Designing conversational AI across the entire B2B SaaS customer lifecycle.

Blip is a LATAM-leading conversational-AI platform (B2B SaaS) running automated and human conversations across WhatsApp, Instagram, Facebook Messenger, Google Business Messages, and web. As an internal product designer I owned conversational product design for three different jobs in the customer lifecycle, with onboarding and support handled in parallel. Not one feature in one corner of one product, but acquisition, activation, and retention, each with its own users, metrics, and cross-functional partners.

On paper they had different owners and different metrics. In practice they were one story: the same funnel underperformed at every stage, and the same asset could fix it. The top depended on paid media and static forms that dropped users; the middle was slow to get new clients to their first win; the bottom was slow, expensive, and human-bound on repetitive work.

So the decision that runs through everything: don't ship three point solutions, design one coherent conversational layer (omnichannel, CRM-synced) and adapt it to each stage's job. That reframing is what let a single working method pay off three times, and it's why onboarding and support could later collapse into one in-product surface.

Marketing paid for the clicks; the static form lost most of them, and the few leads that made it through reached Sales slow and cold.

Before designing, I benchmarked how competitors used messaging to acquire and qualify leads. The gaps, especially that no one used their own product across three messaging channels, became the opportunity.

From first touch to closed revenue: how far the acquisition agent reached, how deeply people engaged, how many became sales-ready leads, and the revenue and efficiency that followed, with each metric spelled out beneath its number.

The conversational agent became a key acquisition asset: R$15M+ attributed revenue, 48.8% YoY PaaS growth, 23% lower CAC, leads closing 3.62× more than inbound.

New B2B clients took too long to reach their first real outcome, and slow time-to-first-value is an early churn signal. Not a big flow; a focused, high-leverage intervention: get each new client to their first win fast.

Read it as evidence of orchestration across systems (HubSpot × WhatsApp), not for cell-by-cell detail, a program of rules, not a screen. Every touch is sequenced to pull a new client toward their first live chatbot before the drop-off that precedes churn, and to hand off cleanly to the support layer when they stall, the seam that later let onboarding and support collapse into one copilot.

Support is where the work crossed from rule-based conversational design into generative AI. I designed a GPT-3.5 layer to resolve common technical queries and create tickets, plus an AI agent-assist that summarizes incoming tickets so humans resolve faster. This pillar proves the generative half of the arc: grounded, bounded, and shipped, not a demo.

Customers faced roughly 68-minute first responses, no real-time human contact, and expert agents drained by repetitive, low-complexity queries: high cost, low CSAT. Many users didn't even know they could open tickets via WhatsApp.

An 18-day discovery (5 semi-open interviews, 5 usability tests, 2 journey maps, affinity mapping) surfaced a pre-journey shaping channel choice, that response time was the dominant and deficient factor, that email authentication friction hurt the chatbot path, and that many users didn't know WhatsApp ticketing existed.

First response fell from 68 to 11 minutes; the GPT layer saved R$90,200 on resolved tickets; CSAT rose 24%, average resolution 25% faster, recurrence +10%. The dynamic FAQ + AI layer drove a large jump in self-service resolution.

Onboarding help and support started as separate surfaces. The systems decision was tounify them inside the platform as a copilot-style widget: a "have a question?" entry point where the user talks to the AI and, when it's genuinely complex, opens a ticket and gets a human. That's the payoff of the reframe: two lifecycle stages, one surface, one conversational layer. The clearest evidence of the altitude, I wasn't designing three chatbots, I was shaping how help works across the product.

Seen end to end, this is the convergence node on the lifecycle spineabove, made real.

What unifies the three pillars is one conversational layer: omnichannel, delivered through native channel components and the in-product widget, and wired to the CRM. On top of that shared layer, the intelligence evolved across the funnel: acquisition ran on rule-based conversational design (intent matching, keyword disambiguation, NLU routing, business rules); support is where generative AI (GPT-3.5) entered, grounded on curated, auto-synced knowledge, with explicit rules for what it answers, when it hands to a human, and how its tone stays on-brand.

The reusable asset is the conversational + CRM layer; the generative model is the newest capability layered onto it, not the whole story. That progression, designing conversational systems before LLMs, then bringing generative AI in where it earned its place, is the range this case is really about.