The Twenty Minute VCTown vs Instinct vs GrokBot | Why the AI Assistant Market Is Not a Bubble
CHAPTERS
- 0:00 – 1:24
Town’s vision: the AI assistant with agent-level network effects
JD opens by framing AI assistants as a top strategic priority for incumbents like Google and Apple. He argues the eventual winner won’t just be a better chat UI—it will have agent-to-agent network effects that create durable stickiness.
- •AI assistants are a top-3 priority for major platforms in the next year
- •Market winner likely determined by defensibility beyond model access
- •Agent-level network effects as the core moat
- •Assistants shifting from tools to primary entry points for digital work
- 1:24 – 2:10
What Town is: an email-and-calendar-native assistant that automates work
JD defines Town as an assistant embedded in email and calendar that observes patterns and recommends automations. The focus is mainstream users with fast time-to-value rather than power-user tinkering.
- •Town lives in email/calendar and automates recurring workflows
- •Learns user habits to recommend background automations
- •Mainstream ICP; product has been in-market for ~3 months
- •Positioning differs from other assistants by workflow entry point
- 2:10 – 4:09
From AI tax prep to assistants: pivot rationale and timing with agentic models
JD explains Town’s pivot from an AI tax business that achieved only partial PMF. The pivot clicked when models became agentic enough to do multi-step work and a prototype showed immediate pull.
- •AI tax product reached insufficient PMF; decision to reset
- •Opportunity spotted: no great AI operating out of email
- •Model capability inflection (agentic workflows) enabled new product
- •Prototype delivered rapid early product-market fit
- 4:09 – 5:49
Can Town survive Google/Apple/OpenAI/GrokBot? Winning before moats matter
Harry presses on cannibalization risk by incumbents and frontier providers. JD says the first job is deep PMF and a mainstream experience—big players copy proven UX rather than discover it first.
- •Incumbents are inevitable competitors; founder anxiety is real
- •Defensibility is secondary until strong mainstream PMF exists
- •Power-user products differ from mainstream adoption dynamics
- •Big players often copy after someone proves a winning UX
- 5:49 – 8:47
The moat thesis: multi-player AI and “agent-to-agent” collaboration
JD introduces Town’s ‘agent-to-agent’ feature where assistants can query coworkers’ assistants for answers. He argues multi-user, multi-agent coordination is underbuilt and could create strong lock-in.
- •Agent-to-agent allows querying coworkers’ assistants for missing context
- •Network effects emerge once whole teams are on the same system
- •Multi-player AI is a largely unsolved product category
- •Other moat theories: custom models, accumulated context, distribution
- 8:47 – 11:01
One agent or many: entry points, privacy boundaries, and specialist backends
They explore whether users will have one universal agent or multiple. JD predicts 1–3 primary entry points, with separation driven mainly by privacy and data-layer silos (work vs personal), while specialist systems may run behind the scenes.
- •Users want minimal entry points—not dozens of app-specific agents
- •Work/personal separation likely persists due to privacy and governance
- •Specialists (e.g., legal) may be routed to domain agents in background
- •Usability layer vs data-layer architecture will diverge
- 11:01 – 16:03
A future trust leap: letting agents decide what data to share
JD’s ‘hot take’ is that within five years people will trust agents to decide what information to reveal in social and work contexts. He argues AI becomes more valuable as silos loosen, shifting compliance from manual policies to model-mediated controls.
- •Agents will refuse sensitive requests (medical history/salary) by default
- •Humans currently act as the privacy filter; agents may assume that role
- •More data access increases assistant effectiveness, especially at work
- •Compliance could shift from labeling/policies to model-enforced boundaries
- 16:03 – 19:30
Error tolerance and agent goal-seeking: autonomy, budgets, and monitoring
They discuss how much mistakes erode trust and whether agents’ goal-seeking is dangerous. JD argues humans already make costly sharing mistakes, and proposes a framework: humans set goals and budgets, separate agents monitor execution, and autonomy is bounded.
- •LLMs may soon make fewer sharing mistakes than humans (reply-all analogy)
- •Goal-seeking autonomy is powerful but risky without constraints
- •Humans should set direction, token budgets, and oversight mechanisms
- •Monitoring may be delegated—but not to the same agent executing tasks
- 19:30 – 22:32
Model routing in practice: consistency of “personality” vs best tool per task
JD explains Town’s model stack and why routing is nuanced. For user-visible outputs, consistency in tone/personality matters; for hidden reasoning steps, they can swap to the best model for cost/performance.
