Source: YouTube — Invest Like The Best Link: https://www.youtube.com/watch?v=JUsb1FYOstA
Pricing: Access Products vs. Work Products
The single most pricing-relevant framework from this interview:
- Access products = seat-based pricing makes sense. The user is the unit. (ChatGPT Enterprise, Figma)
- Work products = outcome-based pricing. The work output is the unit, not the user. (Harvey should charge per contract processed, not per seat.)
- Seat Pricing breaks when the product's core value is doing work on your behalf — the user isn't the constraint anymore, the work output is.
The Zendesk Warning (directly relevant to SupportWire)
"Zendesk is a good example where literally Zendesk prices seats and each seat corresponds to a customer service agent that takes a certain number of tickets. That company should be worried. Why? Because I can have an AI agent sit right next to Zendesk and you can slowly siphon off. Instead of paying for 50 Zendesk seats, you can pay for 20 and have 30 AI agents sitting next to Zendesk."
- Companies pricing on utility-per-seat are most endangered.
- The siphoning happens gradually — it's a two-way door, not an all-in-one decision.
- These companies need to change pricing to outcome-based but it's brutal: going from $20-30/seat to $0.20-$1.00/ticket resolution.
- Many will need to go private to make the business model transformation.
Outcome-Based Selling
- You cannot lead with what your product does anymore. Lead with the outcome you can deliver.
- Palantir's model: "What's your most important business problem? Give us 6 months. If we solve it, pay us. If not, fire us."
- This maps directly to pricing: if you sell outcomes, you price outcomes.
Durability & Stickiness in the AI Era
Five Sources of AI-Era Stickiness
- Network effects — DoorDash is sticky because of the restaurant-dasher-consumer network. You can't vibe-code your way to liquidity.
- Money flowing through you — Financial services + software. Once money moves through your product, switching is regulated and painful. (Toast, Mercury)
- Hardware — Physical devices create switching friction. (Toast gives hardware free but charges to return it.)
- Access to a unique/scarce asset — Sierra's scarce asset is Brett Taylor. He can call any company in any country.
- Hamilton Helmer's Seven Powers — need a few embedded from day one.
Building Durable AI Applications
The CIO test: A Fortune 500 CIO said "Why would I use startups? Gemini has an agent builder. ChatGPT Enterprise has one. I have 1,000 IT engineers who want to be retrained as AI engineers. I'll just use horizontal tools."
To survive this, you need:
- Ownership of a scarce asset (license, regulation, unique data)
- Control point over how people interact with money or data
- Hardware that's hard to replace
- Part of an essential workflow
- Network effects
Systems of Record Are Fighting Back
- In 2025, legacy systems of record (Salesforce, Epic, Filevine, Slack) started seeing AI agent companies treat them as dumb databases.
- Three defensive moves: (1) blocking API access, (2) offering their own agents bundled free, (3) charging for API calls that were previously free.
- Agent companies now have no choice but to build their own system of record. The "system of action on top of system of record" thesis is dead.
- Migration is a 2-year engineering effort — you can't just present a shiny new system. The data is still there.
Product Development Is Changing
The New PM Role
- PMs are now "keepers of the why" — they articulate customer needs at the highest level.
- Actual product is built bottoms-up by engineers, researchers, PMs, and designers all working on code together.
- PMs are checking code into production repos using Claude Code / Codex.
- Companies now have prototyping interviews in the PM hiring loop.
- PM-to-engineer ratio going from 1:3-10 to 1:20.
Designer Role Is Shrinking
- When given the choice between an extra designer and an extra engineer, teams are choosing engineer.
- Design systems are already laid out. AI can leverage the design language to do designs.
- Small number of designers manage design systems company-wide; AI does the rest.
Non-Deterministic Software
- Old world: user does X, Y always happens. Deterministic.
- New world: user does X, Y happens. Slight variation of X, something completely different happens.
- PMs now own evals — they write AI to evaluate the results of AI, because humans can't keep up.
The Only Future-Proof Skill: Judgment
"In an era of infinite productivity, the question is what are the things to be productive on and are we building the right things?"
- AI slop is the #1 concern of every product leader.
- 1,000 AI engineers writing code = mountains of code. Who decides what's valuable?
