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

Summary

JD Grèze positions Town as an AI work assistant embedded in email and calendar that learns existing routines and executes background automations for mainstream users. He argues the category is strategically important for Google and Apple, but says the real moat is likely agent-to-agent network effects inside teams, not traditional software defensibility. He also makes a strong privacy and interface prediction: people will likely have only one to three digital entry points, and agents will eventually decide what data to share across contexts. On economics, he says Town routes across frontier and cheaper models depending on task, remains mostly frontier today, and faces margin pressure on the most valuable workloads where suppliers have leverage. He frames the market as rapidly copying features within weeks, making speed, low-friction onboarding, and sustained product-market fit more important than hype.

Notes

• Town is an email-and-calendar-native AI assistant that watches how people work and automates routine tasks in the background.
• Town says its initial product insight was that connecting email and calendar gives enough context to recommend useful automations.
• The company forces email and calendar connection, accepting about 30% immediate churn to improve recommendation quality.
• Town reports more than 15% of trial users convert to paid, which JD attributes to the quality of the work recommendations.
• JD says Town’s audience includes mainstream users, but the strongest PMF pockets can appear in unexpected personal workflows like families.
• He thinks the AI assistant category is strategically important enough to be a top-three priority at Google and Apple.
• JD says moats are secondary until a product has real traction; category success comes first, defensibility later.
• He believes the eventual moat in AI assistants will be network effects at the agent level.
• Town’s most distinctive network feature is agent-to-agent delegation across coworkers.
• He thinks once a team adopts shared agent workflows, switching products becomes hard.
• JD expects each person will have only one to three AI entry points into digital life.
• He thinks privacy boundaries will keep work and personal data partly separated across those entry points.
• JD predicts agents will eventually decide what data to share with other people without a human intervening.
• He gives a concrete privacy example where an agent would refuse to reveal medical history even in casual social settings.
• He argues LLMs will make fewer accidental data-sharing mistakes than humans.
• The human role, in his view, is to set goals and allocate token budgets, not to micromanage execution.
• Town dynamically routes tasks across model providers based on task type, cost, and output quality.
• Voice generation is treated as a separate choice, with ElevenLabs used where appropriate.
• JD says product performance shifts weekly or biweekly as model capability changes rapidly.
• Town is still mostly using frontier models for core work.
• He says some tasks can already be handled below the frontier, such as simple email labeling.
• Town also uses cheaper non-frontier models from frontier providers for lower-intelligence tasks.
• JD says moving to open-weight models would lower COGS but not improve product quality enough to matter.
• He sees onboarding as a low-friction magic moment: no configuration, immediate value, then curiosity about more use cases.
• Town deliberately brands the assistant as a named agent with an image to build a human-AI relationship.
• He thinks this relationship design is part of the product’s durability.
• JD warns that supplier competition is hardest on frontier workloads, where Town has no pricing power.
• He worries about the economics of competing with model suppliers while paying them for a large share of value.
• JD says the AI assistant market is blue-ocean today because most customers have not heard of the products yet.
• He says many users still use ChatGPT mainly as a search enhancer rather than as a work assistant.
• Town says it monetizes companies for work use cases, with multiple team members reinforcing adoption.
• He argues Apple is disadvantaged because it is not a cloud company and does not have the right data footprint for agents.
• He says Apple’s privacy and on-device strategy puts it away from the frontier.
• JD says business AI cannot be wrong, and human review of every line of code will not return.
• He says production code will be guarded by testing and by other models acting as attackers.
• JD defines success as a customer who keeps paying monthly.
• Town now warns users when routines are burning too many tokens.
• He says users trust the product more when it helps prevent wasteful token usage.
• Town’s pricing tiers are $14, $49, $99, and $199 per month.
• He says the $15 plan is the most subsidized and has the worst unit economics.
• He says the $99 plan is the most profitable overall.
• Usage-based pricing kicks in after plan limits.
• He prefers acquiring paying work users early because AI usage can expand and raise NRR over time.
• Voice matters, but current systems still need to get about twice as good, especially on tone and emotional read.
• JD says the market can now copy working features within two to four weeks, compressing startup learning cycles.
• He says the pace of the market is unlike anything he has seen before.
• He says market learning used to give startups a long head start, but that advantage has compressed sharply.
• JD says Town’s bull case to $100 billion requires 10 million paying users.
• He says Town already makes over $700 per year per user.
• Town’s early beta saw some users spending thousands of dollars per month on compute before pricing discipline existed.
• JD says Town spends at least $75,000 per engineer per year on AI tools.
• The stack includes Devin, Cursor, Claude Code/Codex, and Town itself.
• He estimates annual compute spend around $1 million.
• He roughly equates that spend to three to four engineers, or maybe 1.5 on an equity-adjusted basis.
• JD says skepticism about the market still keeps him up at night.
• He frames the opportunity as a hard path to a large treasure rather than a low-risk bet.