What this covers

This is a conversation between Patrick O'Shaughnessy and Jesse, co-founder and CEO of Decagon, an AI customer service company. It starts with how Jesse found the idea. He describes systematizing ideation through Willingness-to-Pay Customer Discovery: asking prospects what they would pay, whether a boss must approve, and how they would calculate ROI. He says most ideas die at that step. From there it covers how Decagon's agents work, including the move from static answers to multi-step actions, and the idea that a language model replaces the old decision tree.

The middle of the conversation widens to operations and deployment. Jesse says Escalation-Path Risk Containment makes customer service easy to take live, since a company can start with 5% of users. He argues that agreeing on what good looks like is often the slowest part of a rollout, and he describes using a model to read every conversation instead of a sampled few. Other threads include voice, brand personality, a long-run "digital concierge" for each company, and a critique of the forward-deployed engineer model borrowed from Palantir. He also contrasts augmenting engineers with replacing outsourced support labor, and he calls the job-loss story overblown given high turnover in that work.

The most contested material concerns economics and competition. Jesse says application companies can tolerate negative gross margin early because inference costs will fall faster than people expect. He argues that Value Capture at the Final Step favors the application layer, and that an agent improved on a client's data over a year forms a moat. He also questions self-reported productivity numbers used to justify coding agents. The episode closes on investor selection, AI talent competition, and when fine-tuning or proprietary models make sense.

Sharpest takeaway

Decagon's founder argues that customer service and coding are the clearest enterprise AI use cases because they replace or augment labor with measurable ROI and built-in safety valves, and that durable value sits in the application layer built around models rather than in model margins or raw model access.

  • Customer service has easily quantified ROI (resolution rates rising from 15-20% to 50-80%) and a natural escalation path to humans that lowers enterprise deployment risk
  • Founders should find real willingness to pay by asking customers how much they would pay and probing the approval chain, not by relying on stated interest
  • Thin model wrappers fail, but agents built with guardrails, evaluation, observability, and integrations become hard to displace when models change

The claims · ranked86 claims · weighted by value

This asset isn't compiled yet

You're seeing its claims, ranked. Compile it to build the argument threads, weight them, and check each claim against your library — the full view.

0.74

People underestimate how fast AI is improving and how cheaply it will run, because humans are poor at conceptualizing exponential change, so early-stage application companies can accept negative gross margins if they win market share and mind share.

forecasthigh valuecontestednovelty 4/4durability 2/4· Jesse

“things are growing exponentially, improving exponentially, and no one is good at conceptualizing what exponential means”

0.73

In the agentic world, the moat is an agent that has continuously improved from a client's data over a year, making it hard for a competitor to match performance, and this is distinct from simply training a model on the data.

causalhigh valuecontestednovelty 3/4durability 3/4· Jesse

“So when you think about moats in the agentic world, a lot of it is around: if you've been working with a client for a year, has your agent just continuously gotten better by learning from the data?”

0.69

AI replacement of outsourced customer service labor is overblown as a job-loss story because BPO workforces already have very high turnover, so headcount falls naturally, and displaced workers move to the next task AI cannot yet do, such as data labeling.

factualhigh valuecontestednovelty 3/4durability 2/4· Jesse

“I think it's a little bit overblown of AI is replacing jobs and so on. Even the BPOs we talk to, they're not really that concerned because what typically happens anyways is that there's already very high turnover in these BPOs.”

0.65

Coding agent ROI is often justified by polling engineers on how much more productive they feel, but a CEO relaying a self-reported 50% productivity gain to the board is weak evidence, and Patrick's cited study suggests self-reported productivity may diverge from measured productivity.

causalhigh valuecontestednovelty 3/4durability 2/4· Jesse

“Are you confident that that's true? That they can self-report and be accurate?”

0.65

Application-layer AI companies do not need good gross margins early, because model costs will fall exponentially and the priority is market share and mindshare, though enterprise deals should stay healthy because their contracts are longer-term and customer expectations may change.

forecasthigh valuecontestednovelty 3/4durability 2/4· Jesse

“So zero margin doesn't really matter. What really matters is you need to get market share and you need to get mind share of users.”

