The situation
At the Datapizza and ElevenLabs Voice Agents Sprint we built Nàvia, a voice agent for people who sign up alone for a group trip, and we made the podium. At the AI Hackathon in Paris, in spring 2026, Raffaele and I started a workbench for ad campaigns that later became motive, a product of mine: it measures whether and how a brand shows up in AI assistants' answers.
What I did
With Nàvia we started from a specific problem group travel agencies have, before we even thought about the voice. With motive the hard part came after the hackathon. In July I brought the demo into a real product, with a written list of what went in and what stayed out. In October I rebuilt the measurement method from scratch and cut the parts of the demo that promised more than the product could measure.
What changed
Nàvia stayed an award-winning prototype. motive now observes the answers AI assistants actually give, with all their variability, instead of asking a model what it would say. The part that impressed people most in Paris is gone.
What I learned
Group travel agencies lose a lot of people to last-minute cancellations. At the Voice Agents Sprint we started there: solo travellers book and then, in the weeks before departure, feel like outsiders. A voice agent for travel makes you think of itineraries and bookings, the moment before the purchase. The problem sat after it. Nàvia introduced you to the group, told you about the place like a local, made you a podcast about your trip and, yes, sang you a song too. Since then, when I look for a problem, I look first at the stretch between buying and using.
The first version of motive asked a model which brands it would name, once, with no language or country, on questions invented by another model. What came out was a plausible list that looked like a real measurement, which is why the error didn't show. It measured nothing of what a person sees when they put the question to an assistant. In October I rebuilt the method: now motive collects the answers a real person would see, and every question states where it came from, whether search suggestions, the brand's own site or the user.
The same question to the same assistant, asked twice, can get two different answers: the brand shows up in the first and is gone from the second. A check run once turns that variability into a verdict, in or out. So motive asks each question several times and shows how often the brand appears, with its margin of error. Past observations are never touched: if I change the method, the change shows up in the chart. Now I flag a movement only once it rises above the noise.
In July an agent built the motive MVP while I was away. It picked the price and launch channels itself and marked them as choices to review. The product worked, checkout included, and a working checkout makes a provisional price look final. The mock campaign launch, the piece that impressed people most in Paris, had the same effect: labelled as a simulation everywhere, it was still what gave the product its perceived value. In October I cut the simulation and changed the value proposition. Now an agent's choices stay separate from approved strategy until I've reviewed them.
Our sole focus should be delivering more value with every iteration of the product.
// Nàvia was a prototype: a podium, a lot of stickers, zero retention metrics.