Al health coach
Simple
1 May, 2025
As Design Owner of Simple’s AI Coaching, I turned Avo from an AI feature into a core retention loop, shaping activation, contextual guidance, daily check-ins, plans, and proactive coaching that gave users a reason to return and keep progressing. It scaled to millions of conversations and became an award-winning virtual coach
Timeline
2023–2025
Role
Senior Product Designer
Scope
UX/UI, AI Interaction Design, Design System
Product
1
N
, B2C, Wellness
Surfaces
iOS, Android
Simple is a health and weight loss app backed by Palta (Flo Health, Lensa, etc). It’s one of the leaders in the weight-loss category, with 15M+ downloads, $160M+ ARR, 400K+ MAU, and users across the globe.
Food tracking, which I had worked on earlier, answered the first growth question: what comes after the fast? It gave users more reasons to engage and gave us richer behaviour data. Users could log their activity and get scored, but they still needed help understanding what it meant and what to do next.
The challenge
Simple had already squeezed a lot out of fasting and tracking. The business needed even stronger retention, stickiness, and clearer differentiation from cheaper fasting and calorie-tracking apps to keep growing.
The challenge was to create the connective tissue between disjointed trackers so the app felt coached and holistic, not like a dashboard of tools. The product needed to become more valuable than the sum of its parts – and finally deliver on the coaching promise the company had set for itself.
Exploration
The first versions of Avo were still too close to the obvious AI pattern: a chat entry, a few suggested prompts, and a legal flow users had to accept before they got any value.
That was useful enough to prove there was something there, but not useful enough to become a product loop. Avo had to become easier to enter, more contextual, and more connected to the actions users were already taking inside Simple.
I started with the entry problem: how do we make Avo visible without making it feel like another banner? The early work covered Home entry points, unboxing, consent, chat history, suggested prompts, and prompt categories. Suggested prompts mattered more than they looked. Most users don’t know what to ask an AI coach, and three useful questions on day one beat a blank chat box.
The best iOS unboxing variant lifted Chatbot Open by +146% and opens per user by +80%. Across the broader Q3 Avo push — unboxing, UX, passive triggers, and suggested prompts combined — new users opening Avo grew +83% and completion grew +124%.
Useful in context
Once Avo had a path in, the next question was where coaching should actually appear.
The fastest way to make AI feel useful is to move it to where the user already is. Avo had been sitting in chat, waiting. I pushed it into surfaces with existing intent: food and weight first. After a meal log, Avo had something useful to say about it. After a weigh-in, it had context for the trend.
Passive Triggers 2.0 made those entry points more relevant. CTR moved from 8.5% to ~10%, rating improved from 4.43 to 4.55, and messages per dialog grew from 3.2 to 5.5. Median time in Avo nearly doubled — from 25s to 47s.
Smart Meal Feedback and Smart Weight Feedback pushed the idea further: show a useful bit of advice before the user opens chat. Smart Meal Feedback lifted Avo dialogs per user by +9.5%, while Smart Weight Feedback improved weight tracking — Weight Track RR 1D +5.9%, Weight Tracks/User 7D +5.5%, CR to Weight Track 7D +5.4% — and rolled out on Android.
The honest read: contextual coaching reliably moved engagement and behaviour. It did not move product retention on its own. Distribution made the AI more useful, but it did not magically fix the upstream LTV problem.
Daily ritual
Coaching only works if it shows up every day. So the next move was to turn Avo from Q&A into a daily plan loop.
It started as a simple check-in: how was yesterday, what’s the plan today. Then it grew into a recap, plan reasoning, daily tasks, and a Home stack that held it all together. Avo was no longer just a place to ask questions. It became the thing that helped users reflect, plan, and come back.
R2 Home made daily coaching dramatically more visible: D1 retention +6.3%, 7/7 active DAU +8.8%, check-in completion +144%, and tracker engagement up across meal, weight (+18%), and activity (+39%).
Check-in R1 had already shown the loop had legs: D3 retention +3.5%, Avo D1 retention +6%, dialogs per user +12.6%, and check-in D3 retention +21.7%. Card Check-in V2 later made the experience more structured and practical, lifting Daily Tasks Done/user +10%, Weight Tracks +6%, Activity Tracks +7%, and Recap CR +27%.
The important decision came when we tested splitting tasks from the check-in. Tasks alone improved — task completion +30%, Tasks/user +60% — but the broader loop got worse: Check-in RR −10%, Avo RR −5%, Product RR D3 −1.6%.
So we kept them connected. Tasks worked best when tied to the coaching context, not extracted from it. A winning metric inside one component was not a win for the product.
Proactive coaching
A coach who only speaks when asked is not really a coach.
The first proactive layer was check-ins and Home cards: yesterday’s recap, today’s plan, streaks, and useful chunks of coaching delivered before the user opened chat. Recap in Card Check-in V2 alone drove +27% CR.
The second layer pushed further: Avo Patterns. I drafted the PRD and shaped the calibration → analysis → habit challenge → completion loop with PM, science, and copy. Avo watched behaviour over time — late-night snacking, low protein days, missed hydration, processed-food creep — and surfaced recurring patterns as something to act on.
