Conta Azul · 2024–2025

Removing the ERP's main activation barrier with AI-assisted data import

AI-Augmented
Onboarding
Activation
Revenue
Conversational UI

Context: The most consistent finding from the activation journey: before creating the first sale, the user needed customers already registered; before issuing an invoice, services already registered. Manual data entry was the wall between account creation and product value.
Problem: Manual entry of setup data consumed 30–45 minutes of every new onboarding session, all spent before any value-creating action. For most users, that friction was enough to cause abandonment before activation.
Approach: AI-assisted conversational import flow: the user uploads a spreadsheet, the AI extracts and maps the fields automatically, the user reviews and confirms. Setup in minutes, not hours.
Result: +44% feature interest · 22% active engagement · 16% flow completion. Early-stage validation signals, not full rollout.

Results

+44% Feature interest Declared interest
22% Active engagement Started the flow
16% Flow completion Completed the import

Role

I was the Product Designer responsible for this initiative, from initial alignment through the MVP Alpha prototype. The scope included leading the CSD with two product PMs, exploring three UI alternatives, and developing the prototype in direct partnership with the Design Manager.

The exploration of alternatives went through engineering, directors, GPMs and PMs before closing. There was one criterion: the path had to converge with the business’s long-term vision, and it did. AI-assisted onboarding became Conta Azul’s first step in delivering customer value with AI inside the product.

The decision to go with the Chat format (instead of a Wizard or integration with Captura AI) went beyond an interaction decision. It meant handing the scope to the Conta AI team, with support from the Growth team. I navigated institutional uncertainty about how AI would be used in the product: the project documentation noted that this flow could see unplanned changes as Conta Azul Pro’s AI vision evolved. Moving forward with MVP Alpha, instead of waiting for a complete definition, was a calculated risk decision.

Strategy

The OKR was direct: raise 30-day activation for subscribers without assisted onboarding from 27% to 40% (excluding the bulk channel). The core metric was Time to Value, the time until the user recorded their first sales data. Manual registration entry was the biggest blocker on that path.

In ERPs, the standard approach to this problem is CS-guided onboarding: a human walks the user through manual data entry. It scales poorly, because CS cost grows with the user base, and that was exactly what we wanted to replace in the segment without assisted onboarding.

The AI performed one specific mechanical task: mapping the fields in the user’s spreadsheet to the product’s fields. Pattern recognition, not judgment, and the AI solves it in seconds. Separating roles was the most important design decision: the AI handles the technical mapping, the user validates the business data. That resolves trust (the user controls what enters the system) and friction (the user doesn’t do the mechanical work) at the same time.

The real tension was timing: AI integration into Conta Azul Pro was still being defined while we were designing. Choosing Chat anticipated the Conta AI vision, but depended on a product decision still open. I designed MVP Alpha to generate signal before that decision became a blocker.

Process

The sequence was defined before any wireframe: align on uncertainties, research existing solutions, explore models, decide with explicit criteria.

CSD with the product team: a structured survey of Certainties, Assumptions and Doubts with the two PMs. Result: a list of hypotheses about desire, usability and perceived value that shaped the MVP’s learning expectations.

Internal desk research: mapping of previous initiatives in the Inbox and Conta AI developments documented by other designers. This avoided rework and connected the initiative to the long-term vision already present in the product.

AI in SaaS benchmark: a study focused on how interfaces solve AI integration for file import in B2B products. It identified that file import as a core feature was the standard for delivering perceived value quickly.

Exploration of 3 UI models: prototyping and comparing: (1) an Import Wizard (an experience separate from Conta AI); (2) Integrated into Captura AI (within the existing visual experience); (3) AI Chat (a static chat experience, no text input, clicks only).

MVP Alpha with a restricted audience: launched to new users across roughly 10% of the eligible base, covering registration of Customers, Products/Services and Expenses. Learning expectations were defined before launch: Desire (% adoption, main point of abandonment), Usability (understanding of the flow, clarity of the value proposition, ease of review) and Perceived Value (qualitative feedback on the AI experience).

Diagnosis

Solution

Reflection

What this project actually was

A bet on timing: designing trust for AI in a context of critical data before the product’s AI vision was fully defined. The technical design work (splitting AI/user roles, the review screen, the static chat) was a consequence of an earlier strategic decision: generate a validation signal before institutional uncertainty about AI in the product became a blocker.

Going from 27% to 40% activation in 30 days wasn’t going to come from UX fine-tuning. The jump required changing the structure of the problem: removing manual data entry from the path between account creation and product value.

What it meant for the product

This project was Conta Azul Pro’s first AI case delivering direct customer value in production. The MVP Alpha result became a reference for leadership deciding Conta AI’s next steps: which interaction format to scale, where to invest engineering, how to measure adoption in flows that touch critical business data. The role separation tested here, the AI maps and the user validates, went on to guide other squads assessing where to apply AI in the product.

What stayed afterward

The trust architecture developed here (upload → automatic mapping → user review → explicit confirmation) is reusable for any AI flow operating on critical business data. The split between what the AI decides (technical mapping) and what the user decides (business validation) is a design principle, not a one-off solution. The future vision documented in the project (MLP with connection to sales history, active contracts and centralized integration with the chat system) was recorded and can feed subsequent iterations as the product’s AI vision matures.

What I’d do differently

I’d set quantitative success criteria before launch with more rigor: specifically, what would count as sufficient signal to scale vs. pivot. The qualitative criteria (Desire, Usability, Perceived Value) were clear; the numeric thresholds for the continuation decision were implicit. Given the risk of a shift in the product’s AI vision (a documented, real risk, not a hypothetical one), that upfront clarity would have protected the team from ambiguous interpretations of the Alpha results.

On the 16% completion rate: market benchmarks show that import flows in B2B SaaS typically register between 10–25% session completion. The 16% figure sits within that range. I’d make that context explicit from the start, as part of the evaluation framework agreed with stakeholders before launch, not as a defense after the fact.