Context

a16z partner Josh Elman, who has worked in senior roles at LinkedIn, Facebook, Twitter, Robinhood, and Apple before returning to venture investing, joins Ollie Forsyth on New Economies to discuss what separates a consumer AI product people try once from one that becomes part of daily life. Elman's framing is specific and testable rather than abstract: he offers a concrete four-stage funnel (intrigue, adoption with substitution, evangelism, retention) and a specific diagnostic question ("what did you stop doing when you started using this?") for evaluating whether a new consumer product actually has staying power. The episode matters to PMs building consumer products because Elman ties each idea back to real cases (Twitter's retweet design, Discord's and Musical.ly's wedge strategies, Siri's capability-scoping mistakes) rather than staying theoretical.

The Big Idea

Getting a consumer's attention has never been easier, but earning a genuine habit, and the right to expand into doing more for that person, has never been harder, because a new product isn't just competing for novelty, it has to be good enough at one specific thing that people actually stop doing something else to use it.

Elman's diagnostic for testing this directly: ask any user of a product they've adopted, "what did you stop doing when you started using this?" If there's a clear, specific answer, that's evidence of real behavioral substitution, not just trial. If there isn't, the product hasn't actually earned a place in the person's routine yet, no matter how much initial excitement it generated.

Key Insights

Attention is now cheap and instant; behavioral substitution is the hard, still-scarce thing

Elman separates two problems that used to be roughly proportional: getting people to notice and try something (now easy, because content and product demos spread virally through TikTok, Instagram Reels, and YouTube almost instantly when something looks compelling) and getting people to actually stick with it (still hard, because it requires displacing an existing digital habit, not just adding a new one to an already-saturated digital life). He points to the COVID period as the moment that revealed people's practical ceiling for time spent online, meaning new consumer products increasingly have to substitute for something already in a user's routine rather than simply adding incremental time.

The four-stage funnel: intrigue, real substitution, evangelism, retention

Elman's explicit checklist for evaluating whether a consumer product will actually last: first, does it produce a clear, describable value that makes people raise their eyebrows when they hear about it; second, once someone tries it, do they have a good enough experience that it actually replaces something they were doing elsewhere (not just a nice-to-have addition); third, are users excited enough to spread it to others, either organically through network effects or simply because the value is so obvious; and fourth, does usage persist over time rather than fading after the initial novelty. He treats all four as necessary, and treats the substitution question specifically as the one that separates a viral moment from a durable product.

Discord and Musical.ly both won by deliberately not trying to do everything at launch

Elman invested in both companies in 2015, when Facebook, Instagram, WhatsApp, Twitter, and LinkedIn were already dominant, and argues both succeeded specifically because they solved one narrow, different problem well rather than competing head-on: Discord for people coordinating and voice-chatting during multiplayer games, Musical.ly for people creating short, music-backed video stories rather than only consuming others' "back-facing camera" highlight-reel content (his contrast with Instagram). Both companies then expanded gradually: Discord became a place people also just chatted, even without playing games, only after first earning trust as the best place to play and talk during games.

A product earns "the right to do more" incrementally, and Siri's history is the cautionary counterexample

Elman's central prescriptive claim for any assistant or agent product: don't launch by claiming to do everything, launch by doing a few specific things (placing calls, setting timers, sending texts, in Siri's case) reliably, and expand scope only once trust in that narrower scope is established. He uses Siri's own history as the negative case: once users learned Siri could reliably do those few tasks, every failed attempt at something slightly beyond that scope taught users to stop trying broader requests, and each of those failures cost trust that had to be rebuilt narrowly rather than assumed. The lesson: an assistant that clearly explains what it can't do, rather than failing silently or vaguely, preserves more trust for expanding scope later than one that overpromises.

Trust becomes the dominant differentiator only once a category moves from early adopters to the mainstream

Elman's read on adoption dynamics: in the current early-adopter phase of consumer AI, raw utility is enough to get people trying new tools, they'll tolerate imperfect trust tradeoffs for a genuinely useful capability. But he argues that as AI moves toward handling more sensitive parts of a person's life (payments, health information, personal data), trust in how a company uses that data becomes the deciding factor for reaching hundreds of millions or billions of users, not just early adopters. Products that treat user data purely as fuel for their own model improvement, versus products that treat it strictly as a means to serve that specific user better, will diverge sharply once this becomes visible to mainstream, less technical users.

Instant cloneability shifts competitive advantage from "what you built" to "where you're going"

Elman argues that in the current environment, a competitor can copy a visible feature or product within days, not months, which changes what's actually defensible. His position: since anyone can clone what already exists, the real advantage lies in a founder's specific, hard-to-replicate vision of where the product goes next, plus the cumulative relationship, personality, and identity a product builds with its users over time (which a feature-level clone can't replicate immediately). He frames this explicitly as a reason product identity and consistency of vision matter more, not less, as cloning speed increases.

