Product Design Case Study
AI Conversational Chat with Funnel Capabilities
From fragmented chat experiments to a scalable AI product system
The problem all started with the wrong icon.
The problem all started with the wrong icon.
During one of my first weeks as AI Product Designer at Natural Intelligence, I was reviewing existing chat implementations across our products. The team was running an A/B test on avatar icons — trying to find the right visual to represent the AI assistant. One of the variants was a figure with a round head and an antenna on top.
A robot.
We were trying to build an AI product that people would trust, engage with, and ultimately make decisions through — and we were testing a robot mascot to welcome them in.
That image stopped me. Not because of the icon itself, but because of what it revealed: the chat had been built like a support bot, dressed up with a new label. If it looks like a bot, sounds like a bot, and responds like a bot — no amount of “AI” in the product name changes how people experience it. And the data confirmed it. Nobody was using it.
That was the moment the problem became clear. Within the first week of shipping the redesigned experience, roughly 3% of users converted directly through chat — a surface that had previously delivered near-zero conversion. This is the story of how we got there.
Where I Was Coming From
Before this role, I was a product designer at Natural Intelligence, with an additional responsibility in DesignOps — building design systems, improving how designers and developers collaborate, creating the infrastructure that lets product teams move faster and more consistently. I was shipping product as an IC while also thinking about how design scales across the org.
Around early 2025 I started seeing signals I couldn’t ignore. AI wasn’t just changing content and marketing — it was changing what a designer could actually build. No-code tools, AI-assisted prototyping, products you could ship without an engineering team. I took a course, built a working product from scratch, and had what I can only describe as a “there is no spoon” moment. The limits I’d assumed were fixed turned out to be mostly self-imposed. Imagination and time became the real constraints — not tooling, not access, not seniority.
I leaned in hard. My manager and head of design noticed, and when the AI product designer role needed someone to lead it, the transition made sense to both of us.
The chat funnel project was one of the first things I touched in that role. It was, in hindsight, exactly the right problem for someone who thinks in systems and had just started believing that everything is buildable.
Understanding The Biz And More
Natural Intelligence runs comparison marketplaces — Top10.com and BestMoney.com — where high-intent consumers come to make decisions about financial products, insurance, home services, and more. The business model is built on matching the right person to the right brand at the right moment. Conversion is everything.
The company operates across many verticals simultaneously, which means product decisions rarely live in one place. Design, product, engineering, data, and business stakeholders all have a seat at the table — and they don’t always want the same thing. On this project, I worked closely with PMs and engineers, and alongside the designers embedded in each vertical who were responsible for implementing the chat system in their specific areas. They were my ground-level feedback loop — the people who knew what edge cases existed, what constraints the system had to accommodate, and what wouldn’t work in practice.
Chat was supposed to be a key moment in that journey — a way to guide users, understand their intent, and connect them with the right partner. Instead it was underperforming across every metric that mattered. Users rarely started a conversation. When they did, sessions ended quickly. There was almost no repeat usage. The data ruled out placement — chat had prominent positioning on most surfaces. The problem was the experience itself.
And here’s where it gets complicated: Natural Intelligence is a performance-led company. Revenue is optimized, monitored, and fought for across every surface. Chat wasn’t just a UX problem — it had a business expectation attached to it. It was supposed to convert.
But that pressure was part of what had broken it. Chat had been pushed toward conversion before it had earned the user’s trust. The result was an experience that felt transactional and hollow — more like a sales script than an intelligent assistant. Users felt it, even if they couldn’t name it.
The tension I had to navigate wasn’t just design versus engineering, or user needs versus business goals. It was specifically: how do you build something that converts, for a company that converts everything, without destroying the trust that makes AI useful in the first place?
What the Work Actually Showed
I came into this without having designed any of the existing implementations. That was useful — it meant I could look at them the way a user would, without the attachment that comes from having built something.
I started with the data. Low initiation rates, short sessions, high drop-off after the first message. These patterns told me the problem wasn’t awareness — users were seeing the chat. They just weren’t finding a reason to continue.
The qualitative picture filled in the why. Chat responses were long and unstructured, designed for reading rather than scanning. There were no suggested actions, no clear next steps, no sense that the system understood what the user was trying to do. It felt generic in a way that’s hard to pinpoint but impossible to ignore.
I also spent time researching how trust works in AI experiences — what was emerging in the field, what competitors were doing, what patterns were starting to show results. AI UX in 2024 was still being figured out in public. Best practices were forming in real time. That research shaped a core belief I kept coming back to throughout the project:
Structure builds trust more than personality.
Research finding — 2024
A well-organized, clear, scannable response signals intelligence. A friendly tone with an avatar does not. The more we leaned on visual personality to compensate for a weak interaction model, the worse the experience felt.
