All podcasts / Invest Like the Best / Summary

Jesse Zhang - Building Decagon (EP.443)

2025-10-06 - 80 min - source - Read full transcript

Patrick O'Shaughnessy (host)Jesse Zhang

Key insights

Forcing prospects to name a dollar figure is the real discovery signal, not whether they say an idea sounds interesting.

Jesse's process goes past 'would this be useful' to 'exactly how much would you pay, and who has to approve it.' People will politely agree an idea sounds cool, but making them state a number and an approval chain reveals genuine budget and urgency versus polite enthusiasm. Most ideas, when pushed this way, top out at a few hundred dollars a month; the winning idea for Decagon (customer service) surfaced answers in the low-to-mid six figures.

customer-discovery-and-pricing

Talking to one senior stakeholder about many use cases, or one mid-level owner about one use case, yields cleaner signal than generic 'what do you want' conversations.

Jesse tailors interview scope to seniority: a COO can speak credibly to several potential use cases at once, while a VP of a specific function should be pressed deep on the one area they own. Tallying willingness-to-pay across many such conversations, rather than trusting any single anecdote, is what let Decagon converge on customer service as the highest-conviction bet.

customer-discovery-and-pricing

AI agents are winning first at the two ends of the labor-cost spectrum: the cheapest outsourced labor gets replaced, the most expensive engineers get augmented.

Customer service (often BPO-staffed, high turnover, low pay) is easy to fully automate because the work itself is replaceable. Engineers are the highest-paid and most sophisticated users of AI, so coding agents multiply their output rather than eliminate their jobs. The middle of the spectrum, Jesse argues, is where AI adoption is slower and messier.

ai-labor-substitution

Customer service was the first enterprise AI use case to scale because ROI is trivially quantifiable and the product already has a built-in escalation path.

Companies already track conversation volume and current bot resolution rates, so the savings math ('resolve 60% instead of 20%') is simple. Just as important: because human-agent escalation infrastructure (call centers, telephony) already exists, a failure mode is cheap and low-risk, letting enterprises go live on a small slice of traffic instead of betting everything up front.

enterprise-ai-deployment-risk

Once an enterprise trusts the metrics, rollout from a small pilot to full deployment can happen within a single week.

Enterprises typically start an AI agent on 5% of volume, then watch resolution rate, CSAT, and human-reviewed accuracy scores in near real time. If those check out, Jesse says there's 'really no reason' not to expand to 100% almost immediately, because the safety net (escalation to a human) is already proven to catch failures.

enterprise-ai-deployment-risk

The biggest technical gap in enterprise AI is voice-to-voice: it is roughly 8x more prone to hallucination than text because it generates far more tokens per utterance.

Voice-to-voice models carry latency and naturalness advantages over text-to-speech pipelines (they capture tone and cadence, respond faster), but the higher token count per response multiplies error surface. Most production systems today route voice through a text intermediate step so outputs can be checked before being spoken.

enterprise-ai-deployment-risk

Data collected from AI-customer interactions is a durable moat because agents keep improving from lived conversations, not just static training.

Jesse frames the compounding advantage as the agent 'continuously getting better' from the specific, unstructured signal in a year of conversations with a given client - flagging under-covered topics, drafting fixes, and refining guardrails - which is a different and harder-to-replicate asset than merely fine-tuning on the same dataset once.

ai-labor-substitution

Negative gross margins are fine pre-enterprise-scale because compute costs are falling exponentially, but AI companies should still avoid deeply negative unit economics on long-term enterprise contracts.

Jesse argues most people underestimate how fast AI unit costs will keep falling, so optimizing hard for margin today trades away growth for savings that will arrive anyway. The exception is enterprise deals, which are multi-year commitments where a company can't simply assume margins will fix themselves before the contract renews.

ai-economics-and-competition

Being 'just a wrapper' is not inherently a weak business - the software layer built around a model (guardrails, observability, QA, integrations) is what's actually defensible.

Jesse pushes back on the pejorative 'ChatGPT wrapper' framing: an agent is not just a model call, it requires observability into conversations, alerting, unit tests for conversation quality, and enterprise integration work that has nothing to do with AI per se. That non-AI engineering surface is large and is what protects a company from being obsoleted by a model update.

ai-economics-and-competition

The forward-deployed-engineer model, borrowed from Palantir, only makes sense above roughly $1M in customer revenue - smaller deals can't support it.