- Product side: judgment on what to build and evaluating output.
- Engineering side: reviewing AI-generated code for bugs and vulnerabilities.
- Design side: does this make sense in the broader design system?
- Jack Dorsey called the PM role "product editor" — the job is to edit down, not add.
- Rick Rubin: "I'm not a producer, I'm a reducer."
Self-Serve as a Superpower
Larry Page's mandate at Google: everything built for large customers via internal tools must also be available to small customers via self-serve.
Result: small self-serve customers were the most sophisticated users — agencies, entrepreneurs, hustlers who exploit the system in ways you never anticipate.
Definition of self-serve: Customer can onboard AND use the product without ever talking to or engaging with a single employee.
This forces you to:
- Nail onboarding (where most drop off)
- Get to a moment of delight quickly
- Think like a consumer product, even for B2B
Self-serve opens the aperture: 100 salespeople reach 10K customers. Self-serve reaches millions. (Cursor is in every large company — 99.9% bottom-up, not top-down.)
Shifting Risk to the Transaction Level
Two powerful examples of the same principle:
Square: Instead of approving/rejecting businesses upfront (like banks do), accept 95% and run ML risk models at the transaction level. Gate later, not sooner.
Google AdSense: Sergey killed the publisher approval system. Instead: let everyone in, run the JavaScript on every page load, and only start reviewing URLs after they hit 100 impressions. Most things never reach the level where you care.
"Getting somebody to come to you and sign up is one of the rarest things in history. Someone is expressing interest and you're going to put 10 barriers? That's the opposite of self-serve."
Ads Business: Three Ways to Win
- Own a coveted first-party surface with identity + intent data. (Google had intent. Facebook had identity. ChatGPT has both — "the dream of any advertising person.")
- Drive outcomes at a certain cost. Don't own inventory, just deliver results. (AppLovin: mobile app installs at a target CPA. $100B+ company.)
- Be the exclusive provider for large demand sources. (Trade Desk: P&G gives them all non-Google/Facebook display budget.)
What doesn't work: being a middleman on top of Google/Facebook. They learn your capabilities and incorporate them.
CEO Communication: The Weekly Email Format
Three sections:
- Top of mind (60-70% of effort) — Product, business, and team. What's keeping you up at night. Be candid — people will rise to help.
- Performance update — How the company is doing on key dimensions.
- Miscellaneous — Recognitions, customer quotes, off-site announcements.
Key: don't be afraid of repetition. Say it once, twice, three times, four times — that's when it seeps into people's bones.
North Star Metrics
- Should indicate customer value, not revenue directly.
- Must be coupled with check metrics (guardrails) to prevent gaming.
- Square: GPV (gross payment volume) — not revenue, but shows growth in payment processing.
- Facebook: MAU → DAU (engagement signal).
- DoorDash: GMV — but checked against gross margin % and customer retention %.
Hiring in the AI Era
- Hire doers and builders, not managers. Don't hire managers as long as possible.
- Span of control < 10 should not be allowed. Either manage 50 people or be an IC.
- The future: functional experts who build and orchestrate armies of AI agents.
- Always use work projects in interviews. Engineering has always done this. Every other function should too.
- Best candidates reject the premise — they question whether the thing should be built at all.
- Tony Xu's test: Give someone $10-20, ask them to acquire 1,000 DoorDash customers. Nobody comes close. The point is seeing how many different things they try.
Red Flag: Job Hoppers
- 12-18 months then move = immediate red flag.
- Minimum 3-4 years to have impact at a company.
- "Tons of managers wrote to me saying it's an immediate red flag."
Founder Evaluation
Two key questions:
- "Tell me your founding story." — Looking for authentic lived experience that compels them to work on this problem. Not "my buddy and I wanted to start a company."
- "How did you navigate the idea maze?" — Why this solution over five other ways to solve the same problem? Are they students of industry history?
Quotable Lines
- "You either die or you live long enough to become an ads company."
- "Getting somebody to come to you and sign up is one of the rarest things in history."
- "In an era when you can do everything, the question is which of these things matter."
- "The half-life of software today is so short."
- "You can't vibe-code your way to network effects."
- "Jack called the PM role 'product editor.'"
- "Don't be afraid of repetition. That's when it seeps into people's bones."