Jesse Zhang - Building Decagon - [Invest Like the Best, EP.443]

Here is the translated report in natural, professional English.

At a Glance

Jesse Zhang, co-founder and CEO of Decagon, shares a systematic methodology for building an AI customer service company from zero to one. The core theme revolves around how to identify and capture the clearest enterprise AI application scenario through an extremely pragmatic discovery process anchored by the customer's willingness to pay. The most significant judgment in the entire piece is: Jesse Zhang believes that AI agents will first penetrate from the two ends of the labor cost spectrum—the highest-paid engineers and the lowest-paid customer service agents—rather than the middle ground.

Topic Sections

1. Systematically Finding PMF: Replace "What do you think?" with "How much would you pay for this?"

Jesse Zhang argues that the best way to find product-market fit is not through intuition, but via a structured customer interview process that directly quantifies potential customers' willingness to pay. He criticizes the common founder practice of "asking customers what they want," believing this usually fails to yield genuine signals.

• Mechanism Breakdown: Zhang's process begins with high-level exploration with potential customers (e.g., COOs or VPs) to understand their pain points. He then forms a product hypothesis on the spot and asks: "If an AI agent could do X, Y, and Z, would that be helpful?" Upon receiving an affirmative answer, he immediately follows up with: "How much would you be willing to pay for this? " This critical question forces the customer to shift from "this is cool" to a rational calculation of "what is this worth," thereby quantifying the ROI.
• Data Chain: After exploring multiple ideas (data analysis, security, pre-sales, etc.), the willingness to pay expressed by potential customers in the customer service domain (low to mid six figures) was an order of magnitude higher than for all other ideas. This gave them the core confidence to ultimately choose the customer service track.
• Deduction and Falsification: Zhang emphasizes that the signal value of this process far exceeds seeking advice from "smart, old-school founders." He believes that in any technological wave, truly good ideas are extremely rare and non-obvious, and can only be discovered through this kind of deep, quantitative customer conversation.

2. Customer Service vs. Coding: AI Agents are "Eating" the Two Ends of the Labor Cost Spectrum

Jesse Zhang proposes a "labor cost spectrum" framework to explain why customer service and coding are currently the two most successful enterprise AI application scenarios. He argues that the essence of AI agents is to replace human labor, and they will begin by penetrating the highest and lowest cost ends of the spectrum.

• Comparative Analysis:

Feature	Customer Service (Low-End)	Coding (High-End)
Labor Cost	Low (often outsourced, Tier 1/2)	Extremely High (Engineers)
AI Role	Replacement: Directly reduces labor costs	Augmentation: Increases engineer leverage; companies won't lay off staff as a result
ROI Proof	Easy to quantify: Saved operational costs	Easy to quantify: Engineers self-report productivity gains (e.g., 50%)
Deployment Risk	Low: Has a natural "escalate to human" path	Low: Used by engineers themselves, bottom-up adoption

• Unique Judgment: Zhang points out that the success of the customer service scenario is not accidental. It has two underestimated advantages: 1) ROI is extremely easy to quantify ("cut costs by 60%"); 2) Deployment risk is extremely low, because when the AI agent fails, it can seamlessly escalate to a human agent, providing a valuable "safety net" for large enterprises.

3. The Key to Enterprise AI Deployment: Aligning on "What is Good" and Building a "Hybrid" System

Jesse Zhang believes that the biggest challenge in successfully deploying AI agents within enterprises is not technology, but aligning with the customer on the standard of "what constitutes good performance." This requires a sophisticated evaluation and testing system.

• Mechanism Breakdown: Decagon's approach is to build a simulation suite containing 10,000 test cases for each customer, running continuously to quantify the AI's performance. This forces different internal departments within the customer's organization (e.g., CX head, product head) to agree on the "correct answer," thereby translating the vague concept of "good" into a measurable score.
• System Design: Zhang emphasizes that enterprise AI needs to balance "flexibility" and "rigor." He describes a "hybrid" system: for scenarios requiring high accuracy (e.g., financial compliance), the system enforces a fixed process; for scenarios requiring a personalized experience (e.g., checking an account balance), the model is allowed more freedom. This design is key to managing the non-deterministic nature of LLMs.
• Deduction and Falsification: Regarding voice AI, Zhang points out that while Voice-to-Voice models have great potential for naturalness, their hallucination rate is approximately 8 times higher than text models, making them difficult to use directly in enterprise applications. Therefore, the current best practice is a hybrid architecture of "speech-to-text, then text-to-speech," sacrificing some naturalness for accuracy.

