I Replaced My Website's Chatbot with Voice AI — Here's What Happened

#webdev #ai #saas #javascript

Last month I ripped out the chatbot widget from our website. My co-founder thought I'd lost it.

"We spent three months building that chatbot," he said. "Users need it."

"No," I told him. "Users ignore it."

He wasn't wrong about the effort. But I had the analytics to prove my point. And what we replaced it with changed everything.

The Problem Nobody Talks About

Here's the dirty secret about chatbots: they get 2-5% engagement. That's it.

We had a beautifully designed chat widget sitting in the bottom-right corner of our SaaS landing page. It could answer FAQs, route support tickets, even book demos. On paper, it was perfect.

In practice? 97% of visitors never clicked it.

The ones who did would type one question, get a mediocre canned response, and bounce. Support tickets kept piling up. Our "AI-powered" chatbot was basically a glorified FAQ page with a chat bubble skin.

I started digging into why people weren't engaging:

Typing is friction. On mobile especially, nobody wants to type out their question.
Canned responses feel robotic. Users can smell a decision tree from a mile away.
No real actions. The chatbot could tell you where to find pricing, but couldn't take you there.

What We Built Instead

We replaced the chatbot with a voice AI widget. Instead of typing, visitors just talk.

But here's where it gets interesting — this isn't just speech-to-text piped into a chatbot. The voice AI can actually do things on the page:

Navigate to any section or page ("Show me pricing" → it scrolls/routes there)
Fill out forms on behalf of the user ("Book a demo for Thursday" → fills the form)
Take DOM actions — click buttons, open modals, toggle features
Answer questions with full context awareness of what's on screen

It's like giving every visitor a personal assistant that lives inside your website.

The Tech Stack

For the devs reading this, here's what's under the hood:

React for the widget UI and DOM interaction layer
Supabase for real-time session management and analytics
Vertex AI (Gemini 2.5 Flash) for the LLM backbone — fast enough for conversational latency
Custom DOM Action Layer — this is the secret sauce. We built a middleware that maps LLM function calls to actual DOM manipulations. The AI doesn't just suggest actions, it executes them.

The function calling pipeline works like this:

User speaks → speech-to-text
LLM processes intent + page context
LLM returns structured function calls (navigate, fillForm, clickElement, etc.)
DOM action layer executes in real-time
Response is spoken back via text-to-speech

Total round-trip latency: under 500ms.

The Results

We've run this in production for about a month now. The numbers speak for themselves:

Metric	Chatbot (Before)	Voice AI (After)
Engagement Rate	3.1%	41.2%
Sessions	~200/month	500+ in first month
Tool Success Rate	N/A	99.2%
Conversion Lift	Baseline	+47%
Avg. Response Time	1.2s	<500ms

The engagement jump alone was staggering. People are curious about voice. They click the mic icon just to try it. And once they start talking, they stay.

The Multilingual Game-Changer

Here's something I didn't expect to matter as much as it did: multilingual support.

Our voice AI supports 50+ languages out of the box. That includes Hindi, Tamil, Bengali, Telugu, Marathi — basically every major Indian language.

Why does this matter? India has 500+ million internet users who don't speak English as their primary language. For SaaS products targeting the Indian market (or any multilingual market), a text chatbot in English is essentially useless for half your potential users.

Voice changes that equation completely. People speak naturally in their language, the AI understands and responds in kind, and it can still execute all the same DOM actions.

We saw our Indian user engagement go from near-zero to our fastest-growing segment.

What I Learned

Voice removes friction that you didn't know existed. Typing a question requires effort. Speaking is effortless.
Action > Information. A chatbot that tells you "Click the Pricing tab" is useless compared to an AI that just takes you there.
Multilingual isn't a nice-to-have anymore. If you're building for global users, voice is the fastest path to true accessibility.
Speed matters more than accuracy. A 95% accurate response in 400ms beats a 99% accurate response in 3 seconds. People forgive small mistakes but won't forgive lag.