Watch the video with chapters & transcript →
AI TOOLS
In short: Google AI Studio's Build tab can turn a single system prompt into a working conversational voice app, at no cost.
Voice agents are usually pitched as expensive enterprise projects. Ani Björkström, a tech consultant based in Stockholm, demonstrated a different route: building a functional website voice agent for free using Google AI Studio, with a real Stockholm spa website as the test case.
Google Gemini is the consumer version, built for quick chat. Google AI Studio is the right choice when you are creating applications. Its Playground offers Gemini for text, plus tools for images, video, and audio. Two models were available here: Gemini 3 Flash Preview, which is quicker, and Gemini 3 Pro Preview, the most intelligent option and the one used for this build.
The build is a two-stage process: first generate a system prompt, then feed it into a voice app template. The whole flow took minutes.
| Step | Where | Action |
|---|---|---|
| 1 | Playground | Ask Gemini 3 Pro Preview to write a voice-agent system prompt for the target website's URL |
| 2 | Playground | Copy the generated system prompt as text |
| 3 | Build tab | Paste the prompt, select Gemini 3 Pro Preview, choose "create conversational voice apps" |
| 4 | Build tab | Click Build, allow microphone access, and start the conversation |
The Playground also exposes a temperature setting, kept at 1 for this task: raise it for more creative output, lower it for more fact-based answers. The Build tab offers other options too, including image generation, video animation with Veo, and Google Search grounding, plus buttons to view the code, resize for mobile or tablet, download the app, save it to GitHub, or deploy it on Google Cloud.
Impressively, with one gap. The agent opened in Swedish, stopped talking the moment it was interrupted, and switched to English on request. It correctly reported Sunday opening hours of 9:00 a.m. to 6:00 p.m. and listed the saunas, including mixed and female-only saunas and a steam room. But it could not give exact treatment details or an entry price. One corrective prompt, asking Gemini to add detailed treatment and pricing information, fixed this: on retest the agent quoted 595 SEK per person for Sunday entry and 1,190 SEK for two people.
The finished app has a URL you can simply copy and share. The suggested next step is direct outreach: contact local businesses, demo the voice agent you built for their website, and ask whether they want a real agent in their production environment. Those conversations can turn a free prototype into paid consultancy engagements.
No. Gemini writes the system prompt, and the Build tab generates the application. The code is viewable if you want it, and the app can be downloaded, saved to GitHub, or deployed to Google Cloud.
Gemini 3 Flash Preview is quicker, while Gemini 3 Pro Preview is the most intelligent. The demo used Pro Preview for both prompt generation and the app itself.
Iterate with follow-up prompts. The demo agent initially lacked pricing detail; one prompt asking for detailed treatment and price information resolved it.
Read the full video transcript, with chapters, on the video page →
Want this working inside your finance team?