Gokul Rajaram - Lessons from Investing in 700 Companies - [Invest Like the Best, EP.456] — Summary & Takeaways · Colossus (Invest Like the Best / Business Breakdowns)

Theme 1: Fundamental Transformation of Product Building in the AI Era

Gokul Rajaram argues that product development is shifting from "deterministic software" to "non-deterministic software," fundamentally reshaping team structures and workflows.

Historical Context: 10–15 years ago, product managers (PMs) defined requirements, designers designed, and engineers built, with clearly delineated roles. However, over the past few months (late 2025 to early 2026), the maturation of long-running agents has rendered this model obsolete. Rajaram himself experienced a turning point: six months ago, building a video transcription tool with Claude Code repeatedly failed, yet just two weeks ago, he succeeded in only one hour using prompts, because agents are now resilient to failure.
Mechanism Breakdown: AI capabilities advance every two months, making top-down rigid planning no longer feasible. Teams are shifting to a "bottom-up" model: PMs focus solely on the highest-level "why" (customer needs), while engineers, researchers, designers, and PMs collaborate at the code level. Specific changes include:
PMs now use Codex or Claude Code to submit code directly to production repositories (engineers still review, but AI will soon handle reviews autonomously).
Interviews now include a "prototyping" round, forcing PMs to get hands-on.
The roles of PM and designer are converging: design systems are already established, and AI can handle most work based on design language, leading companies to reduce designer headcount and increase engineer headcount. The PM-to-engineer ratio has shifted from 1:3 or 1:10 to 1:20.
Non-deterministic software requires PMs to lead "evals"—judging whether AI outputs are reasonable, sometimes using AI to evaluate AI results.
Implications: Rajaram cites his friend Zach's framework—"the Industrial Revolution targeted goods, the AI Revolution targets services." The pace of product building will irreversibly accelerate, and the core value of PMs will shift from "planning" to "judgment and evaluation."

Theme 2: Defensiveness in the AI Era and the Moat of "System of Record"

Rajaram argues that in the AI era, the defensiveness of software companies depends on the time value of data and the depth of business integration, making Zendesk and Slack more vulnerable than Salesforce or NetSuite.

Mechanism Breakdown: Rajaram categorizes software companies into two types:
Utility-based pricing companies: Such as Zendesk, which charges per seat, with each seat corresponding to a customer service agent handling a certain number of tickets. These companies are the most vulnerable, as AI agents can "sit alongside Zendesk" and gradually replace human seats—users do not need to replace everything at once, but can simply reduce from 50 seats to 20 and supplement with 30 AI agents. This "two-way door decision" makes erosion slow but inevitable.
Data accumulation companies: Such as NetSuite (an ERP system), whose data has "timelessness." Businesses run on NetSuite, and replacing it is "career limiting" due to the extremely high risk of data migration. These companies have more time to build AI agents on their own data and bundle them for sale.
Data Chain: Rajaram points out that Slack has a very short data half-life, whereas customer records, payment data, and loyalty data in ERP systems hold long-term value. The case of Square supports this: even with cheaper products, small merchants are reluctant to migrate because data migration scripts can take months or even years.
Competitive Landscape: In 2025, system-of-record companies began to fight back against AI-native companies. Slack (owned by Salesforce) publicly cut off Glean's access to its API, preventing Glean from building on top of it and extracting value. Other companies adopt three strategies: blocking APIs, bundling their own agents for free, or charging for API calls (e.g., $2 per call). Rajaram concludes that AI-native companies cannot survive long-term solely as a "system of action" and must eventually build their own system of record.
Implications: Rajaram believes that the public software market currently does not distinguish between these two types of companies, leading to an overall undervaluation. Utility-based companies (such as Zendesk) may need to go private to complete a business model transformation (shifting from per-seat pricing to outcome-based pricing, e.g., charging $0.20–$0.50 per resolved ticket).

Theme 3: Sources of Stickiness in the AI Era — Five Defense Mechanisms

Rajaram argues that in an era where "code is infinitely cheap," stickiness comes from five scarce resources, not from the software itself.

Mechanism Breakdown: Rajaram lists five sources of stickiness with specific examples:

1. Network Effects: For example, DoorDash, which consists of a three-sided network of restaurants, riders, and consumers, cannot be replicated through "vibe coding."

2. Money Flowing Through the Platform: For example, Toast (payments + POS) and Mercury (corporate banking), where the flow of funds introduces regulatory and switching costs.

3. Hardware Lock-in: For example, Toast provides hardware for free, and replacement requires physical removal and payment of fees.

4. Unique Assets: For example, Sierra possesses Bret Taylor (Chairman of OpenAI, top-tier salesperson), whose relationship network cannot be replicated.

5. Data Moat from System Records: For example, NetSuite and Salesforce, where data migration requires two years of engineering effort.

Inference: Rajaram cites Harrison Helmer's "Seven Powers" framework, emphasizing that AI-native companies must embed these defense mechanisms into their business models from day one; otherwise, the software's half-life will be extremely short.

