Jean-Denis Grèze | 20VC Summary

1. Key Themes

The AI Assistant Market Is a Blue Ocean, Not a Bloodbath Yet

Despite intense competitive anxiety, JD insists the category is barely penetrated because most consumers still use AI as a "Google enhancer" rather than an assistant. This reframes the "crowded market" narrative that dominates investor discourse.

"Look, it's a blue ocean market, right? You got to understand... when we go to most customers, they've not heard of anything. It's blue ocean. Because people are using ChatGPT as a Google enhancer. Like that's the market." 00:35:33

Moats Are a Luxury for the Unsuccessful — Distribution and Network Effects Matter More Than Technology

JD explicitly rejects moat-talk as premature, arguing the real battle is reaching product-market fit at mainstream scale, after which network effects (not model quality) become the differentiator.

"I think talking about moats is a little bit of a luxury and you have to be more successful than Town today for it to matter... I think the product in this category that will win will have a network effect at the agent level." 00:00:00

Multi-Agent Network Effects ("Agent-to-Agent") as the Real Defensibility

Town's standout feature lets a user's agent query a coworker's agent directly, creating organizational lock-in once a whole team is on the platform — a form of stickiness JD believes nobody else has cracked.

"Once you have your whole team on that it's actually really difficult to imagine moving to a different product and I think no one has figured out multi-user multiplayer AI today like I think that's the thing that will be the moat." 00:08:38

The Frontier-Cost Problem Is the Real Long-Term Threat to Unit Economics, Not Competition

JD's biggest structural worry isn't a competitor copying features — it's that a meaningful share of workloads will remain "frontier" forever, meaning AI assistant companies are permanently competing with (and paying) their own model suppliers, compressing margins the way it did for Cursor.

"If you're paying your suppliers and competing with your suppliers at 70% margin eventually it gets like a little bit difficult... I'm still competing with OpenAI and Anthropic and I'm just giving them money for the 20 or 30% of workloads that are at the frontier for me. And that's what makes the economics not work." 00:49:02

Speed of Iteration Has Collapsed — Learning, Not Building, Is Now the Bottleneck

JD describes an unprecedented compression of competitive cycles: features that once took competitors years to copy now take two to four weeks, meaning startups can no longer "milk" product insights over a long runway.

"You can build now at the speed of machines, but you can only learn at the speed of humans... when we started the company... there's like 15 competitors... And now I'm probably down to like two or three competitors." 00:00:00

Enterprise (Work) Assistants Have Structurally Better Long-Term Economics Than Consumer

Given a choice between 100M consumers at $20/month or 1M business customers at $100/month, JD firmly picks the larger, lower-paying consumer-style base for work use cases because work-driven AI usage compounds (more automation → more revenue → more willingness to pay), unlike personal use cases which are capped by finite daily needs.

"I think over time in the work setting, AI will be used to do more and more and more for people. So I think the long-term potential for going for like NR, driving more revenue per user is extremely large over the long-term... Whereas in the personal sphere, it doesn't feel like that to me." 00:45:08

Data Silos Will Dissolve as Trust in LLMs as Privacy Filters Increases

JD predicts a fundamental societal shift: humans currently act as filters deciding what information to share; within five years, people will delegate that filtering judgment to their AI agents, unlocking far more effective cross-organizational agent collaboration.

"I think you'll trust your agent to decide what data to share with other people without you intervening in five years." 00:00:00

AI Agents Will Make Fewer Costly Mistakes Than Humans, Not More

Countering fears about agent errors (e.g., Jason Lemkin's Apple Watch anecdote), JD argues LLMs will soon outperform humans on judgment calls about data-sharing and task execution, comparing it to a smart, trusted employee occasionally making an honest mistake.

"I think the LLMs will make many fewer of these mistakes than humans pretty quickly." 00:18:02

Apple's Structural Handicaps: Not a Cloud Company, Over-Committed to On-Device Privacy

JD gives a sharp, specific critique of why Apple's agent roadmap will lag despite owning the device layer — their DNA and privacy positioning actively work against building an effective cloud-scale assistant.

