Not an NLP-powered Conversational Design AI Engagement Framework
Product / Conversational Designer | 12 weeks
Just a fix for “we are experiencing higher than normal call volumes” rage
Customer service was clogged with repeat questions: “Do you stock?” “How do I get a refund?” Agents were stuck on loop. Customers were stuck on hold. And a major Aussie supermarket wanted a fix.
So we built a chatbot. Not just any bot. One with actual empathy, logic, and a working memory longer than 10 seconds.
My Role
Pushed through dead ends to finally replace hold music with real help
I led the design of the supermarket’s first conversational interface, from digging through customer data to designing a bot that actually gave people answers.
Designed end-to-end flows from discovery through to scripted interactions and UI
Dived into content audits, data + intent analysis
Built conversation logic in Dialogflow, mapping real pain points to real responses
Led chatbot UI/UX design and prototyping with suggestion chips + fallback states
Approach
From scattered questions to solid flows
The job started with friction. Customers were annoyed, agents were overwhelmed, and the business wanted a new approach.
The project kicked off with Product Vision and North Star workshops to align everyone on outcomes, not features. Then I got stuck on the data. Medallia, site analytics, SEO tools, anything that helped us understand how people asked for help. This wasn’t about flashy tech. It was about linking real needs to what the system could do.
We ran ideation sessions, built situational personas, and used value-effort mapping to get ruthless about the MVP. Refunds? In. Store hours? Yep. Dinner inspo? Not today. I mapped out every conversation flow, wrote scripts with our copywriter, and prototyped in Dialogflow. Fast feedback loops and live user reactions kept things on track and stopped the bot from becoming another digital disaster.
Challenges
Everything looked fine until someone asked a real question
No one speaks the same language
Customers said “Where’s my stuff?” The system expected “track my order.” Bridging the gap meant testing real conversations, not assumptions.
You can’t test what you can’t see
Conversations aren’t wireframes. We built custom tools to track flows, spot dead ends, and fix issues fast.
Let’s launch everything
Everyone wanted the bot to do everything on day one. Value vs effort mapping kept us honest (and sane).
Bots don’t earn trust easily
We designed tone, timing, and fallback logic carefully. Otherwise, it risked sounding like a confused call centre script.
Solution
Made self-serve feel like the obvious choice
Flowchart-first design
Every intent was mapped visually, so teams could spot dead ends before users did.
Scripted like real people talk
Conversations were co-written with a UX copywriter using actual customer language from Medallia and SEO data.
Dialogflow prototyping
Fast iterations with real users revealed where flows fell apart.
UI designed for fallback
Suggestion chips guided confused users. Branded UI kept trust intact.
Results
Fewer tickets. Shorter waits. Better vibes.
Thousands of chats per week
NPS up 2 points in week one
Agent calls slashed
Still learning post-launch
The bot handles order status, product info, and store info without needing a human.
Customers didn’t just use it, they liked it.
Reduced repetitive tickets so humans could handle actual edge cases.
Feedback prompts at the end of chats fuel continuous updates.