- •Town routes across providers (e.g., Gemini/OpenAI for images, ElevenLabs for voice)
- •User-facing voice/text needs consistent personality; routing can break that
- •Anthropic model-family consistency is a practical advantage
- •Back-end reasoning can be optimized aggressively without UX disruption
- 22:32 – 26:45
Economics and margins: SaaS-like hopes, frontier dependence, and open models
Harry probes whether assistant businesses can reach strong margins. JD says many workloads will move down the cost curve (open-weight/smaller models), but the unknown is what share remains frontier—this determines long-term unit economics.
- •Near-term margins often subsidized; cost curve expected to improve in 18–24 months
- •Some tasks (e.g., labeling) don’t need frontier intelligence
- •Key unknown: what % of workloads stay at the frontier long term
- •Engineering time prioritized toward growth and product advantage over cost optimization
- 26:45 – 29:56
Go-to-market leverage: tinkerer-led internal virality and underserved functions
JD describes how adoption expands wall-to-wall: a single power user can build team skills/routines that others get for free. He also notes unexpected pull from roles underserved by AI (EAs, chiefs of staff, HR, recruiters) that operate heavily in email.
- •A ‘tinkerer’ on a team accelerates adoption through shared automations
- •Town aims to reduce reliance on tinkerers via out-of-box role-specific workflows
- •Underserved functions show strong demand and become expansion vectors
- •Fast ‘time to wow’ (briefings/action items) drives individual conversion
- 29:56 – 39:32
How VCs should think about the assistant race: capability parity, TAM, and speed
Harry asks how to evaluate the flood of ‘European Towns/Instincts.’ JD argues the market is expensive and fast; startups must keep capability parity with well-funded labs, have a credible distribution edge, and enough TAM/war chest to survive the arms race.
- •Local winners need a real reason: distribution, regulation, or unique wedge
- •R&D burden is largely ‘keeping up with the Joneses’ on capabilities
- •Capability gaps quickly kill willingness to pay vs bundled lab offerings
- •Market feels like it’s consolidating to a few serious contenders
- 39:32 – 41:50
Apple’s risk in assistants: cloud DNA, on-device constraints, and time lag
JD lays out why Apple could lose: weak cloud DNA and an on-device privacy posture that keeps them behind frontier capabilities. Even with device distribution, he expects Siri improvements to lag best-in-class assistants for a period.
- •Agents improve with broader data access; cloud matters
- •On-device focus trades capability for privacy and speed limitations
- •Apple’s device advantage helps distribution but may not offset capability gap
- •Winning may require cultural/leadership shift in approach to AI interfaces
- 41:50 – 43:42
Cybersecurity and governance: inevitable incidents, evolving best practices
They discuss whether agents create a security time bomb. JD argues we’ve crossed a point of no return where AI-generated code ships to production; the industry will adapt with testing, adversarial evaluation, and regulation akin to environmental controls.
- •AI-generated code at scale is irreversible; humans won’t review everything
- •Security will rely on guardrails: tests, model-based red-teaming, access controls
- •Incidents are inevitable early; best practices will mature over time
- •Enterprise AI requires getting security ‘not wrong’ to maintain trust
- 43:42 – 47:46
Metrics that matter: why ‘token maxing’ is a trap and pricing/whales dynamics
JD defines success as retention of paying users, not token usage. He explains why higher usage can destroy perceived ROI and churn, how Town flags ‘rogue routines,’ and how plan design balances subsidy, profitability, and usage-based overages.
- •Token usage ≠ value; optimizing for tokens can backfire on retention
- •ROI varies widely by role even for identical token cost workflows
- •Town builds trust by warning users about expensive, low-value routines
- •$15 plan most subsidized; mid/high tiers and usage-based pricing improve unit economics
- 47:46 – 59:47
Work vs consumer: where bigger businesses form, parents surprise, and GTM tradeoffs
JD prefers many users paying less, betting that work use cases expand monetization as AI drives revenue outcomes. He also reveals unexpected PMF with families/parents via school email and scheduling—sparking an internal debate about focus vs distraction.
- •Work use cases scale spend with business value; consumer use cases are episodic
- •Example: recruiting firms expand capacity and revenue via automation
- •Unexpected PMF: parents/families managing school and activities via email
- •Strategic debate: consumer virality vs enterprise expansion and monetization
- 59:47 – 1:16:00
Hype vs reality, fundraising ethics, and quick-fire: the long-term bet on AI
They touch on market skepticism driven by large raises without monetization, then JD shares views on PR, ethics in staged valuations, and a quick-fire covering angels, board needs, and AI optimism. The episode ends with JD’s belief that AI reduces toil and expands opportunity if navigated responsibly.
- •Skepticism is partly about subsidy and unclear paths to durable economics
- •JD avoids PR-for-PR’s-sake; prefers user-driven validation
- •Critique of misleading fundraising valuation optics and employee impact
- •Quick-fire: bull case scale targets, board finance expertise, tooling spend, optimism for societal uplift
