ExpertsFMGokul Rajaram - Lessons from Investing in 700 Companies - [Invest Like the Best, EP.456]
Invest Like the Best with Patrick O'Shaughnessy · January 29, 2026
5 Selected Lessons:
framework · 3 reasoning_pattern · 1 case · 1
Five sources of durability for an AI application
Source
Invest Like the Best with Patrick O'Shaughnessy — Gokul Rajaram - Lessons from Investing in 700 Companies - [Invest Like the Best, EP.456] (January 29, 2026) · at 16:34 ▸ · at 20:01 ▸
When this applies
when building any product that could be subsumed by a more powerful general-purpose platform
The takeaway
A product's durability comes from controlling a scarce asset, chokepoint, hardware, essential workflow, or network effect — not from the feature itself.
Why it works
General-purpose platforms (foundation models, horizontal incumbents) will absorb any capability that isn't anchored to something structurally hard to replicate, so survival requires owning an asset the platform can't just build.
One level up
In a fast-moving technology wave, differentiate on structural asymmetries, because feature advantages erode within months.
Where it stops holding
This stops working when the 'scarce asset' is itself commoditized over time (e.g., a license expires, regulation changes, or the control point gets disintermediated).
Detail
procedure:
- Start with a deep, compelling problem in a vertical where agentic AI can now do a human job
- Target a high-value workflow that is complex and requires custom data, not a thin wrapper
- Check whether a CIO/customer could build it themselves using horizontal agent-builder tools
- Secure at least one durability source: scarce asset/license, control point over money or data, hard-to-replace hardware, essential workflow status, or network effects
- Plan to eventually replace the entire system of record, not just sit on top of it
benchmarks: {"time_horizon":"multi-year durability, not a one-year application"}
rationale: Without one of these, you'll either be copied by horizontal agent-builder tools or squeezed by the incumbent system of record.
In their words
So, what that means is you've got to have durability because ultimately, as venture capitalists, or even as an entrepreneur, your time horizon can't be building something that lasts for one year.
Three (and only three) ways to build a profitable ads business
Source
Invest Like the Best with Patrick O'Shaughnessy — Gokul Rajaram - Lessons from Investing in 700 Companies - [Invest Like the Best, EP.456] (January 29, 2026) · at 46:16 ▸ · at 49:14 ▸
When this applies
when deciding how any intermediary/platform business should monetize attention or access to a valuable audience
The takeaway
A platform monetizing attention must choose exactly one of: own the coveted first-party inventory, sell a measurable outcome, or become the exclusive channel manager for a budget — there is no fourth path.
Why it works
Advertisers will only pay for one of three things: direct access to a coveted first-party audience/surface, a guaranteed measurable outcome, or exclusive management of their spend — because these are the only forms of value an ad buyer actually cares about.
One level up
In intermediary businesses, value capture concentrates around a small, enumerable set of buyer-value forms; identify which one you structurally own before building a monetization model.
Where it stops holding
This stops working when advertiser behavior/measurement norms change fundamentally (e.g., attribution becomes impossible, or ad-free environments become structurally preferred — a risk Gokul separately flags with agentic AI interfaces).
Detail
procedure:
- Own a coveted first-party audience/surface users interact with (Google Search, Facebook identity, ChatGPT intent+identity)
- Drive a measurable outcome for advertisers without owning inventory (AppLovin: mobile app installs)
- Be the exclusive agency/distribution partner managing an advertiser's spend across others' platforms (Trade Desk)
rationale: These are presented as exhaustive — 'No other ways of making money' — based on direct ads-building experience at the two most important ad businesses.
In their words
There are three fundamental ways to succeed in the ads business. Three, and only three.
North Star metric paired with check metrics
Source
Invest Like the Best with Patrick O'Shaughnessy — Gokul Rajaram - Lessons from Investing in 700 Companies - [Invest Like the Best, EP.456] (January 29, 2026) · at 51:46 ▸
When this applies
when choosing any single metric to rally a team or company around
The takeaway
Any North Star metric must be paired with explicit guardrail ('check') metrics on both customer value and business health, because incentives drive behavior toward whatever is measured alone.