0.62

Companies losing money on inference are not optimizing margins yet because the effort is better spent on quality and growth; they could likely improve margins in a month of re-architecture, but optimization can wait until costs fall and customers are won.

normativehigh valuecontestednovelty 3/4durability 2/4· Jesse

“What you're really optimizing for right now is just quality and growth. So if you can do that, then the optimization can always come later on.”

0.58

When a company makes it clear at the start of a call that the caller is in a different experience, customers are willing to give the AI a chance; for Oura Ring, the share of customers who repeatedly pressed zero or said agent until reaching a human fell from one in three to one in 20.

factualhigh valuespeaker onlynovelty 4/4durability 2/4· Jesse

“Now it's one in 20 because we just spent a lot of time making the beginning of the process just feel very different and folks are willing to give it a chance.”

0.57

Early on, fine-tuning was widely considered unhelpful because gains were small and models changed so fast that tuned work would be discarded, but as applications mature, narrow tasks like agent routing can use smaller fine-tuned models that improve performance and latency without needing general intelligence.

causalhigh valuespeaker onlynovelty 3/4durability 3/4· Jesse

“You can make that a fine-tuned model. That model doesn't have to be good at math or coding or anything. It's just you just need a smaller model that's just fine-tuned on that.”

0.57

Asking potential customers deep qualification questions, such as exactly how much they would pay, whether a boss must approve, and how the organization would calculate ROI, is basically classic sales qualification in founder form, and it feels less salesy and produces much more signal.

normativehigh valuespeaker onlynovelty 3/4durability 3/4· Jesse

“if you really go deep there, it's almost you're basically asking classic sales qualification questions when in founder form. And because you're a founder, it's just feels a lot less salesy for you to go deeper.”

0.57

Customer service is easy to take live because it has a natural escalation path: the agent hands off to a human if it cannot resolve an issue, and existing call-center infrastructure is already in place, so enterprises can roll out to a small share of users like 5% with limited downside.

causalhigh valuespeaker onlynovelty 3/4durability 3/4· Jesse

“the nice thing with customer service is that you have a escalation path just naturally built into the way the product works. The agent's having the conversation. If for any reason it needs to exit, it'll just escalate to a human.”

0.57

AI agents replace human labor, and adoption spreads along the cost spectrum of human labor from both ends: highly paid engineers, where AI augments them because there is infinite engineering work, and low-paid outsourced customer service, where AI replaces workers.

causalhigh valuespeaker onlynovelty 3/4durability 3/4· Jesse

“The one way to think about it is that AI use cases will start eating the spectrum from both ends.”

0.57

Founders should use the period when investors want to invest but have not yet to test how helpful they will be, because investors are most willing to help then, and their behavior at that stage predicts how helpful they will be after investing.

normativehigh valuespeaker onlynovelty 3/4durability 3/4· Jesse

“during the stage where people want to invest but they haven't yet, that's when they're most willing to be helpful. And so it's actually a great way for you to use that to proxy how helpful they'll be afterwards.”

0.57

AI is behind in unlocking enterprise use cases because of non-determinism and the Waymo effect, where AI must be near perfect despite being objectively better than humans; enterprises should evaluate AI on holistic success rate rather than hunting for individual mistakes.

factualhigh valuespeaker onlynovelty 3/4durability 3/4· Jesse

“instead of evaluating AI in a way where you're just trying to find mistakes, you're evaluating holistically, looking at the success rate.”

0.57

The forward-deployed engineer model, as used at Palantir with $10-25 million deals, works only when customers are massive enough to justify full-time staffing; startups with $50k deals who adopt the model are over-indexing, and a startup must find a middle line between hands-on service and scalable staffing.

causalhigh valuespeaker onlynovelty 3/4durability 3/4· Jesse

“But for you to actually have a FDE model, you need to have massive clients, and most people do not have massive clients.”

0.57

Experienced founders and conventional wisdom said customer service was an obvious use case that incumbents would tack onto, but founders must build their own conviction through direct customer signal because obvious-looking markets can still be underexploited.

normativehigh valuespeaker onlynovelty 3/4durability 3/4· Jesse

“as a founder, you kind of have to build your own conviction and kind of do this process yourself.”