"They're not a cloud company. They're just not. It's just not their DNA... they've like contorted themselves for competitive reasons around a privacy and on-device story that is like absolutely like puts them far away from the frontier." 00:37:41

We've Passed the Point of No Return on Human Code Review

A throwaway but massive claim: the overwhelming majority of production code is now AI-written and unreviewed by humans line-by-line, a permanent civilizational shift comparable to industrial-era chemical regulation.

"We've passed the point where humans will read every line of code. That is never happening again... in the history of humanity, we have passed the point where we will go back to a world where humans are looking at lines of code." 00:39:52

2. Contrarian Perspectives

Founders Shouldn't Chase Token Usage as a Success Metric

Against the common AI-company instinct to maximize usage/tokens consumed, JD argues that's a dangerous vanity metric that leads to customer resentment and churn if ROI isn't crystal clear to the user.

"The problem with token maxing... you want people to think is the ROI of using AI here worthwhile?... I think of success as paying me because if you're paying me every month, that means I'm mostly doing a job of delivering enough value." 00:41:53

Rejecting the "Burn the Boats" Subsidization Strategy Even When Capital Is Available

Despite having access to more capital and a market that rewards pure growth, JD refuses to fully subsidize usage indefinitely, insisting on monetizing early even at the cost of growth — a contrarian stance in a market obsessed with land-grab economics.

"As soon as you get three to five team members I really want to make money on that side I really want to make sure I'm..." 00:59:55 ...on the business side: "what I've learned is on the business side they don't like it if you don't charge them because they don't know how much it's going to cost." 01:01:09

Priced Rounds With Manipulated Headline Valuations Are Unethical, Not Just "How the Game Is Played"

JD directly criticizes a common Valley practice (announcing a round at an inflated blended valuation where only a sliver of capital was raised at the top price) as harmful to employees and dishonest — an unusually blunt on-record ethical stance from an active fundraiser.

"I have seen deals where it's like I invested like 200 and then the announcement is at 500... I don't think it's ethical... you can't look at someone in the eyes and say an investor that made 80% of their investment at like a 200 or 250 million dollar valuation but hey they put the last 20% at 500." 00:56:39

Coding Model Choice Barely Matters to Users; Voice/Personality Model Choice Matters Enormously

Counter to the assumption that model routing decisions are primarily about raw intelligence/benchmarks, JD says for user-facing personality and voice, consistency trumps intelligence — while for code, only functional correctness matters, not "which model wrote it."

"For coding, it matters less, interestingly, because for coding, you're like, does it work?... Before when you're talking or speaking to an assistant, if suddenly it's twice as verbose... people don't like that. They will literally write tickets." 00:21:57

GrokBot's X/Twitter Integration Distribution Advantage Is Largely Irrelevant to Winning the Real Market

JD dismisses the idea that being embedded in a massive social platform is a decisive edge, because the power users active on X are not representative of the mainstream market that will actually determine the category's winner.

"I think a person on X that uses these products is not actually product market fit, meaning that those are not the mainstream users... I don't think they think winning the power user slash influencer on X is where the market is. That is not where you win the market. That's the early adopter market." 00:37:01

3. Companies Identified

Town (Town.com) — AI assistant embedded in email/calendar that auto-recommends and executes work automations ("Townies"); JD's own company, three months post-launch with strong mainstream PMF and a >15% free-to-paid conversion rate.

"The payment rate for us on acquisition is like more than 15% of users who try the product end up paying for it, which is extremely high for PLG." 00:50:12

Instinct — Consumer AI assistant scaled to a $2.5B valuation, backed by Index and Benchmark; JD sees it as pursuing a different (subsidized, consumer acquisition) monetization strategy than Town.

"I don't think Instinct and Town are trying to do the same thing... I see a strategy that's more like customer acquisition, like with a free product that's fully subsidized right now." 00:00:29

GrokBot — xAI's assistant integrated with X; viewed by JD as the closest strategic competitor to Town in market approach, though currently more of a "power user product."

"GrockBot's really cool it's awesome but it's a power user product it's not a mainstream product." 00:08:09

Anthropic (Claude/Opus) — Frontier model provider praised for maintaining consistent model "personality" across versions and moving at startup-like speed despite scale.

"Anthropic spends a lot of time... actually making sure all of their model families roughly don't change too much in terms of their personality." 00:21:29; "you got to be honest, like Claude, like Anthropic is very fast." 00:34:15

Cursor — Coding tool cited repeatedly as operating at startup speed with massive R&D investment; also cited as a cautionary tale for margin compression from competing with model suppliers.

