VivazGPT
A conversational platform that lets CX teams ship AI assistants without writing code.
Visit product
The context
Support teams depended on engineering for any change to the assistant. Every tone or intent tweak became a ticket — and the queue was measured in weeks.
The real problem
The core tension wasn’t the AI generating answers — it was trust: no existing assistant let the support team publish and control content without code, so no one trusted it.
My role
- Designed the end-to-end experience: dialogue flows, intent editor, and human handoff.
- Defined the tone-of-voice guidelines and ethical guardrails with the CX team.
- Ran the moderated tests and the pilot with a support squad.
The critical decisions
Intent editor before the analytics dashboard
The CX team wouldn’t trust data from an assistant they couldn’t control. I prioritized control over metrics — analytics came later.
Show “where the answer came from”
Surfacing each answer’s source reduced operator anxiety and made the human handoff a conscious choice, not a rescue.
Tone guidelines next to the editor, not in a PDF
Persona and tone had to live inside the editing flow. Moving them from docs into the UI is what made the guidelines actually followed.
The AI layer
- When model confidence was low, the assistant offered a human handoff instead of risking an answer.
- Tone-of-voice guardrails kept replies inside the brand persona.
- Knowledge curation stayed with the support team, keeping the source of truth auditable.
Outcomes
- Consistent reduction in average handling time during the CX pilot.
- A meaningful share of requests resolved without human handoff.
- Greater internal-team confidence after consolidating the curation panel.
What I’d do differently
I’d have introduced source transparency even earlier. It moved trust the most, yet only landed after the pilot had begun.