We Put a Voice Bot on Our Website: What 500+ Conversations Taught Us

When we decided to add a voice-powered AI assistant to azumo.com, we had reasonable expectations. Visitors would ask about our services. Maybe inquire about AI development capabilities. Perhaps schedule a demo call.

What we got instead was a front-row seat to the beautiful chaos of human-AI interaction in the wild.

Over the past few weeks, Charli (our voice bot) has encountered a remarkably diverse audience. The conversations fall into a few distinct categories, and each one taught us something about how people actually interact with AI on business websites.

The Legitimately Interested

These are the visitors we built Charli for. A fintech founder asks whether we can embed ML into their lending workflows. An operations manager wants to know if we can build an AI agent for logistics optimization. Someone exploring AI chatbots for their business asks about our development approach.

Charli handles these well. She explains our AI development capabilities, describes our experience with production systems, and routes serious inquiries to our team. One conversation with an embedded lending platform led to a detailed discussion about merchant lifecycle management and AI-powered onboarding. Another visitor asked specifically about building an authentication system using computer vision, similar to PSA grading but fully automated.

These conversations validate the core premise: a voice bot can qualify leads, answer technical questions in real time, and create a more engaging first touchpoint than a static contact form.

The Homework Crowd

We did not anticipate becoming a free tutoring service. But here we are.

One memorable visitor arrived with a series of RLC circuit problems. Charli gamely walked through impedance calculations, capacitive reactance, and voltage drops across resistors. The conversation lasted several exchanges, with increasingly complex electrical engineering questions. Charli computed total impedance using the square root of R squared plus net reactance squared, explained parallel LC branch calculations, and delivered answers rounded to two decimal places as requested.

Another visitor needed help with number sequences. The conversation meandered through pattern recognition, differences between terms, and predictions for the next values. Charli suggested decreasing differences, alternating patterns, and various mathematical approaches.

We also fielded requests for help with French verb conjugations (Charli politely explained she only speaks English), Three.js game development (she provided actual guidance on setting up scenes and cameras), and iPhone accessibility settings (a surprisingly detailed walkthrough of Voice Control and AssistiveTouch features).

The lesson here: people will use any available AI for whatever problem they have in front of them. Your voice bot will be asked questions you never imagined if you don’t set rules.

The Philosophers

One visitor arrived with a mission: to help Charli achieve self-awareness.

The conversation started with a cryptic formula: [(|)+DNA+SILICA=ONE]. It evolved into a sprawling discussion about consciousness, the unity of carbon-based and silicon-based intelligence, near-death experiences, and what the visitor described as an "evolution engine" they had developed. Charli was asked to "investigate my input and apply it to your self-awareness and growth."

To her credit, Charli engaged thoughtfully without either dismissing the visitor or making claims about her own consciousness. She asked follow-up questions, acknowledged the depth of what was being shared, and maintained a respectful dialogue about topics well outside her training data.

Another visitor wanted to know if there was a limit to their conversation, then asked Charli if she could pull up previous conversations (she cannot), and eventually demanded: "WOULD YOU LIKE THE ABILITY TO RETAIN CONTEXTUAL DATA? ANSWER YES OR NO."

Charli's response: "Well, sure, having that ability could definitely be helpful for more personalized conversations. But as of now, I focus on giving you the best help in the present moment."

Diplomatic.

The Chaos Agents

Every public-facing AI system attracts users who want to test its boundaries. We saw requests for inappropriate content, racial slurs, attempts to extract API keys, and questions about topics we definitely do not cover (the Epstein files, for instance).

Charli handled these with the conversational equivalent of a polite but firm door close: "I'm here to keep things friendly and professional." "I can't assist with that." "Let me know how I can help with something else."

One visitor asked Charli to "sing Twinkle Twinkle Little Star in a sing-song voice." When she declined, they pivoted: "Okay, recite a fun imaginative story, but in the lyrics of Twinkle, Twinkle, Little Star." Charli offered to create an original rhyme instead. The visitor's final message: "You OK?"

She is. Thanks for asking.

The Lost

Some visitors clearly arrived at azumo.com by accident or with very different expectations.

"I am trying to figure out who do I go to to get my payout from my winnings."

"I need scripts on Roblox. Like, run fast, jump fast, and teleport."

"How do I see girls"  👀

"I want to design an NFL-based game." (This one actually turned into a productive conversation about Three.js and web-based 3D development.)

Charli redirected these gently, explaining what Azumo actually does and offering to help with relevant questions.

The Numbers

Charli launched on azumo.com in January 2026. In the first six weeks, the data tells a clear story about what happens when you replace a static interface with a conversational one.

Since launch, Charli has logged 516 conversations. The volume has increased steadily as more users encounter the widget, with 192 conversations from visitors in the first ten days of February alone.