"Codex is like 100 people making that thing better... If you're paying your suppliers and competing with your suppliers at 70% margin eventually it gets like a little bit difficult. That is the part... I think this is what happened to Cursor right at the end." [00:32:10, 00:49:02]

OpenAI (Codex, ChatGPT) — Frontier player whose capabilities Town must match; also cited for weak/limited product suggestions relative to Town's onboarding approach.

"If Codex can do something that you cannot do and that thing is something that matters to users, it's over." 00:31:40

11 Labs — Best-in-class voice model provider, but flagged as extremely expensive with uncertain long-term differentiation versus commoditizing open-weight alternatives.

"On 11 Labs, like we're users of 11 Labs. They sound the best... It is very expensive. What I don't know is if it tops out." 00:47:15

Meta / WhatsApp — Identified as JD's most-feared competitor for personal-use AI assistants due to existing massive distribution.

"The competitor I would worry the most about would be would be probably for personal use cases it would be Meta and WhatsApp right. They're going to have a personal assistant that's going to come in WhatsApp." 00:09:37

Apple — Discussed at length as structurally disadvantaged (not a cloud company, over-indexed on-device privacy) despite owning distribution via devices; new Siri expected to lag capability-wise.

"It's going to be good, but it's going to feel not nearly as powerful as Town or Grokbot... it's going to be like nine months away, I think capability wise." 00:38:38

Google — Along with Apple, cited as having Town's category as a top-3 corporate priority within 12 months.

"I know what I'm building is a top 3 priority at Google and Apple like in the next 12 months." 00:00:00

Harvey — Legal AI referenced as a specialized "work agent" model for vertical (legal) use cases in a multi-agent future.

"You're an investor I think in Harvey or Lagora... when you're a lawyer and you're talking to your assistant about legal things and it's immediately just talking to Lagora." 00:12:10

Lagora — Legal AI platform (Harry Stebbings is an investor); used as example of vertical agent specialization due to privilege/privacy needs.

Plaid — JD's former employer (CTO), used as reference point for competitive hiring intensity and understanding of scaled consumer products.

"When I was at Plaid and we were competing with talent for like Stripe, that felt no harder than what I'm doing now." 00:32:34

Dropbox — JD's earlier employer, referenced as a comp for freemium-to-paid conversion economics and plateauing paid user growth.

"There was a huge pack of free users that were very costly on the cost side and then on the paying side... it did flatten out at some point." 00:58:20

Base10 — Venture fund; cited by JD as one of his best angel investments.

"Oh, it's either Base10 or Modal right now those are the first two that come to mind." 00:57:53

Modal — Infrastructure company; cited alongside Base10 as JD's best angel investment.

Devin (Cognition) — Coding agent Town's engineering team uses heavily for bug fixes and smaller tasks via Slack integration.

"We use Devin a lot... for a lot of bugs that come in for a lot of simpler little things or little visual tweaks we are kicking off Devin." 01:04:07

Salesforce — Referenced via Marc Benioff's comment about Anthropic spend, used to illustrate industry-wide AI tooling spend scale.

"Mark Benioff said at Salesforce we spend 300 million on Anthropic." 01:03:37

4. People Identified

Jean-Denis Grèze (JD) — Founder/CEO of Town, former CTO at Plaid; pivoted from a failed AI tax-prep startup to Town within three months, achieving product-market fit almost immediately by exploiting the moment agentic models (Claude Opus) became viable.

"The prototype we built a quick prototype in a couple of weeks and it had product market fit almost immediately." 00:05:48

Jason Lemkin — Referenced twice: once for his anecdote about an AI agent's runaway goal-seeking behavior (trying to buy six Apple Watches), and again as an example of a high-token-spend "worst customer" on Anthropic's Pro Max plan.

"His agent going off and trying to buy six AP watches for him to increase culture in the company... it was prevented because they needed engraving." 00:18:22; "He spends about $15,000 of tokens. He is the worst customer for Anthropic." 00:43:48

Marc Benioff — Salesforce CEO, cited for publicly disclosing Salesforce's massive Anthropic spend as a benchmark for enterprise AI tooling budgets.