Theme 4: Three Profitable Models for Advertising Businesses and the Platform Threat

Rajaram concludes that there are only three successful paths for advertising businesses, and the shift in consumer behavior toward agent-based interfaces poses the greatest threat to existing platforms.

Mechanism Breakdown: The three profitable models:

1. Owning first-party products and users: Examples include Google Search (high intent), Facebook (identity data), and ChatGPT (combining intent and identity data, with multi-turn natural language queries—a dream for advertisers). Rajaram emphasizes that ChatGPT possesses both Google's intent data and Facebook's identity data, and its queries are continuous rather than one-off.

2. Driving specific outcomes: For instance, AppLovin (market cap of $100 billion+), which focuses on the single outcome of mobile app installs, controls both buyers and sellers as well as middleware, and nearly dominates mobile app auctions.

3. Becoming the exclusive demand-side platform for large advertisers: For example, The Trade Desk, where companies like P&G entrust all their non-Google/Facebook display advertising budgets.

Failure Mode: Attempting to act as an intermediary within the Google/Facebook ecosystem. Rajaram warns that these platforms will learn and replicate your capabilities. The Trade Desk does not access Google/Facebook's first-party inventory, and AppLovin primarily operates on "unwashed web pages."
Implications: Rajaram believes that OpenAI's advertising business will give rise to peripheral companies like "AEO (Answer Engine Optimization)," but these firms cannot build lasting value. The greatest threat to existing platforms (e.g., Uber, DoorDash, Facebook, Google) is the shift in consumer behavior toward agent-based interfaces—users delegate repetitive tasks (such as ride-hailing or food ordering) to AI agents, no longer opening native apps, thereby losing advertising exposure opportunities. Rajaram advises platforms to closely monitor the behavior of early users who connect to ChatGPT accounts: Do they open apps less frequently? If so, platforms need to take measures (e.g., making the agent experience less appealing than the native app).

Theme 5: Leadership and Product Philosophy Learned from Four Legendary CEOs

Rajaram shares his experience working with Larry Page, Sergey Brin, Mark Zuckerberg, and Jack Dorsey, distilling a leadership model of "aligning superpowers with company needs."

Larry Page & Sergey Brin (Google): The core is "building the best technology on earth." Larry’s criticism of AdSense was not about revenue, but rather "you account for less than 1% of global advertising." Before AdSense launched, Sergey scrapped the complex approval system and replaced it with real-time content review—only URLs that reached 100 impressions triggered manual review. Rajaram calls this a "lazy but genius onboarding approach."
Mark Zuckerberg (Facebook): Rajaram describes Zuck as "the greatest mind for growth and user engagement." Through "learning by following," Zuck went from an advertising novice to a source of creative ideas within a year. Custom Audiences—which allows advertisers to upload customer data and find similar users—originated from a conversation between Zuck and Zynga CEO Mark Pincus. Zynga complained about being unable to reach "whale users" (who account for 80% of gaming revenue), and Zuck proposed, "Let them upload their whale data into our system." This feature became the cornerstone of Facebook advertising.
Jack Dorsey (Square): Rajaram ranks Jack’s design philosophy alongside Jony Ive and Steve Jobs. Good design is not about visual appeal, but "usable without a manual." Square’s POS system can be downloaded and used immediately, whereas traditional POS systems require days of training. Jack refers to PMs as "product editors," whose core role is to cut rather than add. Square reduced risk from the "person level" to the "transaction level"—accepting 90%+ of merchants but assessing risk on a per-transaction basis using machine learning.
Tony Xu (DoorDash): In interviews, candidates are given $10-$20 and asked to acquire 1,000 consumer customers. No one succeeds, but the goal is to observe how many different methods the candidate tries (e.g., handing out flyers at the gym). Rajaram calls this "an excellent way to screen for action-oriented individuals."

Theme 6: Leadership Communication and Career Advice

Rajaram proposes a standardized format for the "weekly CEO email" and warns that "job-hoppers" are the biggest red flag in hiring.

Communication Format: Rajaram recommends that the weekly CEO email consist of three sections:

1. Top of Mind (most important, accounting for 60-70% of the time): What the CEO is most focused on currently, which may involve products, business, or the team.

2. Performance Update: Key performance indicators of the company.

3. Miscellaneous: Employee recognition, customer quotes, and event announcements.

Rajaram emphasizes that "the more candid, the better," as candor encourages the team to contribute ideas. He cites Eric Schmidt's communication technique: using pure visuals for strategic presentations, because "people don't remember words, they remember feelings."

Career Advice: Rajaram believes that "each job should be held for at least 3-4 years to make a real impact." Switching jobs every 12-18 months is an "instant red flag" in hiring. He predicts that the most scarce skill in the AI era will be "becoming a functional expert and knowing how to orchestrate an army of AI agents to execute that function." A manager's "span of control" should be at least 10 people; otherwise, they should return to an individual contributor role.
Evaluation Method: Rajaram advocates for the "work project" interview. The best PM candidates will reject the given premise, proactively talk to 10 customers, and then say, "We shouldn't build this product." He looks for "agency" and the "ability to question assumptions."

