At a glance
WHAT IT’S REALLY ABOUT
Builders share lessons on AI-native products, agents, pricing, and adaptability
- The panelists describe how improved models (reliability, tool use, context, multimodality, and memory) unlock new AI-native workflows, from hyper-personalized outreach to production-ready app generation and proactive financial guidance.
- They emphasize agentic product design—building systems that can take actions autonomously—while highlighting practical constraints like token spend, latency, and the need for rapid feedback loops and sandboxed execution.
- Multiple pricing approaches are compared, with a recurring pattern of separating token/cost recovery from value capture, because true outcomes-based pricing is attractive but difficult to implement broadly today.
- The discussion stresses adaptability as the core operating principle: teams repeatedly re-architect retrieval and tooling, remove scaffolding as models improve, and rely on internal evals and customer signals to guide rapid iteration.
- Advice to founders centers on curiosity, getting hands-on with real customer pain, not assuming “crowded markets” are solved, and being willing to scrap systems as the technology frontier shifts monthly.
IDEAS WORTH REMEMBERING
5 ideasAI enables “impossible” go-to-market tactics when paired with unique data sources.
Clay’s example of analyzing dumpster colors via Street View to infer competitors shows how multimodal models turn previously unusable signals into automated campaigns and direct revenue actions.
Proactivity is essential in consumer AI products, but defaults can be dangerously expensive.
Silvia’s early always-on “semantic cron” approach led to runaway token spend; the fix was to start with manual tools, observe what users actually do, then codify safe defaults and automation.
Give agents autonomy, but invest heavily in fast feedback and execution environments.
Emergent avoids over-scaffolding around models and instead provides a “laptop in the cloud” (sandbox containers, logs, ports, DB visibility) so the agent can build, run, debug, and iterate like a human developer.
Smarter models can reduce cost because they waste fewer cycles.
The panel notes a counterintuitive dynamic: intelligence improvements can lower total spend by cutting retries, errors, and unnecessary steps, even if per-token pricing isn’t dramatically cheaper.
Enterprise standards (brand voice, compliance, review) lag behind model progress.
Even as prompt reliability improves, GTM workflows still require human review due to reputational risk, meaning product design must accommodate governance rather than assuming full automation is acceptable.
WORDS WORTH SAVING
5 quotesWaste Management is like a trash services company, and they're one of our customers. And, and they use Clay actually to look at Google Street View images of people's homes and their businesses, and they use Anthropic's, um, APIs to actually an- use image an-analysis and analyze the Google satellite views to look at the color of the dumpster outside people's businesses.
— Varun Anand
We made a big mistake early on, which was assuming what the user would want as their semantic crons. And so we had, uh, what you would call like a radar, but it was running twenty-four/seven on a user, and it blew through token usage for us.
— Shane Noor
It's almost like you always underestimate what the new model is, is going to be capable of.
— Madhav Jha
We chose to make it more complicated, but we're covering our costs on one side. So we have one type of pricing that's just kind of low margin. We're not trying to make that much money on it, and it's trying to cover our costs. And then there's another metric that's trying to represent the value that we're creating, and that's more, um, trying to tie to the value that we're creating for customers, and that's high margin.
— Varun Anand
But most of those products don't work. Like, they really just don't work. And if you just, like, use them or talk to people, like, you'll find that they just don't work. And so just destroy them, you know? Like, go for it.
— Madhav Jha
High quality AI-generated summary created from speaker-labeled transcript.