0.57

The biggest learning is that aligning on what good looks like is often the long pole for AI customer service deployments, because enterprise answers, tone, and brand guidelines are nuanced and no single person knows all the right answers, so teams must gather answers from CX leaders and product owners to define the evaluation.

causalhigh valuespeaker onlynovelty 3/4durability 3/4· Jesse

“the biggest learning we had is that oftentimes a long pole in the tent. And again, if by definition, if you can solve this well, it just makes things go a lot easier is aligning on what does good look like.”

0.57

Decagon's only margin principle is not to have negative margins; application-layer companies can capture the most margin because they solve the business problem at the final step of the supply chain, the way a croissant seller at an airport does, and customers judge them on business ROI rather than model costs.

causalhigh valuespeaker onlynovelty 3/4durability 3/4· Jesse

“That last step where you're actually solving someone's problem, that's where you generally can capture the most margin.”

0.54

Rather than relying on a vague sense of what to build, a founder should systematize ideation by ensuring whatever they work on is something customers will really invest money in, and the way to test this is to go deep in customer conversations.

normativehigh valuespeaker onlynovelty 3/4durability 3/4· Jesse

“so I think we just try to systematize the ideation process and it just comes down to, okay, whatever you work on, it has to be something that people will really invest in.”

0.54

Asking a customer how much they would pay forces them to think in concrete terms, which produces an order-of-magnitude estimate of scale; for example a customer with five people doing a task may imply about 20k a year in spend, and most ideas die at this step because there are few good ideas.

causalhigh valuespeaker onlynovelty 3/4durability 3/4· Jesse

“As soon as you force them to think about how much you pay for something, it kind of is a forcing function for finding some order of magnitude, some level of scale.”

0.54

Companies historically sampled a small fraction of customer conversations with a team of about 20 people checking a rubric, but a language model can read every conversation, extract insights, flag the roughly 2% of conversations going badly, and draft suggested agent adjustments.

causalhigh valuespeaker onlynovelty 3/4durability 3/4· Jesse

“now what you can do is you can literally have a language model that reads every conversation and extract whatever info you want from it.”

0.53

In the long-run end state, each company will have a unified AI agent that serves as the front end for the business, for banks, airlines, or telecoms the agent could be the only thing users interact with, authenticated, with memory, taking actions like booking flights and upgrading seats, effectively a digital concierge.

forecasthigh valuespeaker onlynovelty 3/4durability 2/4· Jesse

“If you're working with a bank or airline or telecom company or whatever, the agent could be the only thing that most users interact with.”

0.52

Pure ChatGPT wrappers without enough functionality will not become real businesses, but agents require software layers for design, guardrails, and teaching new skills, which makes them hard to replace with a model update.

causalhigh valuespeaker onlynovelty 2/4durability 3/4· Jesse

“And especially when you get into agents, an agent is not just a model, right? You have to design it. You have to be able to put in guardrails”

0.52

Agent capability progresses through three tiers: answering questions from static knowledge, answering with real-time account data, and taking multi-step actions such as lost-card replacement with address checks and fraud review, and this tier of agentic action is where LLMs produced a step-function improvement.

causalhigh valuespeaker onlynovelty 2/4durability 3/4· Jesse

“And then the third tier is you're actually taking action. ... And that's why there's been such a step-function improvement with LLMs.”

0.52

Many generative AI use cases struggle to take off in enterprises because there is always risk, since models are non-deterministic and something could always go wrong, and employees fear blame from leadership when a product they release fails.

causalhigh valuespeaker onlynovelty 2/4durability 3/4· Jesse

“because there's risk involved. People don't want to feel like something could go wrong for them. When they release your product, leadership's going to get mad at them.”

0.52

Generative AI unlocks super flexible, personalized conversations because older systems required building a gigantic decision tree that forced users down branches with no way back, whereas an LLM abstracts that tree into the model itself, which is the main reason for Decagon's success so far.

causalhigh valuespeaker onlynovelty 2/4durability 3/4· Jesse

“what an LLM does is it abstracts a lot of that tree into the neurons of a language model. So that's really powerful.”

0.52

The most important internal asset for an enterprise deploying AI agents is APIs that let the agent take actions, look up data, and reformat information, because with those, a great experience is likely within the first month; documentation and SOPs can be ingested and converted into Agent Operating Procedures.

normativehigh valuespeaker onlynovelty 2/4durability 3/4· Jesse

“Number one thing is just APIs for the AI to use. So APIs to take action. APIs look up data. APIs to reformat things.”

