Two product challenges. One platform. From broken onboarding to conversion engine.
AI & Experience Strategy Lead for ENSO Co-Living across four cities in Spain and Portugal. Mapped the service end to end, operator backstage included. Phase 1: redesigned the AI enhanced onboarding funnel, killed 60% of planned scope, cut friction 22%, 4x adoption against projection. Phase 2: built a risk-calibrated urgency system for the catalogue that lifted booking conversion +34% (A/B test, 14,208 users).
Co-living onboarding across four cities in Spain and Portugal.
ENSO Co-Living operates in Lisbon, Valencia, Madrid and Barcelona, matching tenants to shared apartments based on lifestyle, language, budget, and timing. When I joined, onboarding was 47 spreadsheets, 3 workflows across four cities, zero scalability.
ScopeRole: AI & Experience Strategy Lead Scope: Onboarding & qualification funnel Markets: Lisbon, Valencia, Madrid, Barcelona Work: Research, service design, IA, interface design, design system
02 The Real Problem
Everyone assumed the existing workflow was correct and just needed to move to digital.
Two weeks of operator shadowing proved that wrong. Operators weren't following the spreadsheets. They'd built informal shortcuts. The spreadsheets documented an ideal process. The operators had evolved a real one.
47 spreadsheets, 3 different workflows across four cities, with no scalability
The qualification bottleneck was trust, not data. Operators needed to feel confident about a person before showing them rooms
~48 hours to first reply. The process gathered data but didn’t answer the trust question
03 Key Insight
Operators were making 3 decisions per applicant, not 47.
(1) Is this person real? (2) Do they fit an available room? (3) Would they be a good roommate? Everything else was administrative noise accumulated over two years of patching.
EvidenceOperator shadowing in Valencia. Quote from operations manager: “I ignore half the spreadsheet. I just need to know: can I picture this person living with the current tenants?”
"I ignore half the spreadsheet. I just need to know: can I picture this person living with the current tenants?"Valencia operations manager, during shadowing
Trust map: who needs to trust whom before a room fills
Operator, per city Phase 1Is this person real, do they fit a room, will they be a good roommate?The gatekeeper. Owns the relationship with every flat. Was answering all three questions off the record, in WhatsApp threads and informal calls.
Current housemates Represented by the operatorWho is moving in with us?Never in the funnel, always in the decision. The operator’s real test was whether they could picture the applicant living with them.
Applicant Phase 2Is this room worth committing to, and is what I’m seeing true?Booking a home in another city, often from abroad and on mobile. Any signal that looks manufactured breaks the deal.
ENSO Phase 2How fast does an empty room fill?Every empty room costs about €4.2K a month. The business clock runs underneath every other relationship.
Phase 1 answered the operator’s questions. Phase 2 answered the applicant’s, under the business clock. Both phases are the same service seen from two sides.
04 Service Blueprint
Same five stages. Two completely different services.
Before designing a single screen, I mapped the shadowing into a service blueprint: what the applicant does, what they see, what the operator does behind the line of visibility, and which systems hold it together. Drawn this way, the problem stopped being “too many spreadsheets”. The trust signal operators relied on lived entirely off the record, in WhatsApp threads and ad hoc calls, so it could not be measured, shared or scaled.
The redesign kept the stages and moved the trust signal into the product. Toggle between both versions.
Service blueprint: applicant to room catalogue
01Reach out
02Share context
03Show who you are
04Get qualified
05See rooms
Applicant
Emails or fills a generic web formRegisters with 4 fields, or signs in with Google
Answers follow-up questions in WhatsAppSets city, dates and living style
Joins a video call, if the operator finds a slotRecords a 1-minute video intro, on their own time
Waits, with no visible statusSees a clear review step, not silence
Gets a first reply after ~48hLands in a room catalogue for their city
Line of interaction
Frontstage
Inbox and web formMobile first registration
WhatsApp threads per operatorPreferences screen
Ad hoc video callIn-app video recorder
NothingSuccess and status screen
Manual message or emailPersonalised room catalogue
Line of visibility
Operator backstage
Screens each inquiry by handNothing to do: data arrives structured
Re-keys answers into a spreadsheetNothing to re-key
Judges fit off the recordWatches the video when it suits the day
Qualifies against 1 of 3 city workflowsOne async review: approve or reject from a single applicant view
Matches rooms through informal shortcutsCity logic applied at matching, not by hand
Line of internal interaction
Support systems
47 spreadsheetsOne flow, one codebase
No shared applicant recordOne shared applicant profile
No trace of the trust signalVideo stored with the profile
3 workflows across 4 citiesOne design system
No funnel analyticsPostHog and Mixpanel on every step
~48h to first reply. Trust decided off the record.Under 5 minutes from application to room catalogue. Trust decided in the product.As isTo be
Red: fail points found in shadowing. Teal: the two moments of truth the new service is built around. Scroll sideways on mobile.
