Case Study - AI Product Design
How I took a vague idea and turned it into a 74-screen AI chat product — from blank slate to acquisition .
Company
My role
Team size
Timeline
🔥 Outcome
Cadena received top ratings from the client and was acquired - a direct validation of both the product direction and the design decisions made throughout this engagement.
01 — Context
Cadena is an AI innovation lab focused on building agentic systems for high-stakes environments — from retail to industrial R&D. Their leadership team includes veterans from Google, MIT, and Harvard. When they came to us, they had something far more valuable than a spec: a clear conviction about the problem they were solving.
The brief was almost entirely verbal. No wireframes, no product documentation, no user flows. Just a sharp insight about what was broken in e-commerce and a vision for how AI could fix it. My job was to turn that conviction into a designed, validated product.
The brief was almost entirely verbal — a sharp insight about what was broken in e-commerce and a vision for how AI could fix it."
02 — The Problem
Think about the last time you tried to buy something online. You searched. You filtered. You opened twelve tabs. You compared. You lost track of what you were even looking for. Then you gave up, or settled.
The modern e-commerce experience is built around catalogues and filters — tools designed for browsing, not buying. Shoppers carry the full cognitive load of finding, comparing, and deciding entirely on their own. The result is high dropout rates, abandoned carts, and buyers who leave without finding what they actually needed.
Cadena's premise was simple: what if a retailer's website could talk to you the way a knowledgeable salesperson would? Not with a script, but with genuine conversational intelligence — asking the right questions, making smart recommendations, and remembering what you told it.
User problem
Too much time, too little confidence
Shoppers spend excessive time searching and comparing — often without reaching a confident decision. Decision fatigue drives cart abandonment.
Business problem
Lost sales at every stage of the funnel
Retailers lose revenue not just at checkout, but at every moment a shopper becomes overwhelmed and disengages — long before they reach the cart.
Market gap
Existing chat tools aren't built for buying
Tools like Tidio and Drift are customer service platforms first. None of the incumbents had built a product where the AI's primary goal was to sell — intelligently and contextually.
Opportunity
AI shopping assistant as a differentiator
A white-labeled AI chat that retailers could brand as their own — one that guided shoppers toward purchase the way a great in-store sales associate would.
03 — Discovery & Research
Before touching Figma, I ran a structured competitive analysis across five direct competitors: REP AI, ADA, LivePerson, Drift, and Tidio. The goal wasn't to catalog features — it was to find the white space Cadena could credibly own.
Four key insights shaped the entire design direction that followed.
Insight 01 — Interaction
NLP is now table stakes
Every competitor had moved beyond scripted flows. Natural language understanding wasn't a differentiator anymore — it was the baseline expectation. Cadena had to meet it and go further.
Insight 02 — UX
Signup walls kill conversion
Competitors that avoided sign-up requirements saw better engagement. Forcing authentication before users could even start a conversation was a consistent drop-off point.
Insight 03 — Innovation
The best tools go beyond the chat window
REP AI could add to cart and complete checkout from within the chat itself. The most forward-thinking competitors were collapsing the gap between conversation and transaction.
Insight 04 — Brand
Humanizing AI increases loyalty
Competitors avoided labeling their assistants as "AI" in consumer-facing contexts. Giving the assistant a name and using online indicators dramatically improved user comfort and retention.
04 — Defining Scope
This was the hardest design problem on the project — and it had nothing to do with the interface. The stakeholders came with an enormous wishlist: AR visualization, video analysis, social shopping features, image search. All compelling ideas. All impossible within budget and timeline for a product that hadn't validated its core value yet.
Rather than push back reactively, I introduced a structured framework: the MoSCoW model. I mapped every requested feature — verbatim from stakeholder conversations — into four buckets: Must Have, Should Have, Could Have, and Won't Have (for now).
Rather than push back reactively, I introduced a structured framework: the MoSCoW model. I mapped every requested feature — verbatim from stakeholder conversations — into four buckets: Must Have, Should Have, Could Have, and Won't Have (for now).
This turned a tense negotiation into a collaborative planning session. Stakeholders could see their ideas were heard and documented. They could also see, clearly, why some of those ideas had to wait.
05 — Key Design Decisions
The most consequential design decision wasn't about visual style or interaction patterns. It was about placement: how and where the AI chat should exist within the e-commerce experience. Three options were on the table.
Option 01
Standard chat widget
A small bubble in the bottom-right corner. Familiar, unobtrusive, low risk. But it signals "support tool" — not "shopping assistant." No differentiation.
Option 02
Full takeover modal
A primary CTA that launches a full-screen experience. Maximum real estate, but invasive — hard to reach back to the browsing context without losing your place.
Option 03
Familiar trigger, bold canvas
The trigger lives bottom-right — familiar and unobtrusive. When activated, it expands to ~90% of the screen, giving the conversation room to breathe. One tap closes it cleanly.
06 — The Solution
The final product centered on a split-pane interface — conversation left, recommendations right — that let the AI earn the shopper's trust through dialogue before surfacing the right products. Every interaction was designed to feel less like a search tool and more like a conversation with someone who actually knew the catalogue.
Conversational search & filtering
The AI asks natural follow-up questions — budget, style, occasion, room context — and refines its recommendations mid-conversation. Designed to feel like NLP, not a form.
Interactive product cards
Three card states — Default, Selected, and Highlighted — let users interact with recommendations on the right to influence the conversation on the left. Selecting a card sends a visual cue back into the chat thread.
Soft authentication
Rather than gating access behind a login wall, the AI prompts users to sign in mid-conversation — after they're already invested — so the session can be saved and picked up later.
07 — Outcome
Within the engagement, Cadena received top client ratings through Superside — a direct reflection of the stakeholder management approach as much as the design craft. By the time delivery was complete, the product had validated its core premise well enough to attract acquisition interest from .
No design project exists in a vacuum. The acquisition isn't just a business outcome — it's evidence that the decisions made in the design process held up under commercial scrutiny. The scope we defended, the research we used to anchor decisions, the UX patterns we chose — all of it contributed to a product that someone thought was worth buying.
08 — Reflection
What I'm most proud of on Cadena isn't the interface — it's the moment the MoSCoW conversation happened. Walking into a room where stakeholders are excited, underfunded, and convinced they need everything is a situation most designers handle badly. Either they fight the client, or they capitulate and build the wrong thing.
The MoSCoW framework gave us a third path: a structured way to make the client feel heard while giving the product a fighting chance. That decision — to spend time organizing thinking before designing — is what made the rest of the project possible.
Cadena also reinforced something I believe deeply: that the most impactful design decisions are rarely visual. They're about sequencing, framing, and helping clients want what the product actually needs.