0.52

Customer service ROI is easy to justify because companies already track conversation volume and can calculate the savings from raising automated resolution from about 15-20% to 50-80%, by taking the total cost and cutting roughly 60% of it.

causalhigh valuespeaker onlynovelty 2/4durability 3/4· Jesse

“If you're able to take that to 50, 60, 70, 80, like that's huge ROI. And it's very easy to quantify.”

0.52

Companies will increasingly use small, fine-tuned models for narrow agent tasks such as deciding which path to take or which data to fetch, because these tasks need little general intelligence, and this improves latency and performance, while frontier models remain necessary for hard reasoning.

forecasthigh valuespeaker onlynovelty 2/4durability 3/4· Jesse

“You can make that a fine-tuned model. That model doesn't have to be good at math or coding or anything.”

0.52

Decagon describes its intensity upfront to recruits so that people self-select into a hardworking culture, and the company attracts people who want to work hard, live near the office, and see the role as the prime of their career, creating strong culture foundations.

causalhigh valuespeaker onlynovelty 2/4durability 3/4· Jesse

“you definitely want to tell people that upfront because you want them to self-select into this culture.”

0.52

OpenAI will move toward applications because API revenue is hard to keep, as models are easy to swap (changing one line of code from OpenAI to Anthropic), whereas switching cloud providers like AWS to GCP is very hard.

causalhigh valuespeaker onlynovelty 3/4durability 3/4· Jesse

“But moving from an OpenAI model to Anthropic model, you just change one line of code.”

0.52

At the end of a discovery exercise most ideas should be abandoned, because in any technology wave only a very small number of good ideas exist, and the founder's job is to find one of them at the right time, which means the right answer is usually non-obvious.

normativehigh valuespeaker onlynovelty 3/4durability 3/4· Jesse

“in hindsight it turns out that in any sort of wave at any time, I would say a very small number of good ideas.”

0.50

Voice-to-voice is more powerful than text-to-speech because it takes in the whole audio of what was said, including cadence and how upset the speaker is, and it reduces latency compared with routing through text.

causalhigh valuespeaker onlynovelty 2/4durability 3/4· Jesse

“The voice-to-voice is powerful because it takes into the entire audio of what you said, so it knows cadence and maybe how upset you are and the tone and everything.”

0.50

An effective discovery call starts with broad questions about projects, time use, and stress, then forms product hypotheses on the fly and asks whether a proposed AI agent would help.

normativehigh valuespeaker onlynovelty 2/4durability 3/4· Jesse

“you start by just doing very high level discovery. Hey, what are the sorts of projects that are ongoing right now? How do you spend your time? What is kind of stressful for you right now?”

0.49

Almost all enterprise AI initiatives are top-down board-level mandates, and a use case gets prioritized only if it can be explained in half a sentence as saving or making money with a quick, visible win, such as saving $10 million.

factualhigh valuespeaker onlynovelty 2/4durability 2/4· Jesse

“pretty much all AI initiatives are very top-down at this point because it is such a board-level mandate.”

0.49

After nearly every fundraising round, Decagon got preempted, which signals a little too much excitement and mania in AI, since it is easy to raise money when there are few AI companies with real enterprise revenue traction, and valuation should not be a major factor compared with business performance and potential.

factualhigh valuespeaker onlynovelty 2/4durability 2/4· Jesse

“So, yeah, pretty much after every single time we've raised a round, we just almost immediately gotten preempted.”

0.49

The best enterprise customers are leadership-level people who are genuinely curious about technology and want to move on AI quickly, cutting through bureaucracy and giving good feedback, and this can be detected in the first conversation, as opposed to people who see AI as a board mandate item.

normativehigh valuespeaker onlynovelty 2/4durability 2/4· Jesse

“You can tell if someone, if they're legit about this is something where I'm going to both push aggressively but also give you a ton of feedback”

0.48

Founders who grew up in competitive math and coding contests are well suited to company building because contests train first-principles problem solving, which transfers to breaking vague business problems into parts, even though the contest problems are not real-world problems.

causalhigh valuespeaker onlynovelty 3/4durability 2/4· Jesse

“what these contests really teach you is you're just solving problems. And those problems of course are not real problems in real life. But the sort of thinking is the same.”

