CONVERSATIONAL DESIGN FOR CRITICAL MOMENTS
Redesigning the Self-Service Experience for Weather
Overview
Weather*, a Brazilian leader in 24/7 assistance, faced a critical efficiency challenge that was directly impacting its profit margins.
With over 22,000 monthly inbound requests via WhatsApp, the existing chatbot experience was unintuitive, leading to high user dissatisfaction. This forced an excessive number of escalations to costly live agents. This situation created a clear business opportunity: reduce operational costs and significantly improve customer satisfaction.
*Fictional name to protect client confidentiality.
My role
Uncovering User Friction
We kicked off the project with a deep dive, conducting user interviews and analyzing existing platform data. The goal was to map key friction points and validate our initial hypotheses for a self-service MVP. The research quickly revealed two distinct user personas: "Novice Users," who required more guidance, and "Power Users" (such as insurance agents and automaker employees), who demanded maximum efficiency.
User difficulties
Amanda Silveira
"I can't get a human. It's horrible to be desperate for help and stuck talking to something that doesn't understand a simple problem. And another thing, the 0800 number they give you doesn't always work."
Elizabete Flores
"They should just let me get help through this same channel."
Marcos Alexandre
"My problem wasn't solved. The bot just gave me a phone number for 'capitals and metropolitan areas,' which isn't helpful to me."
The Root Cause and The Structural Solution
Digging deeper into the 72% drop-off rate, I concluded the problem wasn't the conversational UI, but was architectural. The true root cause was the backend API logic: it forced all users through a single, convoluted data-triage flow, causing mass frustration.
My solution was to redesign the API's Information Architecture itself. I proposed creating two distinct user journeys mapped directly to our personas: a guided flow for "Novice Users" and a "fast-track" path for our "Power Users." After aligning with the product team, this new logic was implemented. This single structural change slashed the human-contact escalation rate from 72% down to 36%.
Bringing the New Experience to Life
With the new API architecture in place, our focus shifted to the quality and intelligence of the conversation itself. To transform the experience from "a bot that doesn't understand" to "an assistant that resolves," we implemented two crucial enhancements:
Conversational Intelligence (AI)
We introduced a Natural Language Processing (NLP) engine to interpret user intent and correct for errors. Through a validation cascade, the chatbot could now understand the intent behind the words, even with typos. This made the conversation drastically more effective at resolving user issues on the first try.
Shifting to an Actionable Metric (CSAT)
I argued that the company's existing metric, NPS (Net Promoter Score), was too broad to measure the quality of a specific, task-based interaction.
Outcomes & Future Outlook
We concluded the project by exceeding our initial goals. More importantly, we uncovered key insights that have directly informed the product's future strategy.
Key Wins
User segmentation was a high-leverage initiative. It required low engineering effort but delivered an outsized impact, significantly improving the perceived value for our key B2B clients.
We learned that in conversational design, mapping the data architecture and business rules before designing any flow is now a non-negotiable step. This process change now saves us weeks of potential refactoring.
The success and flexibility of the solution validated the feasibility of a whitelabel product, opening the door for a promising new revenue stream for the company.
Reflections & Lessons Learned
Our initial discovery was overly focused on conversational usability. We should have conducted a deep technical diagnostic with the engineering team upfront. Realizing the API's technical constraints late in the process cost us valuable time exploring surface-level fixes when the root problem was architectural.
We started with the hypothesis that a single, "one-size-fits-all" flow, even if well-designed, would meet every user's needs. We learned that segmenting our "Novice Users" and "Power Users" wasn't just an optimization—it was a core requirement for the product's success.
We began the project using the company's default success metric (NPS). We quickly realized it was too broad to generate actionable insights into the chatbot's specific usability. The decision to switch to a per-journey CSAT was reactive. Today, I would advocate for defining these more granular metrics from day one.