Decagon Prices AI Support by the Conversation. The Market Is Wages.

Jesse Zhang is co-founder and CEO of Decagon, which sells AI customer-service agents to banks, airlines, telecoms, and large tech companies. The job, as he describes it to Elad Gil on *No Priors*, is to hold an engaging, personalized conversation, resolve it, save the company money, and ideally drive later revenue because customers stay more engaged. As the product has grown, the team talks less about tickets and more about a conversational UI for the brand — a concierge. Zhang previously founded Lowkey, which Niantic acquired; this time he co-founded with college friend Ashwin Sreenivas. Both are second-time founders.

They started where startups usually start: digital natives that will try a new vendor. Rippling and Notion were early partners and helped grind down product details. Demand then pulled them up-market faster than they expected, because that is where the contact volume lives, and because large enterprises moved faster than the old playbook said they would. Gil’s read is that AI has reopened markets that used to be locked. Zhang adds a more specific shift: buying is now top-down. In the past a single team could vet a tool. Now the C-suite and the board are running an “AI transformation,” and customer service is often the biggest area — or the easiest fruit to pick.

The first metric big companies ask for is efficiency: how much of contact-center spend can you take out? Decagon’s case studies land at 60 to 70 percent, a number clean enough to walk into any room. Customer satisfaction sits beside it, often at the same altitude or higher; users should be happier and more engaged, not merely cheaper to serve. The agent itself is a substitute for mundane human labor, not a rip-and-replace of the stack. Whatever CRM and telephony the customer already has, Decagon integrates and does the work a person used to do. The operational extras come for free: always on, multilingual out of the box, no training cycle, no churn, scale when you need it.

Zhang’s thesis as a second-time founder is that a lot of strong technical people are under-commercial. Go-to-market problems are messier than product problems, which is why engineers often avoid them; they are still interesting, still problem-solving, and if you do them well the company grows because you sell more. The first company is where commercial intuition is thinnest. The second time, he says, it is simply easier to tell a good idea from a bad one. He and Sreenivas both came from technical backgrounds, were at similar stages of life, and had each already started something.

Hiring now selects for intelligence ahead of direct experience. Early on they still wanted some experience and did not hire straight out of college for the first large wave; they do now. The same filter runs across engineering, sales, and marketing. Culturally the company is five days in office, with plenty of people showing up on weekends without it being a requirement. Zhang does not want to over-rotate on “996,” but he does want people who see this chapter as a highlight of their career and will put in the time because they get acceleration and hard problems back. Sarah Guo notes that professional athletes never apologize for training six or seven days a week; Silicon Valley went through a stretch of being embarrassed to say that, and the AI wave has brought it back. Zhang’s observation is blunter: the AI companies that are doing well almost all have heavy in-office cultures, at least at this stage.

Decagon is approaching 200 people, opening New York, spinning up Europe, and thinking about a people function for the first time. The work is now org design and operating cadence, not just more ICs. Asked for the counter-instinct advice a first-time founder should take when the nimble team suddenly has product-market fit, Zhang says they hit their stride early enough that it never felt like a before and after. The real adjustment is time horizon. Early on you *should* be greedy: close the deal, close a few customers. Plan too far out and you can burn a quarter and get nowhere. Once you have legs, you both can and must think medium-to-long term, or things start breaking and you realize you should have scaled the unglamorous parts.

That longer view is mostly org design and product roadmap, not capitalization. A lot of core product work closes no customer this quarter and will still leave you fine for now. Skip it, and six months later each incremental customer costs the same or more because overhead never came down. They study later-stage teams that executed this well. Ramp comes to mind; Databricks, especially CEO Ali Ghodsi — Zhang goes as far as calling him his favorite CEO, for the depth of reflection and for being generous with time. The DNA he wants is the same: very strong technical people who apply that intensity to commercial problems and execution.

Not every engineer needs to be commercial. Founders do, and so do the people immediately around them. That is why Zhang tells engineers who eventually want to start a company to join somewhere the commercials are *already* figured out, so they can watch a business work. The common myth is that a smaller, earlier team is closer to a founder’s-eye view. Join pre-PMF, he says, and you mostly see practices failing; you learn what not to do, and most companies never hit the other side. He describes Decagon now as a golden age for that profile: basics working, plenty still unformed, and a lot of it sitting right next to the commercial line. Gil adds that many companies’ golden periods run from 50 or 100 people up toward 1,000 or 2,000; Google from 2,000 to 15,000 had its own magic. Zhang’s metaphor is a founder training a neural net. His first company, started right after college, was two years of negative examples, brute-forced. Sprinkle in positive examples and the learning rate jumps. New offices are a live version of the same problem: SF culture is a living thing, New York needs people to carry it, and Europe is naturally more isolated — “you have to serve wine at lunch.”