0.48

Voice-to-voice models currently have a hallucination rate roughly 8x higher than text because voice generates far more tokens per sentence, and more tokens create more opportunities for errors, which blocks direct enterprise use.

factualhigh valuespeaker onlynovelty 3/4durability 2/4· Jesse

“I think probably 8x higher or something that.”

0.46

Decagon's founders gained conviction in customer service when they tallied the amounts customers said they would pay per idea, and customer service totaled roughly an order of magnitude more than any other idea.

factualhigh valuespeaker onlynovelty 2/4durability 2/4· Jesse

“if you just kind of tallied up even just the amounts that people said and added them together per idea, this was probably like an order of magnitude more than anything else.”

0.46

After every round Decagon raised, investors immediately preempted, but founders should evaluate investors on how well the business is doing and what its potential is, not on the previous valuation, because valuation is a poor basis for a first-principles investment decision.

normativehigh valuespeaker onlynovelty 2/4durability 2/4· Jesse

“pretty much after every single time we've raised a round, we just almost immediately gotten preempted.”

0.44

Many people conflate forward-deployed engineering, which at Palantir means working almost full-time on a single large deal, with ordinary startup hands-on work that does not scale, and the trend is over-indexed in AI startups.

causalhigh valuespeaker onlynovelty 2/4durability 2/4· Jesse

“I think people are a little bit conflating that with what startups do, which is startups are just very hands-on and do things that don't scale.”

0.43

Much of the engineering work for a serious enterprise AI agent has nothing to do with AI itself: it involves observability into conversations, alerting when things spike, QA, and unit tests for conversations before pushing to end users.

factualestablishednovelty 1/4durability 3/4· Jesse

“There's just all this functionality that's there that doesn't really have anything to do with AI really.”

0.38

Voice is the biggest frontier for agents because spoken language has been humanity's primary interface for roughly 150,000 years, and keyboards and typing only emerged in the last 60 or so years, so any agent communicating with humans must handle voice well, which raises a very high bar because the uncanny valley is large.

causalspeaker onlynovelty 2/4durability 4/4· Jesse

“fundamentally, in any sort of agent that communicates with humans, voice has to be a critical factor because that's just how we communicate”

0.35

Coding is especially well suited to AI because it is tokenizable and models are very good at it, and engineers' sophistication lets them extract large leverage from coding agents.

causalestablishednovelty 1/4durability 3/4· Jesse

“It just happens that coding is tokenizable and the models are really good at it.”

0.34

Teams rally around a visible flagpole goal, healthy competition against clear enemies, and specific milestones; Decagon gave everyone Arc'teryx jackets for a revenue milestone, a trivially cheap reward that effectively bonded the team around a shared goal.

normativespeaker onlynovelty 2/4durability 3/4· Jesse

“If you give everyone a clear milestone and this can be pretty insignificant.”

0.34

Extreme parental discipline in childhood, including about three to four hours of piano daily, pulling out of school for competitions, and no TV or video games, formed Jesse's drive, and he argues that this kind of competitive personality is very hard to instill after childhood because the brain is still forming.

causalspeaker onlynovelty 2/4durability 3/4· Jesse

“that level of discipline and also just competitiveness is very hard to gain that after your childhood is over, because when you're in your childhood, your brain's still forming”

0.34

Enterprises in regulated use cases cannot afford the agent to ever deviate, so steps must always occur in a required order, and an AI system must be designed to sit anywhere along a spectrum from free-form to strictly rigorous depending on the use case.

normativespeaker onlynovelty 2/4durability 3/4· Jesse

“If it's a regulated use case, you cannot afford for it to ever deviate. These three steps always have to be followed in this order”

0.34

Any large, hot market that is worth pursuing will be competitive, and this is not specific to AI; anytime there are big growth opportunities, rational people want to go after them, as seen with Databricks versus Snowflake and Ramp versus Brex.

factualestablishednovelty 0/4durability 4/4· Jesse

“Our whole view is that any space that's worth going after, any large hot market, it's going to be competitive. This is not really specific to AI, right.”

0.33

To replace human labor with AI, a company must first know what good human labor is, and the way to find out is to design a process that extracts answers from the people who do know.

normativespeaker onlynovelty 2/4durability 4/4· Jesse

“If you're going to be replacing human labor, you need to know what good human labor is.”