05 What I Changed
Three interventions. Each targeted a specific trust question.
1. Redesigned onboarding as a 5-step funnel answering 3 trust questions
Mobile-first wireframes. Registration → Google flow → Preferences → Video qualification → Success → Room selection.
2. Replaced 12 personality form fields with a single 1-minute video recording
Operators were already doing informal video calls. I formalized and made it async. Video gives more signal in 60 seconds than the entire form.
Before: Registration
Before: Qualification
3. One flow for four cities
City-specific logic pushed to room matching step, not registration. One design system, one codebase, one analytics pipeline.
The roadmap had 13 workstreams. I shipped 5. The 8 I killed would have delayed launch by 4+ months and none of them addressed the core trust problem.
06 The Operator Side
The applicant flow takes five minutes. The operator flow is where trust is decided.
Most onboarding work stops at the applicant’s screens. Here the critical journey was backstage: one operator per city reviewing every applicant on their own schedule. I designed the review as a single view that answers the three trust questions in order, with an explicit outcome for every answer, so no applicant ends in silence.
Operator review: one view, three questions, four outcomes
1
Is this person real?Registration data and sign-in arrive verified and structured. Nothing to re-key, nothing to chase.
2
Do they fit an available room?City, dates and living style sit next to the inventory they could match. The operator judges fit, not data quality.
3
Would they be a good roommate?The 1-minute video replaces the informal call. Same signal operators already trusted, now async and attached to the profile.
ApprovePersonalised catalogue for their city, ready to book.Happy path
Need more signalAsk for a new video or a short call, then back to review.Recovery path
No room fits yetProfile stays qualified and resurfaces when a matching room opens.Feeds the catalogue loop
Not a fitA clear, respectful close. No ghosting.Closure path
Why the operator stays in the loopI also explored AI powered roommate matching to automate step 3. Operators rejected it: it took away the judgment their relationship with each flat depends on. The final design uses the product to prepare the decision and leaves the decision with the person accountable for it.
07 Key Decisions
What trade-offs did I make?
Every decision involved choosing between digitizing the existing process and redesigning for the real one:
Decision
Chosen
Rejected
Why
Video intro
1-min self-recorded video
12 personality form fields
Operators said video gives more trust signal in 60s than entire form
Registration fields
4 fields (name, DOB, phone, language)
14 fields
-22% friction at highest drop-off point
Multi-country
Single flow, city logic at matching step
Country-specific onboarding flows
One system to maintain, not three
Scope
5 workstreams shipped
13 workstreams planned
8 killed items didn’t address core trust problem
08 Phase 1 Results
Scope discipline wasn't a constraint. It was the strategy.
-22%Onboarding friction reduced
4xAdoption vs projection
<5 minApplication to room catalogue
4Cities, 1 flow
ValidationPostHog + Mixpanel instrumented from day one. Every step transition, drop-off, and time-to-complete measured from launch. Video step had 18% lower drop-off than predicted, because users who reached it were already self-qualified.
09 The Service Loop
Onboarding and the catalogue are two ends of one loop.
Seen as screens, Phase 1 and Phase 2 look like separate projects. Seen as a service, they are the same loop. Every room moves through a cycle from the day it gets a target move-in date to the day someone moves in. Phase 1 built a steady supply of trusted, qualified applicants. Phase 2 pointed that supply at the rooms costing the most.
Room lifecycle: tap a stage
01
10 Phase 2: The Conversion Problem
Every empty room is burning money. Static listings weren't helping.
With onboarding solved, the next bottleneck was clear: the catalogue. Every habi costs ENSO ~4.2K€/month when empty. Our listing page treated every room equally. A high-risk vacancy that needed to fill this week looked identical to one available three months out.
No urgency differentiation. No conversion pressure calibrated to actual business risk. In Q3 alone, 23% of rooms went past their target move-in date.
HypothesisIf we surface the right urgency signals at the right time, we can accelerate bookings for high-risk inventory without degrading trust.
11 Risk Framework
Five risk tiers. Progressive urgency, calibrated to real risk.
Five risk tiers drive the progressive addition of urgency elements. Each tier maps to days remaining before vacancy becomes a financial liability. Levels 1 to 4 are additive: each inherits the elements of the level below. Level 5 does not add pressure: it swaps the heaviest perks for a flexible entry date.
Three design principles: progressive disclosure (start clean), truthful signals (real dates, real deals, real perks), scarcity not pressure (context, not panic).
Try it: move the vacancy clock
overdue0153045 days
40 daysL1 Calm
Available 15 Dec3 people viewingENSO DealLast room in this flat
Double room, shared flatValencia, 4 housemates
€640€640 / month
+ Free cleaning+ No exit fee+ Free membership+ Flexible entry date
What the tenant sees. Illustrative card and values
Five risk tiers: from calm (30+ days) to critical (overdue). Each tier progressively activates urgency elements.
