Jaclyn Konzelmann: When AI Can Build More, Choose What Deserves to Exist

Google Labs product director Jaclyn Konzelmann says more capable models have changed both what her team can imagine building and how it works day to day. As more ideas become feasible, the team spends time reconsidering which ones are worth creating.

Rather than starting with a long product requirements document, Konzelmann often starts with three things: a high-level strategy explaining what, why and the opportunity; a presentation for discussing the idea with others; and a prototype to gauge excitement before committing to a production build.

Labs applies a sequence of priorities: users first, Google second and Labs third. Konzelmann considers both newly emerging model capabilities that might make an old idea viable and changing user habits that could create a need for a new kind of product.

She gives Pomelli as an example. Image generation was improving at producing material consistent with a brand, while small-business owners often lacked time to learn elaborate prompts. Her parents had run small businesses, and she saw an opportunity to make the capability accessible in a different product.

A polished one-off prompt is not the same as a product that works for many businesses. The team explored understanding each business so it could produce relevant material while allowing owners to participate as much or as little as they wanted.

Konzelmann says Labs works with marketers, businesses and trusted testers. The team brings small-business owners into workshops to use what it is building and give feedback. The aim is an iterative loop that tests accessibility and usefulness, not merely technical possibility.

Listening does not mean implementing every request literally. Konzelmann says the team should understand the underlying problem and use its knowledge of the technology to find a solution the user might not have imagined.

Strategy documents still take time, even if they are shorter. Konzelmann encourages authors to sit with an idea, refine the narrative and remove material that does not deserve a reader’s attention. Once the idea is clear, she says, it is easier to bring people along and build.

Her rules for AI-assisted writing include reading the output and being able to defend every word. If a document takes longer to read than its author spent producing it, she considers that a warning sign. She uses AI as a thought partner while expecting the author’s own thinking to come through.

Yang asks why Labs runs experiments instead of focusing only on the Gemini app. Konzelmann says separate teams work on the app and future experiments, and the latter can contribute products or lessons. She points to an experiment called CC influencing a daily briefing as one example.

The outcome of each experiment depends on the case, and the team closest to the problem often helps decide whether it should grow, graduate or inform another idea. Labs gives teams autonomy because they know the product and users deeply.

Ending an experiment need not mean failure. Konzelmann describes the principle of keeping the lessons rather than the prototype: teams reflect on what they learned and share it at internal meetings. Code and infrastructure can sometimes be reused by others as well.

She credits leadership with encouraging curiosity and says teams hold themselves to a high standard. When an idea encounters unexpected challenges, she wants people to discuss them safely and work out how to change course. Internal sessions let teams share methods and ask colleagues for help.

Yang asks whether faster AI-generated code has shifted the product bottleneck. Konzelmann says it has made the value of a product a sharper question. An agent may be able to reproduce an isolated feature over a weekend, so the team must ask what makes an experience durable beyond its code.

She mentions network effects and context as possible lenses, while acknowledging that user behavior is changing. Some people may interact with an interface themselves; in other cases a personal agent may use the underlying service on their behalf.

Konzelmann still wants direct interaction when she is creating, learning or being entertained. For tedious work she may prefer an agent. Yang agrees that forms and administrative tasks are candidates, but says he sometimes feels slighted when another person delegates a reply to an agent.

Konzelmann gives an example from a family hike: a GPS glitch spoiled the route recorded in Strava, and repairing the file manually required several steps. Her agent fixed it, but she still values seeing what friends have accomplished. The lesson for her is to remove toil while preserving what people enjoy.

Both speakers like using voice to work away from a screen. Konzelmann records thoughts while walking and later shapes them into a finished document. Many of her strategy documents begin as voice memos.

As a parent of a five-year-old and three-and-a-half-year-old twins, Konzelmann wants her children to spend time outdoors and learn to handle boredom and friction. She also wants AI to help them imagine that they can make things.

Her oldest daughter loves scary stories, so they began making a podcast together. The child chooses voices and character names, helps refine generated images and storylines, then listens to drafts and requests changes. Konzelmann uses an agent to revise the result.

They publish the stories so the child’s friends can listen too. That choice grew from an earlier song they made together: when her daughter asked if it was “real,” she meant whether her friends could share the experience.

Konzelmann has also made a marble-run video and an interactive animal game for her children, and wants to customize them with the children’s input. She wonders how such tailored experiences might make learning more engaging as they grow.

Yang wants children to create rather than only consume media. Konzelmann agrees that she wants hers to be curious, learn and participate actively, not just use AI to make everything easier. Yang adds that he wants children to see how their interests could connect to work.

Looking ahead, Konzelmann is excited for her children to contribute more ideas of their own. At Labs, she sees new models expanding the set of possible products. The ideas she values connect technical capability with a real user need in a way that surprises people.

Yang describes treating AI more like an agent responsible for a whole problem than a tool for a single task. Konzelmann says expectations are shifting, though different users want different things. The product challenge is to show people useful experiences rather than leave them stuck in front of a blank prompt box.