0.32

Many customers under-engage in discovery when asked for numbers because they are not used to thinking about price, and the founder's job is to convert vague enthusiasm into a quantified scale estimate.

normativespeaker onlynovelty 2/4durability 3/4· Jesse

“people are usually a little bit, almost embarrassed or not comfortable going super deep in these questions.”

0.31

The community of people from competitive math backgrounds has large untapped potential; many historically went into trading or academia, but some of that talent could be redirected into company building, since those with this background tend to be more risk-averse track followers.

normativespeaker onlynovelty 2/4durability 2/4· Jesse

“One of my big theses is that this audience of people or this sort of community of people, there's a lot of untapped potential there.”

0.31

Latency remains a hard problem for enterprise voice agents because the system must identify the request, collect materials, and call APIs before responding, and API calls can take about 10 seconds, so agents must use human-like filler such as asking for a moment to look something up.

factualspeaker onlynovelty 2/4durability 2/4· Jesse

“sometimes if you think about how a human does it, you might have to say something such as, "Give me a sec to look that up," because it actually genuinely takes 10 seconds for the API to come back.”

0.31

The talent war at the application layer is less intense than at frontier model labs because there are fewer top-tier researchers, so Decagon hires applied researchers from Harvard, MIT, and Stanford who are in the San Francisco market at any given time, and expanding to a New York office opens another talent pool.

factualspeaker onlynovelty 2/4durability 2/4· Jesse

“Unfortunately, in the application layer it is not as crazy as the Metas and the OpenAIs. There's only so many top level researchers.”

0.31

The best high-profile investors often refuse to help a company until they have invested, which makes it hard for founders to verify helpfulness, so the best investors demonstrate capability through signal over a few months before the round closes.

factualspeaker onlynovelty 2/4durability 2/4· Jesse

“We've definitely talked to a lot of high-profile investors where they're just not willing to help much until they're invested.”

0.31

Frontier labs like Anthropic are building coding applications such as Claude Code, which compete with Cognition and Cursor, and this pushes coding agent startups to train their own models.

forecastspeaker onlynovelty 2/4durability 2/4· Jesse

“now the labs are quite interested in building applications... those will end up being competitors with Cognition or Cursor. And maybe for that reason it is wise for the coding agents to start training stuff as well.”

0.31

The 996 work schedule that is a meme in AI startups is not healthy, and while intensity is needed, a good balance with time to relax is important; the 996 culture is tolerated in China partly because workers have little job security, which gives employers leverage.

normativespeaker onlynovelty 2/4durability 2/4· Jesse

“I don't actually think 996 is that healthy. It happens in China and everyone's super hardcore.”

0.31

The public sees a cool voice demo and assumes voice is solved, which leads executives to expect rapid enterprise adoption, but going live in enterprise is much harder, so adoption is slower than people think.

forecastspeaker onlynovelty 2/4durability 2/4· Jesse

“the general public will just see a really cool demo and be, 'Okay, oh wow, voice is solved now.' And as a result, enterprise should be adopting it left and right.”

0.30

In a competitive market, company culture can be a real competitive advantage; a good culture is pretty hard to replicate and lasts a long time, especially when everyone is geared toward succeeding and working hard.

causalspeaker onlynovelty 1/4durability 3/4· Jesse

“And if you build a good culture that's pretty hard to replicate, it also lasts quite a long time.”

0.30

Companies already have brand guidelines because they train human agents with them, so those guidelines can be applied to AI agents in the same way, alongside examples of top agents' best work, though the preference (for example, whether the agent should apologize) must be taught.

normativespeaker onlynovelty 1/4durability 3/4· Jesse

“The nice part is that they already have all this training process for the humans and you should ideally be able to apply it in the same way as to the AI.”

0.30

Decagon indexes heavily on raw cognitive ability and intellectual throughput when choosing employees, advisors, and investors, and this can be detected by seeing whether someone thinks about the company from first principles during time spent together.

normativespeaker onlynovelty 1/4durability 3/4· Jesse

“What we really index a lot on is just cognitive ability, raw intellectual throughput.”

