To Judge a Consumer AI Product, Don't Ask About Retention — Ask What Users Stopped Doing
A consumer product has to clear four gates: make people curious, actually change their behavior, make them spread it, and keep them. The sharpest test is a single question — once you started using it, what did you stop doing? If there's no answer, it hasn't squeezed into anyone's life yet.
The video won't play here. Listen to the audio instead:
The argument · tap a timestamp to hear it
Early stage is about utility, mainstream is about trust
Josh Elman's judgment: in the early-adoption phase, utility tends to beat trust — people are willing to try new things and to experiment with them. But if a product wants to go from early users to a mainstream market of hundreds of millions or even billions, trust eventually becomes the most important thing. The dividing line he draws is how a company handles data — whether it uses your data to make itself stronger and then trades on it, or only uses it to give you a better experience. He himself stopped using Google search long ago; when he needs to describe what he wants, he goes straight to Claude, GPT, Gemini, and he thinks the basic way people search online has changed permanently.
— Josh ElmanMicrodrama isn't about output volume, it's about creator identity
At Greylock, Josh invested in Musical.ly, and his post-mortem goes like this: Instagram is the "rear camera" — look at what a great place I'm in, what a great latte I'm drinking. Musical.ly was "I want to create something" — even with a pile of unwashed laundry behind you, with dance and music it becomes a work worth sharing. Short video made discovery easy, but after watching you want something deeper, you want a more complete story — that's the entry point for microdrama. AI lets people without a studio, actors or sets make the good story in their head, and the opportunity lies in finding those creators, giving them tools, and pushing them in front of more people; an entirely new story market could emerge. But he stresses it isn't about dumping content into every category — what actually sustains this wave of UGC companies is the relationship between users and creators.
— Josh ElmanToday's AI is still a single-player game
His observation: the vast majority of AI use right now is single-player mode — I have AI handle things in my own life, but it isn't plugged into my group, it doesn't understand the shorthand our group shares; pull it into a group chat and it only helps me with one thing, it can't let me and the same group of people do other things together. Yet group chat itself is already the social form that has genuinely grown over these past years — he's in far more local group chats now than a decade ago. The next step he imagines is agent-to-agent conversation: I have mine ask, yours is being asked too, and when the two match up they bring the people together. He describes a very specific awkwardness — you want to make weekend plans with someone, but you're afraid that sending it out will only get silence, and that kind of unaskable moment is exactly right to hand to an agent to arrange.
— Josh ElmanTo build a big company, first earn the right to do one small thing
He doesn't believe in doing it all at once: to become a very big company, you can't do everything from the start — you have to do a few things extremely well first, then step by step earn the right to do more. Discord is the sample — it first earned the position of "voice chat while gaming," then people started chatting there, chatting more and more, and eventually even people who don't game came to use it, and only then did it slowly become a place you can rely on for more things. The counterexample is Siri: it once did a good job earning the right to make calls, set alarms, set reminders and send texts, but when you want it to do a little more, it often can't. He says the chances for someone to say "try one more thing, try five more things" are limited.
— Josh ElmanVoice is lossy, a pure-voice agent doesn't hold up
Josh likes voice because he thinks while speaking, and speaks faster than he types. But he points out a property that's easy to overlook: voice is lossy — you ramble, you go off track, whereas typing actually forces you to tighten your thinking before you send it. So he deliberately thinks about where voice sits in the whole spectrum of communication: an agent should let you ramble at length, then help you distill what you actually meant and remember the core. He doesn't believe we'll turn everything over to voice; voice will always be a balance, and the concept of a pure-voice agent doesn't really hold up. In quiet settings, or when you want to polish a thought first, he'd rather type. He thinks we do far too little with voice today.