Position Moves

Ticker	Analyst View	Key Data
Zendesk	Risk Warning (Most vulnerable to AI disruption)	Priced per seat, each seat corresponds to one customer service agent handling tickets
Slack	Risk Warning (Short data half-life)	Has cut off Glean's API access
Salesforce	Relatively Safe (System of record)	Customer record data is time-permanent
NetSuite	Relatively Safe (ERP system)	Replacement is "career suicide," data migration risk is extremely high
AppLovin	Bullish (Ad model 2)	Market cap $100B+, focused on mobile app installs
The Trade Desk	Neutral (Ad model 3)	Does not access Google/Facebook first-party inventory
ChatGPT (OpenAI)	Bullish (Ad potential)	Combines intent + identity data, multi-turn natural language queries
DoorDash	Bullish (Network effects)	Three-sided network of restaurants, drivers, and consumers
Toast	Bullish (Hardware + payment bundling)	Provides hardware for free, funds flow through the platform
Mercury	Bullish (Funds flow through platform)	Business banking, regulatory and switching costs
Sierra	Bullish (Unique asset)	Owns Bret Taylor (OpenAI Chairman)
Figma	Bullish (Bottom-up adoption)	Defeated Sketch through a bottom-up movement
Cursor	Bullish (Bottom-up adoption)	99.9% of companies adopt via engineers bottom-up
Palantir	Bullish (Outcome-based sales)	Promises "free if not solved in 6 months, high price if solved"
Square	Neutral (Historical case)	North Star metric is GPV (payment processing volume)
Facebook	Neutral (Historical case)	North Star metric shifted from MAU to DAU
Google	Neutral (Historical case)	Sergey cut the approval system when AdSense launched
Zynga	Neutral (Historical case)	80% of revenue came from "whale users," giving rise to Custom Audiences
Glean	Risk Warning (API cut off)	Slack cut off its data access
Clio / Filevine	Neutral (System of record)	Legal industry system of record
Epic	Neutral (System of record)	Healthcare industry system of record
Atlassian (Jira)	Neutral (System of record)	Product development data

Judgments Worth Remembering

1. Rajaram believes that in the AI era, the only future-proof asset is human judgment. When code generation becomes infinitely cheap, the ability to decide "what to do" and "what is worth doing" becomes a scarce resource. Supporting evidence: AI slop (AI-generated junk code) is the biggest concern for every product team.

2. Rajaram argues that there are only three viable business models in advertising, and only these three. Model 1: Owning first-party products and users (Google, Facebook, ChatGPT); Model 2: Driving specific outcomes (AppLovin's mobile app installs); Model 3: Becoming the exclusive demand-side platform for large advertisers (The Trade Desk). Attempting to act as an intermediary within the Google/Facebook ecosystem is doomed to fail.

3. Rajaram judges that Zendesk and Slack are more vulnerable to AI disruption than Salesforce or NetSuite. Reason: Zendesk charges per seat, and AI agents can gradually replace human seats (a "two-way door decision"); Slack's data has a short half-life. In contrast, ERP system data is timeless, and replacing it would be "career suicide."

4. Rajaram warns that the shift in consumer behavior toward agent interfaces is the biggest threat to existing advertising platforms. If users delegate repetitive tasks (ride-hailing, food ordering) to AI agents and no longer open native apps, platforms will lose advertising exposure opportunities. He advises platforms to closely monitor behavioral changes among users who connect their ChatGPT accounts.

5. Rajaram concludes that stickiness in the AI era comes from five scarce resources: network effects (DoorDash), capital flowing through the platform (Toast, Mercury), hardware lock-in (Toast), unique assets (Sierra's Bret Taylor), and data moats from systems of record (NetSuite, Salesforce). Without these, software has an extremely short half-life.

6. Rajaram suggests that the roles of PMs and designers are converging, with the PM-to-engineer ratio shifting from 1:3 to 1:20. Design systems are already established, and AI can execute tasks based on design language. The core work of PMs is shifting toward "evals"—judging the reasonableness of non-deterministic AI outputs, sometimes using AI to evaluate AI.

7. Rajaram believes that every job requires at least 3-4 years to generate real impact. Job-hopping every 12-18 months is an "instant red flag" in hiring. He predicts that the most scarce skill in the AI era will be "becoming a functional expert and knowing how to orchestrate an army of AI agents to execute that function."

8. Rajaram argues that systems of record companies (e.g., Salesforce, NetSuite) are fighting back against AI-native companies. In 2025, Slack cut off Glean's API access. Other companies are adopting three strategies: blocking APIs, bundling their own agents for free, or charging for API calls (e.g., $2 per call). AI-native companies must ultimately build their own systems of record.

Gokul Rajaram - Lessons from Investing in 700 Companies - [Invest Like the Best, EP.456]

In plain words

At a Glance