Mental Models & Frameworks

The wedge-then-expand growth pattern

A repeatable structure Elman sees across Discord, Musical.ly, and, in his view, the strongest consumer AI assistants going forward: pick one narrow, well-defined use case, execute it well enough that users trust the product for that specific thing, and only then expand into adjacent behaviors once that initial trust is established. He explicitly warns against the opposite approach, launching as a general-purpose assistant or platform that claims broad capability from day one, because a product that hasn't earned trust in something narrow first won't be believed (or forgiven for early failures) when it claims to do everything.

The substitution test: "what did you stop doing?"

A single, concrete question Elman uses to separate genuine behavioral change from mere trial: ask an actual user what they no longer do now that they've adopted a new product. A clear, specific answer (I stopped Googling and go straight to a chat interface; I stopped manually renaming files) indicates the product has displaced an existing habit, which is Elman's bar for durable adoption. A vague or absent answer suggests the product is being used alongside existing habits rather than replacing them, a weaker, less durable form of engagement.

Trade-offs & Nuance

Free-to-use AI products face a cost structure that didn't exist in the previous cloud era

Elman contrasts today's consumer AI economics with the previous cloud-computing-driven wave: once AWS-style infrastructure made hosting nearly free at the margin, consumer products could scale to huge audiences without proportional cost growth, shifting competition purely to product quality. Inference costs for AI products are variable and ongoing in a way fixed infrastructure costs weren't, so giving a product away for free to build distribution now carries a real, continuing cost rather than a one-time infrastructure investment. His resolution isn't a single pricing model, he expects the market to eventually land at an intersection of falling inference costs (from cheaper models, open-source alternatives, and model switching) and rising consumer willingness to pay for digital value (citing mobile gaming and subscription normalization as evidence consumers are more comfortable paying than a decade ago), with ads, subscriptions, and usage-based pricing all coexisting depending on the product.

The best place to start a company isn't necessarily the best place to scale one

Elman explicitly separates two different claims about San Francisco and the Bay Area: he doesn't think it's the best place in the world to start a company (citing Lovable's Swedish origins and Musical.ly's roots in a Chinese team as counterexamples to the idea that great consumer products must start in the Bay Area), but he does think it's an unusually valuable place to scale one, because of the concentration of people who have personally been through the journey from early product-market fit to durable platform. His practical implication for founders: proximity to that expertise matters more than geography of origin, and founders building for a global consumer audience should be wary of over-indexing on Bay Area tech culture as representative of mainstream consumer behavior elsewhere.

Practical Application

Run the substitution test on your own product before claiming real adoption

Before treating early usage numbers as evidence of product-market fit, ask a sample of actual users the specific question "what did you stop doing when you started using this?" If you can't get a clear, specific answer, the product is likely being tried alongside existing habits rather than replacing one, and retention is at higher risk than raw usage numbers suggest.

Scope a new AI assistant or agent product to a narrow set of reliable tasks before expanding

When launching an assistant-style product, resist positioning it as general-purpose from day one. Following the Discord/Musical.ly/Siri lesson, identify the two or three tasks it can do reliably, ship those first, and treat visible, well-communicated failure on anything beyond that scope (rather than silent failure) as protecting the trust needed to expand scope later.

Audit whether your data practices would survive mainstream scrutiny, not just early-adopter tolerance

Since Elman argues trust becomes the dominant factor once a product tries to cross from early adopters to mainstream scale, explicitly evaluate whether your current data usage (training on user data, sharing with third parties) is described transparently enough to survive that scrutiny before you need mainstream trust, rather than retrofitting trust-building after a trust failure becomes visible.

Treat a cloned feature as a prompt to accelerate your roadmap, not defend your current feature set

Given that visible product decisions can now be copied within days, when a competitor clones a specific feature, resist spending significant effort defending or differentiating that exact feature. Elman's framing suggests the more durable response is continuing to execute toward your own specific next step, since a competitor copying your current state can't replicate a roadmap they don't know.

Bottom Line

Josh Elman's core framework, intrigue, real substitution, evangelism, and retention, gives PMs a concrete funnel for testing whether a consumer AI product has actually earned a lasting place in someone's life, and his repeated emphasis on wedge-first, trust-earned expansion (Discord, Musical.ly, and Siri's cautionary history) argues that the products that will matter most aren't the ones that promise to do everything, but the ones disciplined enough to do one thing well first and earn the right to do more.