Why it works
A metric optimized in isolation invites gaming (e.g., driving GMV to zero delivery fees), so pairing it with guardrail metrics on customer and company health prevents the team from sacrificing what the metric doesn't capture.
One level up
Single-metric optimization is inherently exploitable; durable goal-setting requires a metric system, not a single number.
Where it stops holding
This stops working when guardrail metrics are too numerous or vague to actually constrain behavior, defeating the purpose of having a simple North Star.
Detail
procedure:
- Choose a North Star metric correlated with both customer value and business growth (not revenue directly)
- Pair it with a customer-health check metric (e.g., retention)
- Pair it with a company-health check metric (e.g., gross margin)
- Monitor both whenever pushing the North Star metric
rationale: E.g., DoorDash's GMV North Star needed a margin guardrail, or a team could drive GMV up by cutting delivery fees to zero and destroying revenue.
In their words
So you basically want a checkmatric around the health of the customer and a checkmatric around the health of the company that are the guardrails around this Nordstra metric.
Judgment as the future-proof skill in agentic AI
Source
Invest Like the Best with Patrick O'Shaughnessy — Gokul Rajaram - Lessons from Investing in 700 Companies - [Invest Like the Best, EP.456] (January 29, 2026) · at 5:33 ▸ · at 8:48 ▸ · at 13:55 ▸
When this applies
when any production process becomes radically cheaper and faster via automation, flooding you with output
The takeaway
As a task's execution becomes commoditized, the scarce and valuable human skill becomes evaluative judgment, not production.
Why it works
When generation is nearly free and capability frontiers shift every two months, the bottleneck moves from creation to deciding what's worth creating and whether the output is correct.
One level up
Invest your human capital where the automation gap persists: deciding what matters, not doing what's now automatable.
Where it stops holding
This stops working when the domain is so novel or low-stakes that even flawed judgment calls carry no real cost, or when outputs are easily and cheaply verifiable by simple automated checks.
Detail
approach: Ask: given infinite production capacity, what is actually worth producing, and is the output reasonable across use cases?
signals:
- non-deterministic software outputs vary with small input changes
- PMs now own evals instead of specs
- teams struggling to evaluate code/design/PM candidates
rationale: Judgment (what to build, whether output is good) is the one thing that doesn't get commoditized as capability frontiers expand.
In their words
Yes, the one thing I think that's going to be truly future-proof is judgment.
Incumbent system-of-record cuts off API access (Slack/Glean)
Source
Invest Like the Best with Patrick O'Shaughnessy — Gokul Rajaram - Lessons from Investing in 700 Companies - [Invest Like the Best, EP.456] (January 29, 2026) · at 20:01 ▸
When this applies
when a smaller player builds dependent value on top of a larger platform's data or infrastructure
The takeaway
Building critical value on another company's controlled chokepoint is fragile because the chokepoint owner can revoke or monetize access once threatened.
Why it works
Once the platform owner sees the agent company capturing value that threatens its core product, it can cut or price the dependency to protect itself.
One level up
Dependence on someone else's distribution or data access is a liability that converts into existential risk exactly when you succeed.
Where it stops holding
This stops working as a warning when the platform owner is indifferent to or benefits from the dependent company's growth (e.g., true complements rather than substitutes).
Detail
situation: Glean built an AI agent layer on top of Slack's API to search/aggregate Slack data for enterprises.
decision: Salesforce (Slack's owner) cut off Glean's API access to prevent it from siphoning value from Slack's core product.
outcome: Agent companies across verticals are now forced to build their own systems of record rather than live on top of incumbents' APIs.
lesson: Relying on an incumbent's API as your core value layer is a structurally weak position once you threaten their business.
In their words
Slack has done it most publicly. Slack is owned by Salesforce. They cut off access to Glean, where Glean can no longer access Slack data.