Google Cloud - Dial an App
Building the world’s most efficient co-founder, a hyper-fluent AI agent available 24/7 by phone
Introduction
Inside of an urban parklet at Google Cloud Next ‘25 sat 3 phone booths with a large touchscreen and a telephone receiver, beckoning attendees to step inside and build an app from scratch in a matter of minutes. Together with the Google Demos & Experiments team, and with consulting from the stars behind the Customer Engagement Suite product, we built this experience from the ground up.
Over the course of 2.5 days, Dial an App generated over 1 million lines of code, fielded around 3,300 minutes of conversation and generated over 625 apps—ranging from sophisticated tech solutions to saree builders to dogwalking apps.
Callers only had to pick up the receiver to be greeted by Google’s most next-gen voice. So next-gen that we were swapping new voices and models into the experience until the last minute to ensure that it was as fluid, conversational and intelligent as it possibly could be. After greeting, callers were offered two pathways. Already got an idea? Great, let’s make it real in no time. Need help brainstorming? Let’s come up with something brand new in just a few minutes.
The agent was capable of deeply natural conversation, and ready to respond to any caller need. Once it had all the information required to build, a summary was passed on to Gemini, and in just a minute, app prototypes, designed mockups, a product requirements document and actual code and app architecture were output. From nothing to nearly pitch-ready in as little as 3 minutes.

Creative Technology
With a front end and screen powered by AppEngine and a backend where CES ran the show and Gemini swooped in to help generate outputs, we opted to build a Playbook in Dialogflow which ran a few key tools. We needed to minimize tool calls and webhooks to keep the speed of conversation optimal, and in the end the agent was so fast and conversational that we were actually prompting in order to slow it down.
After initial latency between the agent and caller turns, we were able to tinker over the course of a few weeks to really streamline the model and voice so that its speech patterns, tone and voice, and speed were pretty incredible. Many Next attendees were fascinated by how we built and prompted the agent. This showcase definitely moved us into another echelon of prompting expertise.


Callers picked up the receiver, and a physical trigger built into the phone booth by the George P. Johnson team in collaboration with our tech leads signalled to the agent that the call was beginning. From there, we called a tool which began to register the contextual nuances of the caller’s idea. Numerous training example scripts guided the agent on how to sound incredibly fluid while ensuring that it hit the key points needed to build an app, avoiding snags at all costs.
As the agent progressed through its goals and instructions, it checked off the narrative boxes needed and then began to build an output package for the caller. In collaboration with Gemini, which handled the visual design mockups and UI/UX prototypes, CES helped output a PRD filled with pertinent details which would be required to pitch on and build the app, plus a full document of tech specs complete with app architecture.
The package was served up to the caller in clickable form on the touchscreen, and provided for download via a QR code, in a moment of celebration.
Outside of the phone booths, a live updating Big Data Dashboard was translating insights in real time, including minutes analyzed, calls fielded, apps generated, industries of focus, and even app spotlights. We trained Gemini to name each app and create a brief synopsis for the App Spotlight section of the Big Data Dashboard, which would appear once the caller was finished.

Results
In all, Dial an App fielded over 1,300 calls, generated over 625 apps, spoke to callers for over 3,300 minutes, and wrote over 1 million lines of completely automated code for callers’ apps.
CES and our build dazzled callers, who were also able to speak to the Dial an App agent in their native tongues with perfect recognition, be responded to in English, and still have an app built in perfect timing. And all of this took an average of five minutes from start to finish.
Attendees were amazed by the conversation quality, the flow of the discussion, and the prompting that made it possible. The challenge of linking CES and Gemini to AppEngine on the front end proved out some incredible applications for the world of customer service and engagement, and we’re planning to offer Dial an App as a standalone public service accessible to users via mobile or desktop, and even possibly a toll-free number.
In addition, we’ll be moving the Dial an App booths around the country to meet callers at other locations over the next year.