4. Competition, Talent, and Capital: The "White-Hot" Center of AI Entrepreneurship

Jesse Zhang paints a picture of the intensely competitive state of the current AI startup ecosystem, particularly regarding talent and capital. He believes this attracts a generation of founders with a highly competitive spirit and a background in math/programming competitions.

• Capital Side: Zhang admits that "raising money is too easy" in the AI field, and there is "a bit of frenzy." He observes that investors, eager to invest in top companies, are willing to provide help even before investing, offering founders an excellent window to test the future value of an investor. He advises founders to use this stage to evaluate an investor's "cognitive ability" and "willingness to help."
• Talent Side: Hiring is "a team battle" that requires the entire company to "surround" the target candidate. To compete for top talent (e.g., graduates from Harvard, MIT, Stanford), companies need to deeply understand the candidate's personal and family needs and tailor a role for them. Opening a New York office was also a move to tap into a new talent pool.
• Culture: Zhang describes the company culture as "hyper-competitive" and "action-oriented," with a motto on the wall reading, "There is no challenge that cannot be overcome, no enemy that cannot be defeated." He believes this culture is attractive in the AI era because it appeals to those who see entrepreneurship as the pinnacle of their career, are willing to work intensely, and want to build lifelong relationships.

Position Moves

Position	Guest's Stance	Key Data
Decagon	Bullish (Own Company)	Customer willingness to pay is low to mid six figures; customer escalation rate dropped from "1/3 request human" to "1/20"; voice interaction accounts for up to 90-95% for some clients.
Oura Ring	Neutral (Customer Case)	After using Decagon, the proportion of customers requesting a human agent dropped from 1/3 to 1/20.
Chime	Bullish (Excellent Customer)	Described as a "case study-level" customer with a data-driven team and excellent culture.
Cognition	Bullish (Friend's Company)	Would include in their 5-company investment portfolio.
Cursor	Bullish (Friend's Company)	Would include in their 5-company investment portfolio.
Etched	Bullish (Hardware Layer)	As a bet on the hardware layer, would include in their 5-company investment portfolio.
Pika	Bullish (Friend's Company)	Would include in their 5-company investment portfolio.
CHI (Josh's company)	Bullish (Friend's Company)	Building a foundation model for healthcare; would include in their 5-company investment portfolio.
Physical (Lockheed's company)	Bullish	As a bet on physical world AI, would include in their 5-company investment portfolio.
Google	Bullish	Believes it has a strong consumer base and team, giving it an advantage in the AI era.
Anthropic	Neutral	Believes it is weaker on the consumer side compared to ChatGPT and may need consumer data in the long run.

Judgments Worth Remembering

1. Systematic PMF Discovery Method (Jesse Zhang): Don't ask customers "What do you think?", ask "How much would you be willing to pay for this? " This quantitative question forces the customer from emotional judgment to rational calculation, yielding the most genuine signal.

2. The "Labor Cost Spectrum" Theory for AI Agents (Jesse Zhang): AI will first penetrate from the two ends of the labor cost spectrum—the highest-paid engineers (augmentation) and the lowest-paid customer service agents (replacement)—rather than the middle ground.

3. The "Low-Risk" Advantage of the Customer Service Scenario (Jesse Zhang): Customer service is an ideal scenario for enterprise AI deployment because it has a natural "escalate to human" path. Even if the AI makes a mistake, it can seamlessly transfer the call, greatly reducing the deployment risk and psychological barrier for the enterprise.

4. The "8x Hallucination Rate" of Voice AI (Jesse Zhang): While Voice-to-Voice models have great potential for naturalness, their hallucination rate is approximately 8 times higher than text models, making them difficult to use directly in enterprise applications. The current best practice is a hybrid architecture.

5. "Test Suite" is the Bedrock of AI Deployment (Jesse Zhang): Before deploying an AI agent in an enterprise, an evaluation system containing 10,000 test cases must be built collaboratively with the customer to align on the standard of "what is good performance." This is a prerequisite for successful deployment.

6. Pre-Investment Help is a "Touchstone" (Jesse Zhang): Founders should use the phase when investors are "most eager to help" before investing to test them. If they are unwilling to help at this stage, they are even less likely to be useful after the investment. This is an excellent window to assess an investor's value.

7. The Threshold for the "Forward-Deployed Engineer" Model (Jesse Zhang): For the hyped "forward-deployed engineer" model to be economical and scalable, the customer contract value needs to be at least $1 million; otherwise, it cannot scale due to human resource bottlenecks.

8. The "Moat" in the AI Application Layer is the Data Flywheel (Jesse Zhang): The longer the collaboration with a customer, the stronger the AI agent becomes by analyzing vast amounts of conversation data (e.g., identifying 2% of anomalous topics) and continuously self-optimizing, ultimately forming a data moat that is difficult for competitors to replicate.