The same room card at each risk tier. Visual density increases only when warranted by actual vacancy risk.
Annotated card anatomy. Every element is conditional, activated by the risk engine.
All five risk levels with design rationale. Level 1 is clean and minimal. Level 4 carries the full weight of urgency signals, deals, and perks.
The urgency system as a service: four layers, one feedback loop
Experience
City landingRisk calibrated cardRoom detail confirms perksPerks in checkout summaryBooking
↓ renders
Business logic
Risk score per roomTier assignmentSignal composer, additiveServer side card template
↓ reads
Data
Inventory and target move-in datesPricing and tier discount rulesFunnel and experiment data from PostHogDaily risk recalculation15-minute cache
↓ measured by
Feedback
Feature flagged A/B testCTR, booking rate, time to bookPer tier, city and deviceGuardrails: cancellations, tickets, NPS, discount depth
No risk logic runs in the browser. The card only shows what the data layer can prove, which is what keeps every signal truthful.
The booking decision flow. From city landing to checkout, urgency signals compound across every step.
Mobile screens: full urgency listing, filtered results, and empty state handling.
12 Validation
A/B experiment: 14,208 users, 28 days, PostHog.
Controlled experiment. 50/50 split. Variant A (risk-based cards) vs Control (static cards).
Card Click-Through Rate+45.8% uplift
Control
8.3%
8.3%
Variant
12.1%
12.1%
p < 0.001"View rooms" clicks / total card impressions
Booking Conversion Rate+33.8% uplift
Control
2.4%
2.4%
Variant
3.2%
3.2%
p = 0.003Completed bookings / unique listing viewers
Median Time to Book-38.7% faster
Control
6.2 days
6.2d
Variant
3.8 days
3.8d
p = 0.008Days from first listing view to completed booking
Users14,208
Duration28 days
Split50/50
SignificanceDay 16
PlatformPostHog
13 Funnel Analysis
Where the uplift compounded
Card click-through was the primary lever (+45.8%), but gains compounded through every step downstream.
5-Step Conversion Funnel
ControlVariant (risk-based cards)
Card impression
100%
100%
baseline
↓
Card click
8.3%
12.1%
+45.8%
↓
Room view
6.1%
8.4%
+37.7%
↓
Checkout start
3.8%
5.1%
+34.2%
↓
Booking
2.4%
3.2%
+33.8%
Key insight: The uplift didn't just happen at card level. It compounded through every step. Room view rate +37.7%, checkout start +34.2%. The urgency signals pre-qualified intent. Users who clicked were more likely to book.
14 Guardrails
Urgency only counts if it doesn’t cost trust later.
A conversion lift that turns into cancellations, tickets or regret is just a delayed cost. So the service had to be measured past the booking, at the moments where manufactured urgency usually shows up.
Early cancellation rate4.8%4.6%No increase in people backing out after booking.
Support tickets per week1211Signals were clear enough not to create questions.
Post-booking NPS6265Slight improvement. No booking regret.
Average discount depth13.6%Within margin. Monitored per tier.
Control vs variant, same 28-day experiment.
The test that shaped the principlesEarly explorations used countdown timers. Four out of five testers said it felt like a hotel booking site pushing them. That is where “scarcity, not pressure” came from: show what is real, like real deals and real availability, and never invent time pressure.
15 Phase 2 Results
Result of the A/B test
+34%Booking conversion
ValidationPostHog A/B test. 14,208 users, 28 days, 95% significance on Day 16. All guardrails passed. No increase in cancellations, support load, or booking regret.
16 Reflection
Once I framed the process as 3 trust decisions instead of 47 data fields, the right product became obvious.
The patternThe brief is almost never the problem. "Digitize 47 spreadsheets" was a symptom described as a task. Scope discipline wasn't a constraint. It was the strategy. Shipping 5 things that matter beats shipping 13 things that cover the roadmap.
What failedAI powered roommate matching failed because operators felt it undermined their trust relationship with tenants. "Smart" features that remove human judgment from high-trust decisions create resistance, not efficiency. Also underestimated design system investment, and custom components couldn't be reused when supporting the expansion to London.
One signal I would rebuildThe test measured the whole bundle, so I cannot say how much of the lift each signal carried. The one I would rebuild first is the viewer count: today I would use signals that stay true when nobody is watching, like calendar scarcity, saves this month or a deal with a real end date.
What I'd do differentlyStart the blueprint with the operator, not the applicant: the backstage is where the service was really broken. Design mobile only from day one, since 73% of catalogue traffic was mobile. Track cross-device journeys from launch. And test urgency signals one by one instead of as a bundle, to know which ones carry the lift.