Why a Good Business Fails to Become a Unicorn in the AI Era | AI Startup | Former Plaid CTO | Tow...

Original Media · 21m · 中文 · YouTube · Best Partners TV

TLDR

This episode explores how Jean-Denis Grèze (JD) pivoted his startup Town from AI tax preparation to an AI assistant for email and calendar after realizing the tax market couldn't support venture-scale returns. The conversation covers Town's product strategy, the shift from vertical to horizontal AI applications, the importance of neutral model providers and data ownership, and how AI agents will disrupt advertising and distribution models.

For you

PRO members can sign in and request more information — or their own version of this digest, tuned to what they care about.

Sign in to unlock

Surprising Insights

•

A viable $1M ARR business was deliberately shut down because it couldn't reach venture scale — a rare founder discipline.

•

The tax pivot window was only ~1 month (mid-October) due to the annual cycle; missing it meant waiting a year.

•

JD compares Town to a 'token allocation machine' — as token costs drop, the range of automatable tasks expands non-linearly (Jevons paradox).

•

Most small businesses use LLMs only as better search, not automation, because product experience isn't a text box.

•

Agent traffic already exceeds human traffic; companies like Walmart (no ad business) may gain from agent-friendly policies while Amazon loses.

Key takeaways

•

JD shut down a growing tax automation business ($1M ARR, projected $2.5-3M) because incumbents could adopt AI faster and the annual sales cycle prevented viral growth.

•

Town pivoted to a horizontal AI assistant for email/calendar after customers complained about meeting prep and repetitive email tasks.

•

Product strategy: prioritize consumer-grade UX for end-users (not buyers), stay model-agnostic, and let customers own their data/context.

•

AI agents will shift value from 'applications' to 'compute'; winning requires continuous adaptation to model improvements, not feature moats.

•

Distribution is uneven: early adopters are 'tinkerers'; mass adoption requires products that deliver value without user prompting/engineering.

•

Agent traffic will break the ad-supported web model; neutral, user-paid agents are needed to avoid marketing bias.

•

Enterprises will need private agents to mediate internal privacy and external negotiations at AI speed.

Notable quotes

“那一刻感觉就是失败,感觉自己真的搞砸了。”

▸ Notable quotes · 2:13

“如果创始人自己都不再相信结果,每天走进办公室、看着别人的眼睛说话,就变得很难。”

▸ Notable quotes · 2:33

“变化的速度就是人的速度。”

▸ Notable quotes · 6:35

“Town 像一台 token 分配机,让人自己决定把 token 花在哪。随着 token 价格下降,可以用 token 的地方会越来越多。”

▸ Notable quotes · 8:04

“你需要的,是一个中立的、属于你自己的智能体。它没有别的资金来源,你的钱包就是它唯一的资金来源。”

▸ Notable quotes · 18:00

“供应聚合的价值大大降低了,因为人的注意力是有限的,而现在它可以被自动化替代。”

▸ Notable quotes · 19:50

Chapters

•

0:00 — Origin: Ben Horowitz's message and the tax business: Ben Horowitz told Alex Rampell that Town changed his life. Rampell was surprised because Town's founder JD (Jean-Denis Grèze) had been running a promising SMB tax automation business, which he then shut down.

•

1:00 — Why the tax business was shut down: AI compressed the advantage window: incumbents could automate tax prep for existing clients faster than a startup could acquire new ones. The annual sales cycle, lack of daily engagement, and no viral loop made venture-scale impossible. JD and co-founder decided to pivot in a one-month window (mid-October).

•

3:40 — Pivot to email/calendar AI assistant: Customer feedback during tax product usage revealed pain points: meeting prep, forgotten context, repetitive emails. Town started building an AI agent for email/calendar workflows around Opus 4.5 release, targeting high-volume external-facing roles (sales, BD, recruiting).

•

5:00 — Product philosophy: consumer UX, model neutrality, data ownership: JD emphasizes consumer-grade UX for users (not buyers), model-agnostic architecture (switch providers in a day), and customer ownership of data/context. No training on customer data. Teams product mediates internal privacy via AI.

•

9:00 — Horizontal vs vertical, distribution challenges: Town starts horizontal to learn from thousands of users, then lets market pull verticals. Customization at scale via automated pattern detection (not manual FDE). Distribution is uneven: only 'tinkerers' currently unlock value; mass adoption requires delivering magic without user engineering.

•

17:00 — Agent traffic, advertising, and the future web: Agent traffic exceeds human traffic. The ad-supported contract is breaking. Neutral user-paid agents are needed to avoid marketing bias. Platforms like Amazon (ad-dependent) are disadvantaged vs Walmart. Supply aggregation value drops as attention is automated.

•

19:55 — Conclusion and open question: JD's path: kill a good old-world business, pivot fast, build neutral agent with user-owned context. Open question: when your agent conflicts with platform's agent, do you choose free-but-biased or paid-neutral?

Speakers

•

大飞 — Host

•

兰普尔 — Guest/Interviewer

•

JD (Jean-Denis Grèze) — Founder of Town

⋮