Talkvisor · LIVE PROJECT
From Off Switch
to Automation
© Talkvisor
ROLE
Founding Product Designer
TIMELINE
2 Months
TEAM
Engineers, PM, Data Science
SKILLS
AI Prototyping, Interaction Design, User Research, Design Systems
PROBLEM
Merchants turned on automation, then turned it right back off
Talkvisor's AI could answer most customer questions on its own, but merchants didn't trust it enough to let it. They switched autosend on, got nervous, and switched it off again. With no way to test the AI, see what it knew, or understand why it answered the way it did, keeping everything manual felt safest, so the automation they were paying for sat unused.
The Pattern
Merchants kept toggling autosend on and off instead of leaving it running.
Root Cause
Not a weak model. A trust gap. Merchants couldn't test the AI, see its reasoning, or set limits on it.
Merchant Signal
One merchant kept messaging our support widget with store updates they wanted the bot to learn.
SOLUTION
Trust, built one step at a time
Instead of asking merchants to flip one big switch, we broke trust into steps they could take at their own pace: test the AI somewhere safe, understand every answer it gives, set the rules for what it can send on its own, and hand it more as their confidence grows.
CORE FLOWS
Test it. Understand it. Set the rules.
Each feature answers a question merchants kept asking before they would let the AI act on its own: Does it know my store? Why did it say that? What will it send without me?
AI Playground
A no-consequences sandbox to chat with the AI, flag wrong answers with a thumbs down, and teach it the right one.
Why This Response?
Every AI reply shows how the question was classified, where the answer came from, and why it was sent on its own.
AI Tuner
Merchants choose what goes out on its own, from draft only to autosend everything, and which topics always come back to them.
RESEARCH
Watching merchants hesitate
Session recordings showed the same moment again and again: merchants moving their autosend level up and down during setup, unsure how much to hand over. Together with the merchant who turned our support widget into a makeshift training channel, the signal was clear. People wanted to teach and check the AI before trusting it, so we gave them a place to do that. The first Playground had no sync controls, so merchants kept typing things like “refresh” or “rescan my store” into the chat, not realizing they could start a sync themselves. We added a What Talkvisor Knows panel that shows what the AI has synced from their store, with a Scan my store button right where they were looking for it.
BUSINESS IMPACT
Results
89% of merchants who used AI Playground stayed, a much higher retention rate than merchants who didn't. Explainable, reversible AI controls shipped in the product, alongside translated conversations so global stores could trust replies in their customers' language.
Strategic directions
Three bets on earning trust
The principles behind every automation decision.
Let Them Test First: give merchants a safe place to try the AI before it ever talks to a customer. Explain Every Answer: show the reasoning behind each reply so the AI feels like a teammate, not a black box. Trust Is a Dial, Not a Switch: let merchants hand over more, or pull back, at any time, with sensitive topics always routed to them.
Let Them Test First
Give merchants a safe place to try the AI before it ever talks to a real customer.
Explain Every Answer
Show the reasoning behind every reply so the AI feels like a teammate, not a black box.
Trust Is a Dial, Not a Switch
Let merchants hand over more, or pull back, at any time. Refunds, angry customers, and complaints always come back to them.
reflections
Key Takeaways
Trust Is Built in Steps
Merchants didn't need a smarter AI first. They needed small, safe ways to watch it work before letting go.
Workarounds Point the Way
The Playground started with one merchant using our support widget to train the bot. Watching how people bent the product showed us what to build next.
Collaboration Compounds
Working closely with engineers, PMs, and data science through multiple iterations shaped automation that feels both smart and human, something no single round of design could have gotten right alone.
U-Haul Homepage
U-Haul Truck Reservation