No Priors Ep. 132 | With Decagon CEO and Co-Founder Jesse Zhang
Summary
Decagon’s enterprise wedge is unusually legible: automate high-volume support while preserving or improving customer satisfaction, with reported case studies cutting contact-center or operations spend 60–70%. The company began with digital natives such as Rippling and Notion, then was rapidly “pulled up market” into banks, airlines and telecoms as boards and C-suites turned AI adoption into a top-down mandate.
Customer demand—not a preconceived founder thesis—selected the market. Decagon’s founders explored ideas through disciplined customer conversations; support was the only one repeatedly attracting six-figure contracts when the company had zero ARR. The apparent obviousness of the category mattered less than the concrete signal: customers said, “I would literally pay you money because I can justify it.”
The product substitutes for mundane labor without initially replacing the enterprise stack. Decagon integrates with existing CRM and telephony systems, performs the tasks expected of a human agent, is “awake 24/7,” requires little training and has “no churn.” The longer-term ambition is broader: become the conversational interface through which consumers interact with a brand.
The guest’s proposed differentiation is the thick enterprise software layer around the model, not trying to outbuild the model labs. Labs may move into applications, but likely begin with more self-contained, consumer-primary products, with coding among the areas they tackle before complex enterprise support. Decagon is concentrating on observability, monitoring, conversation analysis, testing and simulation—and on letting business users change agent logic without waiting for engineers.
AI-agent pricing expands the addressable market from software seats toward labor and services budgets. Decagon charges for an allotment of conversations, including or excluding conversations that require a human, matching customers’ cost-per-contact model; per-minute pricing would perversely reward longer calls. The framing is that AI-agent providers remain “a grain of sand” because the broader services TAM can migrate into software.
The endpoint is a unified concierge that handles support, purchasing, upselling and proactive outreach—and eventually communicates with consumers’ own agents. The guest thinks that world is “basically here,” though agent-to-agent customer-service interactions are not yet operating at scale. Agents may develop more efficient protocols, but communication should remain rooted in natural language because both sides must still interact with humans.
Execution is the company-building thesis: speed, intense in-office work, and commerciality at the founder and immediate-leadership level. Approaching 200 employees, Decagon is adding organizational structure, a people function and international offices while studying operators such as Ramp and Databricks. The strategic shift is from short-term deal-closing to scalable product investment; work deferred today may become substantially harder six months later.
Deep dive
1. Enterprise urgency made customer service an unusually fast AI wedge
The Decagon guest describes the company as an AI customer-service agent for organizations with large contact volumes: it holds personalized conversations, resolves issues and reduces operating costs. As the product broadens, Decagon increasingly sees it as a brand’s “conversational UI”—or, in its preferred language, a concierge.
The initial customers were digital-native companies such as Rippling and Notion, which moved quickly and helped Decagon iterate. Large enterprises arrived sooner than expected because that is where most of the large contact volumes were, and because many proved more willing to adopt AI than conventional enterprise-sales assumptions suggested.
Adoption has also become a “top-down motion.” Rather than one team quietly vetting software, boards and C-suites are directing an “AI transformation”; customer service frequently appears to be one of the lowest-hanging fruits because buyers can connect large contact volumes to a clear operating-cost baseline.
The primary scorecard is economic: how much of the contact-center or operations spend can be cut? The guest cites successful case studies showing 60–70% reductions, but says customer satisfaction is commonly measured at the same level—or higher—because efficiency does not count as success if users become less happy or engaged.
The agents are intended as substitutes for mundane human labor, not as an immediate replacement for the enterprise stack. They integrate with the CRM and telephony systems a customer already uses, perform the tasks expected of a human and can scale because they are always on, require little training and experience no employee churn.
2. Commercial signals selected the idea and now shape the organization
Decagon did not begin with a fixed customer-service thesis. The founders tested ideas through customer conversations; support stood apart because multiple prospects offered six-figure contracts while the company was still at zero ARR.
The objection was that customer-service automation looked “such an obvious idea” that somebody else must already have owned it. The guest’s answer is empirical: once inside a seemingly obvious market, its operational nuance becomes visible, while two people attracting conversations and purchase commitments is itself strong evidence that the problem is worth pursuing.
The second-founder lesson is that technical talent can become more commercial. Go-to-market problems are “more hairy” and less appealing to some engineers, but they remain problem-solving; mastering them lets a strong technical team sell more and grow faster. Building intuition for a good idea was much harder during the first company.