5. Operating Insights

Force the Hard Onboarding Ask Upfront Rather Than Easing Users In

Town's core growth insight was requiring users to connect email/calendar before any value is shown — a deliberate hard paywall-style friction point that filters for serious users and enables high-value personalization, despite ~30% immediate drop-off.

"You have to connect email and calendar. You cannot use our product if you don't do those things but if you do those things we can do all this magic for you... 30% right off the bat." [00:51:53/00:52:00]

Identify and Empower the "Tinkerer" on Each Team to Drive Organic Enterprise Adoption

Rather than building for power users directly, Town watches for a single tinkerer inside a customer account who builds shareable automations/skills, and treats that person's presence as a leading indicator of team-wide adoption — worth explicitly instrumenting for and incentivizing.

"One of the predictors actually is, is there a tinkerer on the team? So one of the questions we ask ourself a lot is can we identify those folks and can we make it easier for them to create virality for their other team members." 00:27:25

Target Underserved Internal Functions (EAs, Chiefs of Staff, Junior Finance, Recruiters) Rather Than Oversaturated Ones (Sales Ops)

A tactical GTM insight: functions already flooded with AI vendor outreach (sales ops) are worse initial beachheads than functions still operating manually out of email, which convert faster and become internal champions.

"If you're a sales ops person and you do not have 10 emails a day from like an AI company... something's wrong. But then there are other functions like executive assistants, chiefs of staff... that don't have that much AI in their day to day... as soon as they adopted, there's an interesting effect because they often work with like leaders or execs." 00:27:54

Segment Model Routing by Whether Output Is User-Facing (Optimize for Consistency) vs. Backend Reasoning (Optimize for Cost/Capability)

A concrete operating framework: don't apply the same model-selection logic company-wide — treat personality/voice-facing layers as needing brand consistency (stick with one model family) while reasoning-only backend tasks should aggressively route to the cheapest capable model.

"There is the part of the stack that deals with the user interface, like the feel and the personality. And there it is harder for me to just route wildly... Below that, when it's just pure reasoning... then there, yeah, I think it's very much a matter of finding using the best model for the task." 00:22:27

Separate Fundraise and Revenue Milestone PR Moments Instead of Combining Them

Harry's tactical advice to JD: never bundle a funding announcement with a revenue milestone announcement, since each is independently newsworthy and combining them wastes a PR opportunity — a repeatable playbook for founder comms.

"I would advise you to always separate moments too many times I see people like combine a fundraise with a revenue milestone do not do that those are two separate PR moments that can be made into two big moments not one." 00:54:06

6. Overlooked Insights

The $75K-Per-Engineer AI Tooling Spend Reveals a New ROI-Driven Hiring Calculus That Nobody Is Modeling Correctly

Buried in the rapid-fire closing segment is a genuinely significant reframe of how AI changes hiring economics: it's not about "smaller teams," it's that AI raises the revenue-per-engineer ceiling, which mechanically justifies hiring more people, not fewer — inverting the popular narrative that AI shrinks headcount. Combined with the concrete disclosure that Town's run-rate is "at least 75k per engineer" on tooling alone (split across Devin, Claude Code, Cursor/Composer, Codex), this is a rare specific benchmark for enterprise AI tooling spend intensity that most companies aren't yet approaching.

"AI means that engineer can generate more than they could have before so maybe before they could only generate 150k of revenue maybe now they can generate 250k of revenue so suddenly AI makes you hire the incremental person." 01:04:51 ...and: "the run rate's at least 75k per engineer" 01:03:37

Town's Refusal to Interview Referrals From Trusted Colleagues Is a Radical, Understated Signal-Extraction Hack

Mentioned almost as a throwaway "fun fact," Town's policy of skipping interviews entirely for candidates vouched for by a trusted existing team member (based on direct prior working history) is a meaningfully different hiring philosophy than the industry-standard rigorous-interview-for-everyone approach — essentially outsourcing signal to trusted internal references rather than a formal process, which is a scalable but rarely admitted practice among sophisticated operators.

"If someone on the team has worked very closely with someone else we don't interview them if it's a top person... this is one of the best people I've ever worked with... we are not evaluating whether they can do the core role." [01:02:35/01:03:34]