0.29

How many use cases to pitch in one conversation depends on the buyer's scope: a senior executive like a COO can be asked about many use cases, while a VP should be asked about one.

normativespeaker onlynovelty 2/4durability 3/4· Jesse

“You have to target it based on what that person owns. If that person's very senior, you can ask about multiple.”

0.29

For a services-heavy AI business, the main constraint is not margin dollars but the inability to hire enough good people quickly, so a business fully constrained by talent cannot scale.

causalspeaker onlynovelty 2/4durability 3/4· Jesse

“It's just prevents you from scaling. No one can hire good people that fast.”

0.28

Competitive contests and company building both have fairly objective results, and that objectivity creates constant motivation to improve, even though success requires long preparation.

causalspeaker onlynovelty 2/4durability 2/4· Jesse

“because your results are fairly objective, there's constant motivation to improve.”

0.28

Jesse leans more toward commercial elements of ideas than his intellectually driven friends, and believes that if optimizing for the highest likelihood of success, founders should index on the commercial side, while acknowledging that other paths exist.

normativespeaker onlynovelty 2/4durability 2/4· Jesse

“But in my opinion, if you were just trying to optimize for the highest likelihood of success, I think you should really index on the commercial side”

0.26

Jesse's first company, Loki, built high-performance video capture software for games and exited in 2021 after a bumpy journey; the product had no market after three months of work, and when both co-founders burned out and left, he felt the psychological toll was greater than Decagon's.

factualspeaker onlynovelty 1/4durability 2/4· Jesse

“you just work really hard for three months and then you realize that obviously had no market.”

0.26

The kindest thing Zhang's parents did was impose extreme early discipline on piano and math that formed his drive, then stop imposing career choices once he was in high school, which he sees as the reason he became self-driven and succeeded.

causalspeaker onlynovelty 1/4durability 2/4· Jesse

“Basically, 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.”

0.26

Decagon selected for people who want to be in the office, many live near it, and it works to make the office a place people enjoy because they spend substantial time there, though Zhang says culture is never fully solved.

normativespeaker onlynovelty 1/4durability 2/4· Jesse

“We have a lot of people that live right next to the office. And part of that is because we've selected for people that being in the office with other people.”

0.26

Across Decagon's interactions, chat and voice are roughly balanced in raw customer count, but large Fortune 100 enterprises are about 90-95% voice and 5% chat because they have long been called by customers.

factualspeaker onlynovelty 2/4durability 2/4· Jesse

“A lot of them are 90, 95% voice and 5% chat.”

0.26

Decagon's product includes a testing or simulation suite where, for example, 10,000 tests are built and each is run five times on a continuous basis, giving a measurable score of how well the AI is performing before going live.

factualspeaker onlynovelty 2/4durability 2/4· Jesse

“we're going to build out 10,000 tests and each one is going to be constantly running five times.”

0.26

Many investors call Decagon's customers for expert calls, and some customers have made significant money from these calls, and some people claiming to be users are not known to Decagon, so reference calls contain noise and underwriting must rely on deep research.

factualspeaker onlynovelty 2/4durability 2/4· Jesse

“Unfortunately, I think what we've seen is that there is a lot of noise there because a lot of people just lie on customer calls.”

0.26

The minimum customer revenue to justify a forward-deployed engineer model is probably about one million dollars.

factualspeaker onlynovelty 2/4durability 2/4· Jesse

“Oh, probably a million. Yeah, yeah, yeah.”

0.26

No company wants to let go of its engineers because it now has coding agents, since there is effectively infinite engineering work to do, so coding agents augment highly paid engineers rather than replacing them.

causalspeaker onlynovelty 2/4durability 2/4· Jesse

“There's infinite engineering work to do. So you're just augmenting them.”

0.26

Early in the AI wave, fine-tuning was widely dismissed because it yielded few gains, was irreversible, and would be thrown out with each new model release, a view that held while applications were still being figured out.

factualspeaker onlynovelty 2/4durability 2/4· Jesse

“Almost no one was doing fine-tuning... Another big reason to not do fine-tuning at that time is that the models are changing so fast.”

0.26

The continuous improvement loop is best described as agent improvement via evals and guardrails rather than reinforcement learning model training, and it can be improved by compiling more evals and more guardrails about what the agent can and cannot do.

definitionspeaker onlynovelty 2/4durability 2/4· Jesse

“And it is not RL model training, more agent improvement via evals/guardrails.”

