviewed
Also
Funnel analytics of the live MVP, from install to booked session
Support emails and store reviews, tagged by theme
Nobody in acute crisis recruited; a psychologist on call in every session
Interviewees who had thought about therapy but never went
“I’m not sure my problem is big enough.”
People paid or booked only after three weeks
The median upgrade came on day 19
New users who left at sign-up, before a single message
Free conversations that hit the message limit, often mid-thought
Only 1 in 25 of those users upgraded
People came before an event they dreaded, almost never after it
And scores about themselves read as a grade: “I don’t want to be rated for how I feel”
Users who ever opened the specialists tab
The riskiest moment, psychologists said: a chatbot reply to someone in crisis
Design sprint with the founders and a clinical psychologist
Journey map from first launch to first booked session, MVP vs v2
Effort vs. Impact matrix for 40+ ideas
Talk first, sign up later
Evoke follows up after an event you mentioned: "How did the interview go?"
Limits that never cut a conversation
Exercises sorted by how you feel
A way to a specialist from inside the chat
I tested the flows in Figma with 8 MVP users on the baseline's five tasks, with a psychologist on call.
v2 kept the MVP's core: conversation, exercises, a growing profile and specialists. I reworked the five places where the funnel lost people and moved every screen onto a new design system: one typeface, blue and pink ramps and Material UI's 8px grid. A desirability test with 30 people chose between three directions: the MVP read as "impersonal" and "sterile", the winner as "calm" and "trustworthy". A second designer helped under my direction.
Decision 01
MVP scores like "Emotional expression: 80" read as a report card and gave no reason to return. v2 keeps only numbers that describe, not grade, and when you mention an interview or a talk with your boss, Evoke asks afterwards how it went. The question comes right in the notification, "How did the talk with your boss go?", and follow-ups for any topic can be turned off in one tap.
Decision 02
Most users aren't ready for a therapist's office; the chat is where they find out. When a topic keeps coming back, Evoke suggests a specialist in the chat, with prices one tap away, and the user approves the summary the psychologist sees. In a crisis the card leads with a crisis line. All four psychologists signed off on the flow.
Decision 03
For evenings when people don't want to talk. Exercises are sorted by how you feel, not by technique. Photos won for one recommended exercise but turned into noise in a list of 80, so the mixed version stayed. Recommendations are signed "Rekomendacje Evoke", so the app speaks in one voice.
Decision 04
The MVP counted "4 of 200 messages" above the chat and cut people off mid-thought. v2 plans count conversations, and a started one can always be finished; the upgrade comes after it ends. 3 of 8 testers couldn't tell the plans apart, so the comparison now leads with what changes for you.
Decision 05
The MVP asked for a login before the first word, and 38% left right there. In v2 one screen comes first: what Evoke is, that it isn't a crisis service, and consent to process what you share. Then the conversation starts, and saving it with Apple, Google or email comes after. In testing, a sign-up prompt after three messages felt like a paywall, so it moved to the end of the first conversation.
Mockups, specs and a Material UI theme went to development with psychologist-reviewed scripts. v2 is being built now, and the site is collecting a waitlist.
v2 isn't live yet: the live MVP against the v2 prototype, same five tasks, 8 participants each. Screens are counted in the flows.
Screens before the first message:
5to1
Explained correctly what the paid plan adds:
3 of 8to7 of 8
Task success, reaching the booking screen from a chat:
2 of 8to6 of 8
Task success, finding an exercise for how you feel:
4 of 8to7 of 8
Average task ease, SEQ 1–7:
4.1to5.8
SUS score:
58to76 (+18 points)
v2 cohorts from the first three months against MVP cohorts of the same length.
30-day retention: 3.1% → at least 5%, the primary goal
First message sent: from 54%
Follow-ups opened and returns within two days: new, no baseline yet
Upgrades within 30 days: from 0.8%
Specialist bookings: from 0.4%
Conversations cut by the limit: from 17% to none
Bondata
An employee-engagement analytics platform for HR leads and managers
(view)
Dishcover
A culinary app with easy, step-by-step recipes for every taste
(view)
Your Vet
In progress
