The Twenty Minute VCTown vs Instinct vs GrokBot | Why the AI Assistant Market Is Not a Bubble
At a glance
WHAT IT’S REALLY ABOUT
Why AI assistants will win: networks, trust, and cost curves
- Town pivoted from an AI tax product to an email-and-calendar-native AI assistant after prototypes showed immediate product-market fit as models became more agentic.
- JD believes the assistant category is not a bubble because a winning product could become the primary interface to digital work, but the market is still early and mainstream PMF is not yet fully solved.
- He expects durable moats to come less from model advantage and more from multi-user agent network effects, accumulated integrations/context, and strong product opinions about human–assistant relationships.
- Economically, the big uncertainty is what fraction of tasks will remain at the frontier (expensive, supplier-controlled) versus shifting to cheaper/open models, which determines whether margins can approach SaaS-like profiles.
- The next wave of risk and opportunity is trust: agents will handle more autonomous actions and data sharing, creating both massive productivity upside and heightened cybersecurity/privacy stakes.
IDEAS WORTH REMEMBERING
5 ideasThe winning AI assistant may be the one with agent-to-agent network effects.
Town’s current defensibility thesis is that assistants become more valuable when they can securely collaborate across coworkers (e.g., one person’s “townie” asking another’s for an answer), creating switching costs that single-user assistants don’t have.
Users will prefer a small number of front-door agents, with specialist agents hidden in the stack.
JD argues you won’t want a different agent per app; instead you’ll have 1–3 “entry points” shaped mainly by data separation (personal vs work) and privacy/compliance needs, while specialist systems sit behind the scenes via tool/model routing.
Trust will shift from humans manually gating data to agents autonomously managing disclosure.
He predicts agents will increasingly act as privacy filters—deciding what to share and what not to—similar to how humans mentally arbitrate sensitive vs appropriate information today, but with fewer mistakes than people make (e.g., accidental reply-all, leaking spreadsheets).
Model routing is constrained as much by user-perceived personality consistency as by raw capability or cost.
Town routes across providers (OpenAI/Anthropic/others; ElevenLabs for voice) to optimize for ROI, but consistency of “personality” and UX can limit routing because users notice tone/verbosity changes and perceive it as quality regression.
The core margin risk is the share of workloads that remain frontier—where suppliers set prices and may compete directly.
Like many AI apps, Town is mostly using frontier models today; JD expects many tasks (email labeling, routine scheduling, repeated workflows) to migrate down the cost curve to cheaper/open models, but the unknown is what % stays “frontier,” which determines whether margins can resemble SaaS.
WORDS WORTH SAVING
5 quotesYou can build now at the speed of machines, but you can only learn at the speed of humans.
— Jean-Denis "JD" Grèze
I know what I'm building is a top-three priority at Google and Apple, like, in the next 12 months. Not a top-10 priority, like a top-three priority.
— Jean-Denis "JD" Grèze
I think you'll trust your agent to decide what data to share with other people without you intervening in five years.
— Jean-Denis "JD" Grèze
The product in this category that will win will have a network effect at the agent level.
— Jean-Denis "JD" Grèze
We've passed the point where humans will read every line of code. That is never happening again.
— Jean-Denis "JD" Grèze
High quality AI-generated summary created from speaker-labeled transcript.