Jesse thinks the FDE pattern is currently over-hyped among AI startups: Palantir's version staffs engineers full-time against $10-25M deals, but many startups apply the same model to $50K accounts, which doesn't scale because good engineers are scarce and headcount becomes the bottleneck rather than dollars.

ai-economics-and-competition

Fundraising demand for hot AI companies feels frothy: rounds get preempted almost immediately, based more on momentum than first-principles valuation.

Jesse notes that after every Decagon raise the company was nearly instantly preempted for the next round, which he says can't be fully rational since a prior valuation shouldn't heavily determine a new investment decision. He treats an investor's pre-investment helpfulness (intros, market insight, hustle before they've committed) as the best predictor of how helpful they'll actually be as a shareholder afterward.

ai-economics-and-competition

Recruiting top talent is treated as a whole-team sales process, including learning about a candidate's family and life goals, not just a hiring-manager task.

Jesse describes talent acquisition at Decagon as requiring the 'whole team to swarm' around a sought-after hire, mirroring enterprise sales tactics - understanding what a person actually wants out of their career and shaping a role around it, rather than relying on a standard pitch.

startup-culture-and-talent

A culture of intensity and competitiveness is treated as a deliberate hiring filter, not a byproduct.

Decagon signals a high-intensity, aggressive culture (a wall quote about 'no enemy that can't be defeated') explicitly so candidates self-select in or out before joining. Jesse links this directly to the math/coding-Olympiad backgrounds common among his generation of founders, arguing that objective, competitive results as a kid built the same problem-solving instincts now applied to company-building.

startup-culture-and-talent

Companies

Decagon - Jesse's company - an AI customer service agent platform that auto-resolves issues across chat and voice.
Low Key - Jesse's first company, high-performance video capture software for gamers, which he started right after graduating and eventually sold in 2021.
Ramp - Episode sponsor; also cited as an example of a hot, competitive market where rational founders pile in once the opportunity is visible.
Rippling - Matt McGinnis, an ops leader at Rippling, was one of the early customer-discovery contacts Jesse credits with validating Decagon's use case.
Oura - Case study customer (the wearable ring company); before Decagon, 1 in 3 callers hung up without engaging a bot, after it was 1 in 20.
Cognition - Coding-agent company; Scott Wu (Cognition) is a friend Jesse cites as a fellow high-intensity founder from the math-Olympiad world, and Cognition is one of Jesse's picks for a hypothetical five-company AI portfolio.
Cursor - Named alongside Cognition as one of the clearest AI-coding businesses and a pick for Jesse's hypothetical portfolio.
Sierra - Named as a direct comparable/competitor in the AI customer-service category alongside Decagon.
Etched - AI hardware/chip startup Jesse names as his hypothetical portfolio bet on the hardware layer.
Pica - Video-model startup built by a friend of Jesse's; named as a hypothetical portfolio pick.
Chai - Foundation-model startup (healthcare-focused, built by Jesse's friend Josh) named as a hypothetical portfolio pick.
Physical Intelligence - Robotics/foundation-model company Jesse names as a hypothetical portfolio pick (referred to informally as a friend's 'physical' company).
Palantir - Decagon co-founder Ashton's prior employer; source of the 'forward-deployed engineer' model Jesse discusses and pushes back on for early-stage startups.
Google - Jesse says he is 'very bullish' on Google among the Mag 7 because consumer reach feeds the data flywheel AI needs.
OpenAI - Discussed as both a model provider (commoditized, easy to swap) and a competitor moving into applications (e.g. coding tools).
Anthropic - Discussed as a model provider; Jesse notes it has weaker consumer distribution than ChatGPT, which he sees as a long-term disadvantage.

Techniques and frameworks

Willingness-to-pay discovery interviews - Jesse's systematic idea-validation method: ask prospective customers exactly how much they would pay, who has to approve it, and how they'd justify the ROI to leadership, before building anything.
Staged AI rollout (5% to 100%) - Enterprises test an AI agent on a small slice of volume with an escalation-to-human safety net, then expand fast (often within a week) once resolution rate, CSAT, and accuracy metrics check out.
Tiered AI capability ladder - Decagon's framework for customer-service maturity: tier 1 static Q&A, tier 2 Q&A grounded in real-time account data, tier 3 agentic multi-step actions (e.g. replacing a lost credit card).
Eval/simulation suites (AOPs) - Decagon builds thousands of test cases run repeatedly per customer, and converts client SOPs into 'agent operating procedures' (AOPs) so 'what good looks like' is explicit before going live.
Labor-cost spectrum framing for AI use cases - Jesse's mental model for where AI agents win first: the cheapest labor (BPO-style support) gets replaced, the most expensive labor (engineers) gets augmented, because both ends have the clearest ROI.
Forward-deployed engineer model - Borrowed from Palantir; Jesse argues it only makes sense above roughly $1M in customer revenue, and that many startups over-index on it for deals far too small to support dedicated engineering staff.