Decagon did not walk in with a preconceived customer-service company. Zhang had empathy for the problem from Lowkey, a consumer product with a lot of users, but the method was commercial: talk to a lot of customers, run a disciplined process for evaluating ideas. Support turned out to be the use case. The real tell was not a TAM slide. It was people willing to sign six-figure contracts while the startup was still at zero ARR. It was the only idea they explored where buyers said, if you did this, I would literally pay you, because I can justify it internally. The counter-argument at the time was that the idea was too obvious — someone would have done it already. Once you are inside it, Zhang says, you see how much more nuance there is than the outside narrative. The sheer fact that people would sit down with two founders and pay them was signal enough.

Gil puts the historical pattern on the table. After Microsoft shipped an OS it forward-integrated into Office — Excel and PowerPoint had been separate companies. Google followed general search with vertical search. Anthropic already has Claude Code and has talked about finance; OpenAI tried to buy Windsurf to get into coding directly. API businesses, he adds, are not automatically low-margin: AWS was dismissed the same way and now throws off enormous cash. Zhang agrees labs should push into the application layer. OpenAI’s revenue and margin, in his view, mostly come from ChatGPT because they own the customer; the API is more of a long-term wedge. They will likely start with consumer and prosumer apps that are self-contained, and only later go deeper into enterprise. He does not think it is useful for Decagon to obsess over what the labs will do next.

The more enterprise you are, the thicker the software on top of the model. Observability and monitoring across conversations, learning from those conversations, dissecting insights, a testing and simulation suite for QA — none of that is “the model.” That is what Decagon is building. The labs have other rooms to clear first; coding is probably one of them. Collaboration is also on the table: the company already has good relationships with the large labs.

Differentiation so far is a young, intense team that moves fast on product and on go-to-market, plus a product bet against classic SaaS. Salesforce Agentforce, Google, and a set of AI-native players already crowd the category. Traditional enterprise software is a technical endeavor: bring in consultants, spend technical resources, grow something powerful that becomes slow and expensive to change. Zhang’s view is that what LLMs unlock is the opposite — non-technical business users building, iterating, and analyzing the agent themselves. Up-market customers like that. When engineering owns the whole deployment, the current approach fits less well; even then, he finds engineers do not want to be on the hook for every small logic change. Let them own how the agent talks to systems and APIs, and offload the business logic to operators. He respects the Salesforce-shaped businesses. He does not think that configuration model is right for the AI era. On their side, they want to keep differentiating on execution.

Gil frames the industry shift as software-as-a-service becoming labor- or cognition-as-a-service: a platform fee plus utilization, because you are augmenting an agent rather than renting a seat in a tool. Zhang says the right meter is use-case specific. A coding agent might reasonably bill on GPU use or cores. Support has a tangible output — the conversation — and buyers already think in cost per contact or cost per conversation. There is no seat to sell. Pricing by the minute of a call is “kind of weird” and would incentivize the agent to keep people on the line. So Decagon sells an allotment of conversations for the term, and the customer burns it down. That is, in his view, clearly the right model for the space. It also rewrites TAM. Per-seat revenue is capped by headcount. Per conversation, Gil says, you are talking about salaries, not seats — the people working in the sector. Zhang goes further: the entire services revenue pool is now in the market, because you are turning service into software. Decagon plus every competitor plus everyone else building AI agents is still a grain of sand.

The product metaphor is concierge, not ticket deflection. As engineers who had never worked a contact center, they assumed customers would want one system that could have any conversation. In practice, a hotel’s reservations team and its customer-service team are different people with different budgets. The long-term direction is still a unified concierge: the default way a user talks to the brand, good enough that consumers stop opening the app or the website. Five-year success, for Zhang, is working with the largest companies, powering conversations for major brands, and getting there with the kind of sharp product and go-to-market he currently admires at Databricks and Ramp. Gil recalls Larry Page saying he wanted a billion-dollar company and meaning revenue, not market cap — insanity at the time, a rounding error later. Zhang will not set a number that far out. Databricks is at single-digit billions of revenue and still calls itself early. Ambition is there; the useful unit is one step at a time.

The agentic future, he thinks, is basically here. Consumer agents already order DoorDash. At some point one will call an airline to reschedule a flight and talk to Decagon’s agent. Near term, they will still speak natural language, because every agent has to stay compatible with humans — a human support agent, a human customer. As agents get denser, more efficient protocols may appear, two systems spinning tokens at each other until the job is done. It will stay rooted in language; he does not see a world of 100% agent-only interaction soon. They are not yet seeing agents writing in at scale. The deeper change is the shape of the conversation: today’s support is mostly reactive — I have an issue, fix it. Over time it widens into purchasing, upsell, and proactive outreach when something is wrong. Those are the moments a personal agent earns its keep: it knows your preferences, knows what to give on if an item is out of stock, and settles it with the brand’s agent. Gil quotes an old colleague’s frame: technology is what the richest people in a society already have, made available to everyone. Roman private baths became household bathrooms. Bill Gates has staff to buy clothes and book flights; everyone else will have agents. Zhang doubts Gates’s people are booking the flights themselves, but he takes the framework. Brand concierges talking to personal assistants will show up sooner than people think.