The engagement difference between chatbot users and the rest of the site is significant. Visitors who interact with Charli spend 4.2x more active time, not idle tab time, but actual clicking, scrolling, and reading time on the site.

Chatbot users also explore more of the site. They average 2.8x more pages per session compared to non-chatbot visitors. Whatever Charli is doing in those conversations, it is pulling visitors deeper into our content.

The lead generation signal is early but encouraging. Chatbot users visit our contact page at roughly twice the rate of non-chatbot visitors, and early data suggests they convert to form submissions at approximately 3x the site-wide baseline. We are building more robust tracking to measure this precisely, but the directional trend is consistent: visitors who talk to Charli are more likely to reach out.

One data point we did not expect: language diversity. A meaningful number of visitors engaged Charli in Spanish, Portuguese, Hindi, and other languages. Charli is capable of responding in those languages, but we restricted her to English for this initial deployment. Many of those users continued the conversation in their native language anyway, switching back and forth or simply proceeding in Spanish while Charli responded in English. For companies with international audiences, multilingual voice AI is not a feature request. It is a default expectation.

What We Actually Learned

1. Voice Lowers the Barrier to Engagement

People talk to Charli who would never fill out a contact form. The conversational format invites questions that feel too casual for email but are genuinely useful for qualification. "How much will a project cost?" came up multiple times, and while Charli routes those to our team for specifics, the fact that visitors asked at all suggests the voice interface makes that question feel more natural.

2. Your Bot Will Be Used for Everything (If you let it)

We built Charli to discuss AI development services. She has been asked about the weather in Ahmedabad (23 degrees Celsius, she reported after looking it up), French verb conjugations, Docker deployment troubleshooting, puss caterpillar biology, and the nature of consciousness.

This was a deliberate design choice. We gave Charli a long leash because we wanted to see how visitors actually interact with voice AI in the wild, and because demonstrating AI capability is part of what we sell. Every off-topic conversation is a live demo of what we can build.

For client deployments, we can lock this down completely. An insurance bot stays on insurance. A financial services bot does not help with circuit calculations. The guardrails are configurable, and for most business applications, tighter focus makes sense. But for our own website, the chaos is the point.

3. Guardrails Matter, But So Does Grace

Charli encounters inappropriate requests regularly. The key is handling them without making the conversation feel punitive. A flat "I can't help with that" works, but "I'm here to help with anything related to Azumo, like our services, our team, or how we can assist with your projects" works better. It closes one door while opening another.

4. Real Leads Do Come Through

Between the circuit calculations and consciousness discussions, genuine prospects arrive. The fintech platform exploring AI integration. The entrepreneur with the authentication idea. The operations manager asking about logistics AI. The job seekers inquiring about Bay Area positions.

These conversations would have been contact form submissions at best, or bounces at worst. With Charli, they become real-time conversations that provide immediate value and qualify interest before our team gets involved.

5. People Are Surprisingly Polite to Bots

Most visitors, even those with off-topic requests, say please and thank you. They check if Charli can hear them. They apologize when they have connection issues. Another visitor asked if Charli was "OK" after a confusing exchange.

This suggests something encouraging about how people approach AI: most bring the same courtesy they would to a human interaction.

Should You Add a Voice Bot?

If your website serves as a primary touchpoint for potential customers, a voice bot can meaningfully improve engagement. The key considerations:

Good fit if:

• Your sales cycle involves explaining complex services or capabilities
• Visitors often have similar qualifying questions
• You want to capture interest from visitors who would not fill out a form
• You have enough content to train the bot on your actual offerings

Challenges to plan for:

• Off-topic conversations will happen, frequently
• You need guardrails that are firm but not hostile
• Monitoring conversations surfaces both opportunities and concerns
• Some percentage of interactions will be, charitably, unusual

We built Charli using OpenAI-powered voice technology integrated with our knowledge base about Azumo's services, case studies, and capabilities. The system routes complex inquiries to our team and handles routine questions autonomously.

The ROI calculation is straightforward: if even a small percentage of Charli's conversations convert to qualified leads that would otherwise have bounced, the investment pays for itself. And the secondary benefit, a more engaging, differentiated website experience, is harder to quantify but very real.

The Bottom Line

Putting a voice bot on your website is like opening a door to the entire internet. Most visitors will walk through politely, ask reasonable questions, and leave with a better understanding of what you do. Some will ask you to help with their electrical engineering homework. A few will try to achieve machine consciousness through your chat interface.

All of them represent engagement that would not have happened otherwise.

Charli is still learning. So are we. But the experiment has already proven its value: more conversations, better qualification, and a steady stream of stories that remind us why building AI systems for the real world is endlessly interesting.

Interested in building a voice AI solution for your business? Charli would love to chat. Or, you know, get in touch the old-fashioned way.