We built an AI chat and voice bot to handle first contact with new clients

We built and implemented a voicebot and chatbot on our website. The goal was twofold: to test conversational AI in a real setting and to improve how we collect and evaluate project ideas from potential clients. Although this was an internal project, the end-users were website visitors looking for a tech partner.

Project Goal

We wanted to create a tool that could automatically collect key details about a client’s project idea. This helped us:

1. Quickly assess whether we could offer value.

2. Provide rough estimates faster, which benefited users.

3. Filter out inquiries far outside our scope.

The project started as a voicebot only, but we added chat functionality based on platform capabilities. 

The bot was added as an optional channel, alongside our existing contact form and email. The goal was to make it easier for visitors to engage and for us to qualify leads faster.

Challenges and Solutions

We faced several key challenges.

Prompt Instability

The prompt tool in the AI platform was unstable. It sometimes erased or cut prompts we were working on. To avoid losing progress, we backed up the last working version before each editing session. This let us recover quickly if anything went wrong.

Bias in LLMs

We found that the base models had built-in biases. For example, the bot would answer a geopolitical question about one region but refuse a similar one about another. This showed us that LLMs still have flaws. We chose to disable responses to sensitive topics and instead asked clients to book meetings for such cases.

User Reluctance to Share

Some users didn’t want to give project details to a bot. Still, our goal of booking meetings was met, as these users contacted us directly instead.

Results and Key Learnings

The chatbot and voicebot gave us useful insights into both user behavior and the tech itself:

• Text Chat Dominance: Most users preferred chat over voice.
• Client Value: Users got quick estimates, helping them decide faster.
• Business Benefit: The tool helped filter out clients whose needs were far from what we could deliver.
• Tech Limits: The project showed that LLMs have built-in biases, which need careful handling.

Future Plans

We’re considering further development. One idea is an AI video avatar that could hold audio conversations without needing users to turn on cameras. This might speed up the estimation process even more.

Try the bot yourself – it’s in the bottom right corner!