Only activation designer for two years. Accountable for the OKR, not a backlog.
I was the only product designer on Conta Azul’s activation team between 2023 and 2025. I mapped the pain points of the first-time experience screen by screen with CS, facilitated 2025 prioritization with the GPM and the tech lead, and designed the initiatives: 31 over two years, measured in activation, revenue, tax document issuance and feature discovery.
Strategy
A solution existed that worked. It didn’t scale.
30-day activation was at 30%, against a 60% target. Customers who went through CS-assisted onboarding activated more: 1.56x in invoice issuance, 2.03x in reconciled charges. But CS only reached 7.7% of new customers (472 of 6,111, from January to April 2024), and the customer base was growing faster than the team.
Attacking one screen at a time would barely move the needle. I stopped treating activation as a feature queue and started treating it as a system, with one question: how do I make the product do on its own what CS did?
Process
Hypothesis and success metric before any screen.
Before proposing a solution: customers, to find the real pain points; the product from the inside, to understand the business; and operations teams, who saw friction no dashboard showed. AI in discovery and synthesis, to gain speed. Every initiative starts with a hypothesis and success metric defined before design.
Hypothesis-Driven
Mixed-Methods
Metrics-Driven
AI-Augmented
Diagnosis
The problem wasn’t the screen. It was the onboarding model.
What the business called activation was, at the same time, a measurement problem, an organization problem, a product problem and a data problem. Nine methods, in the order they happened: stakeholder interviews, correlation-based redefinition of activation, AI-assisted secondary research, a 20+ platform benchmark, CS observation and user interviews, support tickets, flow mapping across four scenarios, service blueprint and native onboarding mapping. They produced four findings.
Evidence: Group and 1:1 onboarding sessions observed without intervening, interviews with users at different stages, support tickets, and 12,887 meetings with leads and customers, transcribed and synthesized with AI.
Finding: Every session started with the same five questions, and the product didn’t answer any of them on its own. The first 30 to 45 minutes went to registration data entry. In the tickets, NFS-e prerequisites were the main driver of contact in the first 30 days. The customer didn’t know why to do it, where they were, or how to do it. CS solved all three for them: entered the data, said what to do, and showed where to find it.
CS observation: sessions mapped to the journey stages
Initial hypothesis: Improve the first-steps screen, eleven setup steps: fewer steps, a progress bar. That’s where I started.
Evidence: In the screen-by-screen mapping with CS, across four entry scenarios (new trial, direct purchase, end of trial, returning former customer), most people skipped the step-by-step flow. The list ignored the interest declared at signup and asked for eleven setup steps before the first result. New users went through 4+ handoffs between teams, and native onboarding flows fired simultaneous triggers with contradictory language.
Decision: No adjustment to the screen solved it. I proposed removing the screen, moving the essential questions into signup, and taking guidance into the product itself, at the moment of the action. The negotiated middle ground, an optional First Steps 2.0, was still in testing when I left.
Current-flow mapping: four entry scenarios and failure points
Evidence: Interviews with Product, CS, Marketing Lifecycle and Data Science, with three questions: what activation means to you, why users don’t activate, what’s already been tried.
Finding: Each team measured something different, and every initiative was optimizing for a different goal. The official yardstick measured setup (digital certificate, municipal approval, business account), not value. Entire segments that would never issue an invoice or a charge counted as not activated by definition.
Decision: I took the argument to Growth, CS and Data leadership, who started a correlation study between feature usage and retention above 12 months. Activation became a set of criteria covering the entire journey: signups, account integration, sale, reconciliation and cash flow. That justified tackling several points at once.
Evidence: Service blueprint across the entire lifecycle: user actions, touchpoints, frontstage, backstage, systems and teams. In parallel, a 20+ SaaS platform benchmark (Claude for the tables, Mobbin for the screens) and secondary research on activation, trial and empty states.
Finding: First artifact where every team saw the full journey together. More than 200 opportunities came out of it, prioritized by frequency, severity and ownership. From the benchmark, three patterns: sample data eliminates empty-product paralysis; contextual activation beats checklists; segmented onboarding completes 30–40% more than linear tours.
Service blueprint: complete activation journey
Solution
The product started doing what CS did. One initiative per function.
From the 200+ opportunities came 31 initiatives over two years, prioritized collaboratively. The six with evidence, grouped by the CS function they replace:
CS filled in the data for you
Before: Before the first sale, the user needed to register customers; before the first invoice, services. Manual data entry was the wall between creating the account and seeing value.
Decision: I explored three models: entry through setup, integrated into onboarding, and conversational chat. The MVP simulates the chat through actions: the user uploads a spreadsheet, the AI extracts and maps the fields, and they review and approve before anything is saved. Hours of data entry became minutes, and the model became the foundation of the product’s AI.