The lightweight proactive cards landed. Patterns didn’t.
The experiment showed no product retention impact across iOS new, iOS existing, and Android existing. Only ~30–40% of users received a pattern in 7 days, and challenge completion ran ~5–13% on iOS. Final recommendation: do not proceed with rollout.
Users who engaged deeply found real value — better awareness, motivation, habit change — but discovery and placement failed the idea. It was still important work. It pushed Avo from reactive chat toward proactive coaching, and showed where the idea outran the execution.
Making Avo feel like a coach
By the time Avo was becoming central to the product, the chat surface itself was the weakest link.
The debt was real: inconsistent navigation, heavy chat bubbles, long unscannable responses, awkward scrolling, weak hierarchy, generic chatbot vibes. I initiated and led the cleanup as a product-quality foundation, not as an experiment.
The shipped cleanup made Avo more readable and app-native: larger type, better line height and spacing, fewer and lighter bubbles, real-time response animation and haptics, app-aligned navigation, blurred nav bars, standardised rating and error states, and a unified single-chat experience with Avo moved into the bottom nav.
There was no isolated A/B test. This shipped as the foundation that every later Avo feature — triggers, check-ins, patterns, avatars — inherited. It made Avo readable, persistent, and scalable. That was the point.
Some of the richer interaction ideas stayed as concepts. That was fine. The right deliverable here was not a flashy future-chat prototype. It was a cleaner, more durable coaching surface the rest of the product could build on.
Coach identity
Utility gets you a useful tool. Identity gets you a relationship.
Once Avo was doing real coaching work, it needed a face. I shaped the avatar system, coach personalities, and vibes with an external AI generation specialist, exploring how a coach could feel chosen and recognisable without tipping into childish, fake, or uncanny.
The interesting constraint was research-driven. Coach Avo Avatar research showed realistic, human-like avatars build trust — but risk misleading users into thinking Avo is human. Age-appropriate, diverse coaches mattered. Some users actively preferred slightly non-human avatars to keep the AI nature explicit.
The shipped picker offers 13 coaches plus Classic Avo. Selecting one plays a short video, locks in a per-avatar tone of voice in the system prompt, and gives the coach a distinct identity. Avatar customisation rolled out to iOS new users.
Early signals were promising: Chatbot Rating D1 +2.4%, Avo Retention D1 +3.6%, Streak Progress D1 +3.1%. The All Hands readout flagged +6% D7 retention and +2.5% chatbot rating as proof the concept worked.
Conclusion
Tracker retention
+34%
Cumulative
Product retention
+8%
Cumulative
British Dietetic Association
Innovation Award
Food Score and Feedback
Avo became a layered coaching system: activation, contextual coaching, daily planning, proactive intelligence, UX maturity, and identity.
Activation: Chatbot Open +146% in the best iOS variant; opens/user +80%.
Contextual coaching: Passive 2.0 moved CTR from 8.5% to ~10% and messages/dialog +72%; Smart Weight Feedback improved Weight Tracks/User 7D +5.5% and rolled out.
Daily check-in: R1 lifted D3 retention +3.5%; R2 lifted check-in completion +144% and 7/7 DAU +8.8%; Card V2 lifted Recap CR +27%.
Proactive coaching: recap drove +27% CR; Patterns were tested and not rolled out.
UX cleanup: shipped as product foundation; not A/B measured.
Coach identity: All Hands flagged +6% D7 retention and +2.5% chatbot rating; avatar customisation rolled out to iOS new users.
Avo became Simple’s core coaching surface, named Best Virtual Health Coach in the 2025 and 2026 MedTech Breakthrough Awards, alongside Johnson & Johnson, OURA, Medtronic, and Teladoc. By early 2026, Avo handled around 5M monthly dialogues and 23M monthly messages.
Worked: contextual triggers, daily check-in loop, R2 Home, proactive recap and plan cards, UX cleanup foundation, avatar picker.
Learned: Patterns were strategically right but operationally weak — adoption was too low and retention didn’t move. Tasks worked best inside the coaching loop, not separated from it.
Cut: splitting check-in from today’s plan; Avo Patterns rollout; some richer interaction concepts that didn’t justify the engineering cost.
Avo scaled, won awards, and improved engagement and product-retention signals across activation, daily loop, and contextual coaching. It did not single-handedly solve Simple’s broader LTV ceiling — that was a structural product question, not a coaching one. Knowing the difference is part of the job.
What I’m proud of is that we turned a generic AI surface into a real product system. Avo became easier to enter, more useful in context, more present on Home, more readable as an interface, and more ownable as a coach.
It wasn’t one magic redesign. It was layers.
Acknowledgements
Built remotely with a team spread across the world: Elena Deshina, Alisa Korchazhnikova, Ivan Bakanovskiy, Val Scholz, Dimitri Nikogosov, Ro Huntriss, Josie Porter, Karina Delaine, Dmitrii Mochalov, Ruslan Dzhafarov, Alex Ovs, Hemank Sabharwal, David Johnston, and Alesia Privado.
Special thanks to Hemank and Alex. Working with you made this whole project genuinely fun.