20VC: The $100 Billion AI Assistant Race: Town vs Instinct vs GrokBot | We Spend $75K Per Engineer on AI Tools | Why the AI Assistant Market Is Not a Bubble & AI Assistants Will Replace Every App on Your Phone with JD, Founder of Town

JD, former CTO of Plaid and founder of Town, presents an in-depth analysis of the emerging AI assistant market and Town's position within it. Town is an AI assistant living in email and calendar that recommends and executes automations based on user behavior patterns. After initially building an AI tax preparation company that achieved some product-market fit but insufficient differentiation, JD pivoted to Town after recognizing the opportunity in the email and calendar space when models like Claude Opus became capable of real agentic work.

On competitive dynamics, JD argues that talking about moats is premature—the priority is achieving deep product-market fit with mainstream users. He contends that products like Grokbot are powerful user products rather than mainstream products. The real defensibility, he argues, will come from network effects at the agent level, specifically through a feature called agent-to-agent where one user's AI assistant queries a colleague's AI to find information, creating strong switching costs once entire teams adopt the platform.

JD addresses the cannibalization threat from frontier model providers (Google, Apple, OpenAI, Anthropic) directly. He acknowledges these are top three priorities at major tech companies but maintains that these players must copy someone's successful mainstream product first—they won't innovate their way to mainstream adoption. He identifies Meta and WhatsApp as more concerning competitors due to their existing distribution advantage. The future, he predicts, will involve one to three entry points into digital interaction (personal, work, and possibly hardware), separated primarily by privacy and data silo considerations rather than multiple different AI systems.

On the agent-human relationship, JD makes a provocative claim: within five years, users will trust their agents to decide what data to share with other people without human intervention. He explains this through information theory—LLMs are more effective with more data access, and agents will be better than humans at maintaining appropriate privacy boundaries around truly sensitive information (medical, salary history) while allowing legitimate information sharing. He illustrates how agents equipped with company-wide email and calendar access could identify the right person to facilitate introductions, dramatically improving business outcomes.

On model infrastructure, Town uses a routing approach based on cost-effectiveness and consistency. For user-facing outputs, particularly voice and personality, JD maintains consistency across model families because switching models creates noticeable discontinuities that degrade user experience. For pure reasoning tasks not directly shown to users, he's more flexible about routing. He details how different model providers serve different needs: Anthropic for consistency, 11 Labs for voice, and others for specific tasks.

JD discusses economics extensively. Currently, Town's tasks run mostly on frontier models, but he predicts this will shift as capabilities push to open-weight models and smaller specialized models. The concerning endgame scenario is if 20-30% of workloads permanently require frontier models—in that case, competing with suppliers while paying them 70% margin creates unsustainable economics. He dismisses token maximization as a metric, instead viewing success as users paying consistently because they perceive clear ROI, and emphasizing the importance of informing users about token-expensive rogue routines.

On pricing, Town offers plans at $15, $49, $99, and $199 monthly with usage-based overages. The $15 plan is most subsidized and least profitable; the $99 plan generates the strongest economics. JD deliberately pursues paid business use cases from day one rather than subsidizing to chase consumption metrics, believing this ensures real product-market fit and prevents dependency nightmares where customers later discover unexpected costs. He rejects fully subsidized growth models, arguing that customers value knowing their costs and that charging forces important market feedback about value delivery.

On network effects and expansion, JD observes that teams with power users who build custom skills and integrations see faster adoption, since all team members inherit those automations. Identifying and empowering these tinkerers can create viral adoption loops. Certain underserved functions—particularly sales operations and recruiting—show surprising adoption because existing AI solutions haven't effectively reached their email and calendar-centric workflows.

Regarding the AI assistant market broadly, JD argues it's a $100 billion opportunity if a company can achieve 10 million paying users at current economics (~$700 per user annually). The competitive landscape is intense but stratifying rapidly—he estimates only 2-3 startup competitors truly matter now, down from 15 at the beginning of the year. Companies like Cursor, Grokbot, and Instinct are moving at startup speeds despite having advantages from larger players or existing distributions.

On infrastructure and tooling costs, Town spends approximately $75,000 per engineer annually on AI tools (Devon, Cursor, Codex, Claude), though this breaks down to roughly 1.5 engineers worth of value after accounting for equity. JD views this as clearly positive ROI given the gold mine of product improvements, integrations, and features lying in front of the company.

JD directly addresses the skepticism around AI assistant valuations, noting that Instinct raised to $2.5 billion with no monetization while Town has both revenue and growth. He expresses strong ethical objections to the practice of announcing fundraises at inflated valuations while actual investment occurred at lower prices, viewing this as misleading to employees and investors alike.

On the speed of market development, JD identifies the central challenge: companies can now build at machine speed but can only learn at human speed. This means that as soon as a feature works, competitors replicate it within weeks, eliminating the traditional advantage of being first to market and extracting insights over months. This acceleration is unprecedented in his experience and creates immense pressure to execute.

Final major themes include JD's belief that the future will be paid-for assistants rather than ad-supported (citing token costs and misaligned incentives), his perspective on hiring (trusting existing strong employees to recommend top performers without extensive interviews), and his optimistic vision that AI will reduce toil and enable people to live better lives with more freedom and capability.