AI-assisted import: spreadsheet upload and data extraction
Before: First-time users opened the product and saw nothing: empty states announced absence without saying what to do, and the main action button went unnoticed.
Decision: I redesigned every empty state with the same structure: what the area is for, what the first action is, a direct CTA. I separated four situations the product treated as one (true empty, filtered empty, permission-based empty, no search results) into design system components. And I rebuilt +New with more contrast and a menu that surfaces Product and Service upfront.
Empty states: all redesigns, same structure
+18%Sales created
+20%Invoices created
+12%New sale started
+20%New invoice started
A/B test: Oct–Nov 2024
1,392Sales startedExclusively via the new entry points
New entry points, Dec 2024, 10 days · absolute volume, not A/B
Before: Help only appeared at login, before the user knew what they needed, and there was no way to revisit it.
Decision: I replaced it with banners that appear on first entry into each area. Each one answers the question at that point, mapped from the most frequent onboarding pain points.
Contextual banners on first entry into each area
+21%Total feature usage
+109%CTA clicksOverview · first days
+19%Revenue creationFirst days
797Validated interactionsContextual help · financial accounts
Before: CS was the onboarding: every new user depended on a scheduled session, and the user base grew faster than the team could absorb.
Decision: Instead of copying the linear session onto the screen, I designed 12 independent flows in Appcues, triggered by the context of the action and split by business type (services, commerce, retail). Rollout to 100% of new services users in November 2024.
Onboarding at scale: segment-guided flow
+40%Sales created
+37%Invoices created
+35%Invoices started
+20%Sales started
+19%Invoices sent
A/B · Services Cluster · Oct–Nov 2024 · impacted base: 50% of new services trials
+130%Sales created
+116%Invoices created
+112%Invoices initiated
+89%Sales initiated
+84%Invoices sent
Segment: users who completed at least one of the 12 flows · rolled out to 100% of new services users on Nov 4
CS showed you where to find things
Before: Critical functions were buried in a menu that showed everything to everyone. The ideal search, a single field across the whole system, would require refactoring a fifteen-year-old frontend that no one was going to prioritize.
Decision: I chose a layer on top of what existed, Spotlight-style: index what already exists without changing the old screens. I delivered it in phases, measuring before opening the next one: MVP, V2 with descriptions, V3 with iconography and classification by type. Alongside it, I reorganized the first level of the menu and simplified the top bar.
Global search: contextual results
+600%Access to reconciled Pix
+233%Access to issued Pix
+100%Access to payment link
+41%Access to purchases
+37%Access to sales reports
Small base · functions almost no one found: the number proves discoverability, not interest
Before: More than 80% of people who went through 1:1 training with CS activated. But scheduling depended on CS reaching the user, and almost no one knew the free training existed.
Decision: On first access, the product started offering scheduling for a free session by video call, with screen sharing. Instead of CS chasing the user, the user finds the training at the right moment.
Training scheduling: invitation and platform flow
Impact
With no CS in the path, activation went from 27% to 40%.
Per-initiative metrics are above. Here, the indicator and the organization.
Activation went from 30% to 51.6%
The rate hovered between 28.8% and 36.2% when I started; the target was 60%. In October 2025 it reached 51.6%. I didn’t hit the 60%, and the yardstick changed along the way. The new segment measures only what the product does on its own: subscribers who activated at least five criteria without going through CS went from 27% to 40%. No single initiative explains the movement; it comes from the set.
New team created
The 200+ opportunities from the blueprint became the backlog of a new team, dedicated to experience quality and owning parts of the product that previously had no owner.
CS freed from onboarding
Standard onboarding started working without human support and without growing the team. CS moved to complex accounts, at-risk customers, and specialized support.
Reflection
What this project was
The brief was activation: a metric to raise, a list of initiatives to deliver. The work that weighed most came before any screen: slowing down, mapping the system, and turning CS’s interventions into product capabilities. AI covered ground a whole team would (thousands of meetings listened to, tickets synthesized), and the method was documented and adopted by other designers.
What I’d do differently
I’d pull the data team in during the first week and define how to read the result before designing. I’d write the experiment calendar together with the backlog: twelve fronts running in parallel without governance cost us the ability to read the whole. One of the fronts was paused for colliding with another team’s test, and I lost the ability to isolate each initiative’s contribution to the 72%. I stand behind each number in the metric it measured. How much each one was worth in the aggregate, I won’t estimate.
Contact
Let's work together?
Available for full-time · remote · freelance
Open to new projects, opportunities and conversations.