0.25

Open source and fine-tuned models are not yet smart enough for everything, so there will always be heavy usage of frontier providers like OpenAI and Anthropic, and the near-term outcome is a balance of frontier models and smaller fine-tuned models.

forecastspeaker onlynovelty 1/4durability 3/4· Jesse

“I do think there's still always going to be a huge usage of the OpenAIs, Anthropics of the world because you just need intelligence, you need the best models.”

0.25

A founder's most useful practice is honest introspection about personal strengths, then designing the company around those strengths, such as rational problem solving and execution.

normativespeaker onlynovelty 1/4durability 3/4· Jesse

“one of the best things you can do when you're building company or anything is just have pretty good introspection on where your strengths are.”

0.25

The first stage of starting a company is by far the hardest because it is about finding direction, which is exploratory and cannot be reached by grinding toward a known goal.

causalspeaker onlynovelty 1/4durability 3/4· Jesse

“I think the first stage is by far the hardest because you're kind of just finding direction.”

0.24

Large enterprises want to go live on everything within weeks, but staged rollouts are used to catch problems quickly, using metrics like resolution rate, customer satisfaction, and human-reviewed accuracy, and a 500-person support org may have mid to high six figures of conversations a year.

factualspeaker onlynovelty 1/4durability 2/4· Jesse

“Very fast. I would say even for large enterprises, within weeks, everyone wants to just go live to everything.”

0.24

Early failures in customer-AI interactions were often unanticipated edge cases, such as an agent praising a ticket customer's plan to bring eight homeless people to an event, so products need ongoing tuning to find the right balance between guardrails and flexibility.

factualspeaker onlynovelty 1/4durability 2/4· Jesse

“And the agent was like, "Oh, my God, that's so awesome that you're thinking about doing something nice for the community." Things like that, where it's like, okay, I will not expect that to happen.”

0.22

The highest amount anyone said they would pay for an idea during the ideation sessions was on the order of low to mid six figures.

factualspeaker onlynovelty 2/4durability 1/4· Jesse

“Yeah. So the time was probably on the order of low to mid six figures.”

0.22

Jesse is bullish on Google because AI use cases depend on consumer data and Google is stronger on the individual consumer side, while Anthropic has done less well on the consumer side compared with ChatGPT, which he thinks matters long-term for new data.

forecastspeaker onlynovelty 2/4durability 1/4· Jesse

“I just think that with AI use cases having individual consumers is so important because that's where all the data comes from anyways and Google is much stronger there.”

0.22

Zhang would build an AI portfolio around talent-dense areas, including coding (Cognition and Cursor), hardware (Etched), video models (Pika), healthcare foundation models (Chya), and physical AI (Physical Intelligence), accepting that hardware is higher variance but high potential.

forecastspeaker onlynovelty 2/4durability 1/4· Jesse

“would want to take a bet on the hardware layer even though it's a much higher variance.”

0.21

Customer service is the enterprise use case with the most traction, along with coding, because the combination of easy ROI measurement and built-in escalation makes it adoptable.

factualspeaker onlynovelty 1/4durability 2/4· Jesse

“Those two things I think are one of the big reasons why it's probably the, I would argue the use case with the most traction at the enterprise. And coding is another one.”

0.18

During intense company building Jesse sleeps about four to five hours a day, which he says is not good because his brain does not function well on less than eight hours.

factualspeaker onlynovelty 1/4durability 1/4· Jesse

“it's probably four to five hours a day. And it's not good because I'm not someone whose brain functions that well on less than eight hours.”

0.18

Chime is a strong example of a digital-native company rolling out AI agents, because its entire organization is sophisticated about AI, tracks everything, makes thoughtful updates, and gives strong product feedback.

factualspeaker onlynovelty 1/4durability 1/4· Jesse

“you have folks at Chime for example, super impressed with the entire team.”

0.18

Zhang sleeps about four to five hours a day during intense weeks, despite believing his brain performs poorly on less than eight hours, and frames the strain as gratitude for interesting problems and a good team.

factualspeaker onlynovelty 1/4durability 1/4· Jesse

“This week in New York, it's probably four to five hours a day. And it's not good because I'm not someone whose brain functions that well on less than eight hours.”