— Josh ElmanInference cost is back, and consumer payment doesn't come up front
He walks through the history of cost: in the early days, building a website meant buying servers, renting racks in a data center, hiring people to maintain them — his first week at LinkedIn, a dozen or so people across the whole company were called in to move machines from one rack to another. With the cloud and the AWS era, that heavy upfront capital disappeared and the product itself became more important. Now the cost is back, and it's a variable cost — inference cost, not fixed infrastructure. The hard part is consumer distribution: willingness to pay doesn't always show up at the start. Enterprise customers can put the expense on the company card; consumers pay out of their own pocket. The good news is that mobile games and subscriptions have cultivated paying habits, and with open-source models and the ability to switch models at any time, costs are coming down — he bets the two curves cross in the next few years.
— Josh ElmanFour gates: fresh, switch, spread, stay
The order in which he looks at a product is four steps. First, can it be explained in one sentence — the thing it gives that didn't exist before — and make someone raise an eyebrow; AI has made this step easier. Second, once people actually come and try it, is the experience good, can they figure out how it fits into their life, and it usually has to replace something they were already doing elsewhere — which is why he especially loves to ask: what did you stop doing once you started using this? A good answer means it can genuinely be adopted. Third, will users spread it for you on their own — sometimes it happens naturally through network effects, sometimes it spreads as fast as ChatGPT did, because it gives answers directly and you don't have to Google anymore. Fourth, will people stay; only with network effects, with a history that gets harder to move away from the more you use it, or with a creator ecosystem, have you really caught something.
— Josh ElmanFeatures can be copied in days, direction can't
He acknowledges that in today's competitive environment, good ideas get chased fast, whether the rival is another startup or a big company. His dividing line: others can copy what you've already built, but they can't copy where you're going, and they can't figure out where you're going before you do — so the best thing a founder can do is stay on their own path and live out that vision. He goes further: building an AI product today isn't just building a capable tool, it also needs personality, a relationship, an identity users are willing to identify with, a set of ways of interacting with you that gets more personal over time, and a point of view of its own. Users will say: I tried that one, but it's missing something that's genuinely special.
— Josh ElmanIn their own words · checked verbatim
It's never been easier to get consumers' attention. What's harder, I think, is getting me not just to try something, but to actually stick.
Josh Elman0:00
I think in these early days of adoption, utility often trumps trust. People will try new things and they'll experiment with it. But if you truly want to break into the mainstream and really get to the hundreds of millions or billions, trust ends up becoming paramount.
Josh Elman10:07
It's not that you can slop lots of things in lots of different categories, but it's really that relationship with the creator that is what's powered the most recent wave of these incredible user-generated content companies.
Josh Elman25:27
You have to start by doing a few things really, really well and earn the right to do more.
Josh Elman35:44
I'm not yet a believer though that we're all gonna move to voice for everything. I think it's always gonna be a balance and I think the idea of a voice-only agent doesn't really make sense.
Josh Elman38:45
But we also know that for consumer distribution, the willingness to pay doesn't always come up front.
Josh Elman43:51
And so I always like to ask people, what did you stop doing when you started using this new thing?
Josh Elman45:57
And the hardest part there is if you're that founder who has the idea and the insight, even if other people can copy what you did, they can't copy where you're going. They can't figure out where you're going ahead of you.
Josh Elman50:10
Figures
| GPT weekly active users | 900 million | 7:06 |
| Josh Elman's tenure at Apple | about 6.5 years | 12:08 |
| Year Discord and Musical.ly began growing | 2015 | 26:28 |
| Headcount at LinkedIn when Josh joined | about 15 people | 42:50 |
Glossary
- earn the right to do more
- Do one small thing extremely well and build trust first; only then will users authorize the product to take over more of their behavior.
- inference cost
- The variable cost incurred each time a model is called to generate a result, unlike the fixed server overhead of traditional cloud services.
- microdrama
- Very short serialized episodes distributed on platforms like TikTok and YouTube, whose production barrier AI is lowering.
- personal agent
- AI that handles a user's private affairs — schedule, health, finances — as opposed to an efficiency agent for work settings.
How to listen
Founders building consumer AI products and early-stage investors watching the sector, especially teams stuck on retention and pricing.
The opening career retrospective and the closing show promo can be skipped.