Summary

Jesse Zhang, co-founder and CEO of AI customer-service company Decagon, walks Patrick O'Shaughnessy through how the company found its wedge, why customer service and coding have become the two clearest enterprise AI use cases, and what he's learned building against a swarm of well-funded competitors. The conversation opens on culture: the aggressive wall-quote ethos at Decagon's offices, and Jesse's theory that a generation of math- and coding-Olympiad kids (himself included, alongside friends like Scott Wu of Cognition) has translated competitive, objectively-scored childhoods directly into founder instincts. He contrasts the emotional weight of his tougher first company, Low Key (video-capture software for gamers, sold in 2021), with the more systematic, less anxious process he and co-founder Ashton used to find Decagon's direction.

The core of the episode is Jesse's discovery methodology: rather than asking prospects what they want, he pushes them to name an exact dollar figure and approval chain for a hypothetical solution, treating it as a sales-qualification exercise in founder's clothing. This process, run across many prospective customers testing several unrelated ideas at once, surfaced customer service as the outlier - willingness-to-pay tallies there dwarfed every other idea by an order of magnitude. He explains why customer service scaled so fast at the enterprise level: ROI is trivial to quantify against existing conversation-volume data, and the human escalation path that already exists (call centers, telephony) makes a staged rollout low-risk, letting companies expand from a 5% pilot to full deployment within a single week once resolution-rate, CSAT, and accuracy metrics hold up.

Jesse frames AI-agent adoption as eating the labor-cost spectrum from both ends: cheap, replaceable support labor gets substituted outright, while expensive, sophisticated engineering labor gets augmented rather than replaced, because engineers are best positioned to leverage the tools. He details the technical frontier still unsolved - voice-to-voice models remain roughly 8x more hallucination-prone than text because of the much higher token count per spoken response - and the tiered complexity ladder Decagon's agents climb, from static Q&A to real-time account lookups to fully agentic multi-step actions like replacing a lost card.

On business model and competition, Jesse pushes back on the "ChatGPT wrapper" critique, arguing the non-AI engineering work (observability, QA, guardrails, integrations) is what actually defends a company from being obsoleted by a model update, and predicts model labs will keep moving up the stack into applications because API margins are structurally thin and easy to arbitrage. He is comfortable running lean margins pre-scale, betting that compute costs will keep falling exponentially, but draws a hard line against negative margins on long-term enterprise contracts. He also cautions against over-applying Palantir's "forward-deployed engineer" model to deals under roughly $1M, calling it a current overreach among AI startups.

The back half covers fundraising (rounds get preempted almost immediately, which Jesse reads as slightly irrational momentum-chasing, and he treats pre-investment investor helpfulness as the best signal of post-investment value) and recruiting (treated as a whole-team, family-inclusive sales process). Jesse closes, per the show's traditional final question, on his parents' unusually disciplined but non-overbearing approach to raising him and his sister, which he credits as the root of his current drive.

Notable Quotes

"If you really go deep there, it's almost like you're basically asking classic sales qualification questions, but in founder form. And because you're a founder, it feels a lot less salesy for you to go deeper." - Jesse Zhang

"AI use cases will start eating the spectrum for both ends... On one end, engineers have the sophistication to really leverage it well... On the other end, it is more of the replacement sense." - Jesse Zhang

"Generally people say wrapper in a derogatory way... but if you have enough software built around the models, that's where you can actually almost capture the most value." - Jesse Zhang

"It just seems way too easy to raise money right now... If you're just thinking of first principles, the previous valuation should not be a super big factor in a new investment. It should be how well the business is doing." - Jesse Zhang

"My parents just pounded a sort of lazy, wanted-to-play-around kid into someone that was just very, very driven. Then, to their credit, they just completely laid off." - Jesse Zhang