The robot icon wasn’t the problem. It was a symptom. The real issue was that the whole system had been designed around the wrong mental model.
Building Blocks, Not a One-Off
The brief I gave myself was simple: don’t design a chat. Design a system that powers chat across everything.
Natural Intelligence runs multiple products across multiple verticals. Any solution that lived in one place would get rebuilt, diverged from, and eventually abandoned somewhere else. I’d seen that pattern up close from my time in DesignOps. The only way to create real change was to build something reusable — a set of components and rules that any team could adopt, adapt, and extend without starting from scratch.
The system had three layers:
This is where my DesignOps background became directly useful. I wasn’t just designing screens — I was designing a design system for a conversational product. The building blocks had to be flexible enough to serve different verticals with different needs, but consistent enough that a new team could pick them up without a briefing.
I also worked closely with the designers embedding chat into their specific areas. They were the ones who surfaced edge cases, constraints, and things the system needed to handle that I wouldn’t have seen from the center. That feedback loop was essential — it meant the system got stress-tested before anything was built.
The Hard Calls
Not everything went smoothly. Two specific challenges shaped how the project evolved — and both taught me more than the parts that went well.
The foundation problem
After I completed the initial design system — the full thing, built around the interaction model and principles I’d defined from scratch — I found out that a different design had already been partially developed by engineering. It wasn’t scalable in the direction we needed, but it existed, it had investment behind it, and we were going to build off it.
That was a frustrating moment. I’d done the work, and now I had to throw out parts of it and adapt to a foundation I hadn’t designed. But I didn’t take it too hard. Practically speaking, doing something the second time is always faster — you’ve already solved the hard thinking, and now you’re just applying it with new constraints. I adapted the system to fit the existing foundation and focused on making it scalable from there. The principles stayed intact. The execution shifted.
It’s a common tension in product work: ideal design versus what’s already in motion. The skill is knowing what to hold onto and what to let go.
The trust vs. conversion tension
Natural Intelligence is a performance-led company, and that culture extends into design conversations. People are opinionated — about placement, hierarchy, copy, interaction patterns — and AI chat, being new territory for both the company and the industry, brought out especially strong opinions. Everyone had a view on how aggressive we should be, how prominent the conversion moments should be, where the sidebar should sit, how the flow should sequence.
I genuinely appreciate that kind of input. Design shouldn’t happen in isolation. But AI experiences have a different set of rules. Users come to comparison sites already high-intent — they want help making a decision, not a sales pitch. An assistant that feels pushy doesn’t just underperform, it actively destroys the trust that makes it useful. That’s not a UX opinion, it’s a behavioral one.
When I pushed back, I didn’t do it from instinct alone. I did my homework — found examples from comparable AI products, referenced UX laws and established heuristics, and built the case from evidence. Not every stakeholder agreed immediately. But grounding the conversation in research rather than preference moved it forward faster than any debate about taste would have.
The goal was always a middle ground — an experience that the business could get behind and that users would actually trust. That balance took longer to land than the design itself.
What Shipped — and What's Coming
The foundation is live. The conversation engine, the interaction model, the shell system — these are running across products today.
Within the first week of the new AI experience, we saw roughly 3% of users convert directly through chat. For a surface that had previously delivered near-zero conversion, that was a meaningful signal. Not a final number — we’re still measuring — but a clear indication that the direction was right.
Users are also telling us, through behavior and feedback, that doing everything in one place is better than being bounced between surfaces and sites. The fragmented experience that existed before wasn’t just a design problem — it was friction that cost real decisions.
What’s next is the part I’m most excited about. We’ve designed and prototyped the agentic layer — the ability to complete actions inside the chat itself. Lead capture. Scheduling. Payments. Intent routing to the right partner without leaving the conversation. Partners have seen these prototypes and the response has been strong. The vision of chat as a complete decision and action environment, not just a guidance tool, is already showing traction before it’s fully shipped.
Connecting the Dots
Two things stayed with me from this project.
The first is about AI product design specifically: the patterns we inherited from support chat are the wrong starting point. AI creates different expectations — users expect structure, guidance, and intelligence, not a faster version of a help widget. When you design against those expectations, trust breaks immediately. When you design with them, something clicks. The difference isn’t always visible in the UI. It’s in how the interaction feels from the inside.
The second is more personal. The “there is no spoon” shift I had early in 2025 — the understanding that imagination and time are the real constraints — showed up directly in how I approached this work. I wasn’t asking what was possible given the tools and the org. I was asking what the experience should be, and working backward from there. That’s a different starting position. It changes what you propose, what you push for, and what you’re willing to defend.
This case study is, in part, evidence of that shift.
