Case study
Chatbot UX
Implementing a real-time hybrid chatbot on the Choices platform to improve user engagement, answer common questions in the moment, and reduce drop-off caused by unanswered queries.
- Role
- UX/UI Design · UX Research · Design Strategy · Accessibility Consulting
- Timeline
- 3 months
- Tools
- Figma · Miro · UserZoom
- Outcome
- Response time 12h → minutes
- AI
- Conversational UX
- FinTech
Overview
The Choices platform had a support problem. People had questions mid-flow and nobody was there to answer them. The only option was email with a 12-hour wait. Most left instead of waiting. I designed a hybrid chatbot to handle the most common questions in the moment and hand off cleanly to a live agent when it couldn't.
Problem & UX Research
Support requests kept coming in around the same handful of questions. The platform had no way to handle them in the moment.
What Research Showed
- The same questions kept coming up in every session: products, fees, transfer methods, retirement policies. Five topics covering the majority of support volume.
- Users hovered around the FAQs but left without acting. Passive information delivery wasn't enough. They needed someone to talk to, not more text to read.
Strategy & Discovery
The core decision was which chatbot model to use. Three were on the table: pure rule-based, pure NLP, and hybrid. Pure rule-based was too rigid for any variation in how users phrased questions. Pure NLP hallucinated on financial details we couldn't risk. The hybrid model handled the common intents accurately and used NLP to manage variation without falling apart.
Design Process
Chat Widget Placement
The widget was positioned in the bottom-right corner for consistent, low-friction access. This follows established conversational interface conventions, reducing the cognitive effort of finding support without competing with primary page actions.
Quick Action Buttons
Predefined quick-action buttons were added for the most common intents, allowing users to initiate queries without typing. This reduced the barrier to first interaction and surfaced the most valuable chatbot capabilities immediately on open.
Handoff to Live Support
The handoff to a live agent was designed as part of the experience, not an afterthought. Agents got the full conversation context so users didn't have to start over. Users saw a clear message telling them what was happening, so it felt intentional rather than like the bot giving up.
Feedback Loop
After each response, the chatbot asked a probing satisfaction question. This real-time feedback mechanism enabled continuous improvement of responses over time and helped identify intent coverage gaps early after launch.
Results
After launching the chatbot, performance was tracked over 60 days, demonstrating improvements across all key indicators.
- ↓
- Bounce rate, unanswered queries no longer the primary abandonment driver
- ↑
- User engagement and conversion rate post-launch
- 12h→min
- Average response time reduced from 12 hours to minutes
Learnings
- Placement and discoverability matter as much as response quality. A well-written bot that's hard to find doesn't help anyone.
- The highest-leverage decisions weren't about intent coverage or NLP accuracy. They were about what happens when the bot reaches its limit and whether users feel helped or handed off.
Conclusion
Conversational design is a trust problem before it's a flow problem. Treating the handoff as a first-class experience and building a feedback loop from day one produced a system that improved after launch and held user trust even when it couldn't resolve things on its own.
Continue exploring
More work that pairs rigor with craft
Open another case study or return to selected work on the homepage.