Decagon hires first for intelligence rather than exact prior experience, applying that philosophy across engineering, sales and marketing. Early on, experience still mattered: the company did not hire straight from college for its first fairly large group of hires, but now does. The office is five days a week, with weekend attendance common but not required, and the company looks for people who see it as a “highlight of their career,” where extra effort brings career acceleration and interesting problems.
Commerciality is most important for the founders and people immediately around them, not necessarily for every engineer. The guest’s advice to engineers who may eventually start companies is counterintuitive: a post-product-market-fit company where commercial execution is visible may teach more than a pre-PMF team where they never see that side in action and effectively learn what not to do. The first company supplied roughly two years of “negative examples”; positive examples accelerate the learning rate.
3. Scale requires retiring the early-stage “greedy mindset” at the right time
Approaching 200 employees, Decagon is adding leaders, organizational structure and a full-time people function. It has an office in New York and is spinning one up in Europe; each office’s culture can become “its own living thing,” so the company has to be deliberate about carrying its San Francisco culture into new locations, especially where local norms differ and the office is more isolated.
The strategic transition is from optimizing for the next customer to allocating resources across a medium- and long-term roadmap. Early on, a “greedy mindset” is useful: get the deal across the line instead of spending a quarter planning. Once the business has footing, longer-range investment becomes both possible and obligatory.
Core product work illustrates the tradeoff. It may close no customer today, but failing to do it means every future deployment requires the same effort—or more as overhead accumulates. Six months later, the company may regret the omission precisely when installing the missing foundation has become harder.
The guest studies later-stage teams that have scaled execution well, including Ramp and Databricks. The goal is to learn from positive examples while recognizing that Decagon still has much to figure out, rather than relying only on the failures that characterized the first company.
4. Enterprise workflow depth is the defense against model-lab integration
The interviewer’s platform-shift framing invokes Microsoft absorbing applications such as Office after launching its operating system, and Google adding vertical searches. In the AI context, the interviewer points to Anthropic already providing Claude Code and OpenAI having tried to buy Windsurf: platform providers may forward-integrate into major applications.
The guest agrees that labs have strong reasons to push into applications, where owning the customer captures more value than supplying model APIs alone. The guest speculates that an API business may function more as a wedge than as a long-term profit center, while acknowledging that infrastructure businesses such as cloud providers can achieve enormous scale and generate substantial cash.
The guest expects labs to begin with more self-contained, consumer-primary applications and may move into enterprise later, with coding probably among the areas they tackle first. Decagon’s defense is the “thicker” software layer required by enterprise support: conversation observability, monitoring, learning from conversations, insight extraction, QA testing and simulation. Decagon already has strong relationships with larger labs and may collaborate with them, but the guest sees little value in spending excessive time predicting their moves while so much application infrastructure remains unbuilt.
Productization and execution are Decagon’s sharper differentiation from Salesforce and Agentforce, Google and other AI-native players. Nontechnical users should be able to build, iterate on and analyze an agent themselves; engineers can retain ownership of API connections and system interactions while offloading logic-building to business teams. If engineering owns the entire customer-service deployment, this approach may fit less well, but even involved engineering teams may not want to handle every small change. “We just don’t think that’s the right approach for the AI era.”
5. Output pricing leads toward a universal, agent-connected concierge
A customer-service agent has a tangible unit of output: the conversation. Decagon therefore sells an allotment of conversations for a contract term, which customers burn down; the allotment can cover any conversation or only those that do not require a human. Seats do not fit an autonomous worker, while per-minute billing would create the “weird” incentive for an agent to prolong calls.
The interviewer’s TAM implication is that output pricing escapes the employee-seat ceiling and instead maps to the people and salaries in the relevant function. The guest extends that argument: “the entire services TAM” can migrate into software, leaving Decagon, its competitors and the broader agent market collectively “a grain of sand” relative to the opportunity.
A current organizational complication is that a hotel’s reservations and support conversations may belong to separate teams with separate budgets, even though the consumer experiences one brand. The eventual product should unify those interactions and could become the default interface, reducing the need for many visits to apps and websites.
The guest thinks agent-to-agent interaction is “basically here” in consumer contexts—for example, a personal agent already ordering DoorDash, or eventually rescheduling a flight by speaking with an airline’s service agent. Such interactions are not yet happening at scale in customer service. Initially both agents will use natural language for human compatibility; later they may exchange information more efficiently, while expanding from reactive support into purchases, upsells and proactive outreach when an issue is detected. “It’ll be here sooner than later.”