Person
2 episodes · 41 quoted claims
Jesse Zhang started a company after graduating from Harvard, which was later acquired by Niantic, before he left to found Decagon.
“started a company right out of school. Uh, that company was eventually bought by Niantic, and then I left to start this company.”
Jesse Zhang grew up in Boulder, competed in math contests and studied computer science at Harvard.
“grew up in Boulder, did a lot of math contests, stuff like that growing up, studied C S at Harvard.”
Jesse Zhang says Decagon's customers first measure what fraction of conversations the AI agent can handle, then customer satisfaction and NPS.
“So for our customers, it's always the same. It's what fraction of total work, i in this case, like conversations, can the AI agent do?”
Zhang says Bilt Rewards, nearly a year into using Decagon, restructured its customer support team.
“And then now, uh, basically, we're almost almost a year in at this point. They've been able to really restructure their customer support team.”
Zhang says Bilt Rewards' published case study quantified savings of around 65 agent headcount.
“So far, uh, it's around sixty five agents of just like headcounts.”
Zhang describes the orchestration layer as combining multiple models, using evaluations that measure each model's strengths, around a customer's business logic.
“the orchestration layer is kind of how you can use all these different models together. You probably have evals set up that measure how good each model is at certain things.”
Zhang says Decagon uses language models to read every conversation, find knowledge gaps and categorize trends, since humans cannot review millions of conversations.
“If you have a million conversations, it's like, okay, like, no one's reading all those.”
Zhang says instruction following, rather than quantitative reasoning, is the type of model intelligence that matters most for Decagon's agents.
“the type of intelligence that matters the most, we would probably describe it as instruction following.”
Zhang says Decagon is experimenting with voice-to-voice models, with OpenAI doing significant work in the area.
“So nowadays you have the the voice to voice models that we're playing around with. Uh, OpenAI's doing a lot of work here.”
Zhang says voice-to-voice models have low latency but some production uses need extra computation to fetch data and make multiple model calls.
“Voice to voice latency is great. Uh, sometimes though, with these production use cases, you do need the extra computation cycles to, you know, fetch data, do multiple model calls.”
Zhang says he angel invests in several fellow contest-alumni startups, and their founders in turn invest in Decagon.
“I angel invest in a lot of the companies you just listed. Uh, a lot of their founders are angel investors in our company.”
Zhang says his co-founder Ashwin grew up in India and also took part in math and coding contests.
“my co-founder Ashwin also similar background. Um, he didn't grow up in the US. but in India, he did a lot of these contests as well.”
Zhang said Anthropic's computer use demo is, in his opinion, not yet ready for production use.
“we've seen the computer use demo from Anthropic. Uh, probably, in my opinion, not production ready yet.”
Zhang says buyers of customer support agents often test them on 1% of volume and scale up, benchmarking against alternatives.
“Let me just put you out into production for, you know, 1% of the volume and build up from there”
Zhang says incumbents such as Salesforce are also pursuing AI customer support agents and will be alternatives buyers benchmark against.
“a lot of the old school companies, like Salesforce, are gonna be this is a very exciting space for them too.”
Zhang argues that because models are non-deterministic, buyers will be slow to trust generative AI agents for security work that must catch every small event.
“because the models are in inherently non-deterministic, it's very hard for buyers to, like, really trust a gen AI solution there, and so especially, like, agentic solution.”
Zhang cites text-to-SQL companies as an example where agents seem promising but the return on investment is hard to measure because humans must still monitor them.
“there were a lot of, like, text to SQL companies, like, stuff like that, where you could kind of see it working.”
Zhang cites coding agents as an example of a use case that can be rolled out gradually without being perfect from the start.
“I think coding agents is, like, a good example of this, where, um, you know, you can just, like, section off some tasks for them, and, like, they'll do it.”
Jesse Zhang's first company, Low Key, built high-performance video capture software for video games.
“We basically built high performance video capture software for video games.”
Zhang says his first company, Low Key, exited in 2021, which he called a fortunate time.
“And again, I think we exited at like a very fortunate time. It was 2021.”
About a year and a half into Low Key, Zhang's two school-friend co-founders burned out and left, leaving him alone.
“They both got burnt out and then decided to leave.”
Zhang says his Decagon co-founder Ashwin previously started a company that was acquired.
“He started a company before that got acquired.”
Before choosing customer service, Decagon's founders explored ideas including data analysis, security, pre-sales and operations.
“we were looking at things ranging from like data analysis to like security to, you know, pre sales to ops stuff.”
Zhang says the highest annual price prospective customers named during Decagon's early discovery calls was in the low to mid six figures.
“the time was probably on the order of low to mid six figures”
Zhang says simple chatbots resolve about 15 to 20 percent of conversations, so raising that to 50 to 80 percent produces an easily quantified return on investment.
“If you're able to take that to 50, 60, 70, 80, like that's huge ROI, and it's very easy to quantify.”
Zhang says large enterprises typically launch Decagon's agent on one surface area for about 5 percent of users first, with escalation to humans as a safety net.
“We're just going to choose this one surface area and release it to 5 of the user base.”
Decagon's testing and simulation suite can run around 10,000 tests, each repeated about five times, to score an agent's performance.
“hey, we're gonna build out 10,000 tests. And each one's gonna be constantly running like five times.”
Zhang recounts an early Decagon customer, a large ticketing platform, where the agent responded oddly to a user who could not find their ticket.
“In the early days, we have a customer that is essentially like a large ticketing platform.”
In a case study with the smart-ring maker Oura (named Aura Ring in the transcript), Zhang says one in three customers used to keep demanding a human agent, a rate Decagon later reduced to about one in twenty.
“before having any sort of gen AI system, one in three customer that came in would just not bother saying anything and just keep jamming agent until they got to one.”
Zhang estimates that voice-to-voice AI models hallucinate about eight times more often than text-based approaches.
“I think probably like eight x higher or something like that.”
Zhang says many Fortune 100 customers of Decagon handle about 90 to 95 percent of their conversations by voice and the rest by chat.
“A lot of them are ninety, ninety five voice and five chat.”
Decagon uses 'agent operating procedures' (AOPs), which are standard operating procedures written for AI agents.
“We call them AOPs, agent operating procedures, they're just SOPs, but for AI.”
Zhang says that when Decagon started about two years earlier, almost no one was fine-tuning models and commentators often said fine-tuning did not work.
“When we first started, this was about two years ago, people were still figuring out applications.”
Zhang argues that switching between AI model providers such as OpenAI and Anthropic requires changing about one line of code, which limits pricing power for model APIs.
“It's not like moving from AWS to GCP is like very hard, but moving from a open AI model to Anthropic model, you just change like one line of code.”
Zhang names the financial technology company Chime as an example of an enterprise customer whose leadership and data-driven approach impressed him.
“you have folks like Chime, for example, had super impressed with the entire team top down.”
Decagon rewarded employees with branded Arc'teryx jackets for hitting a revenue milestone.
“Like last year, we had for our revenue milestone, we told everyone we'd get them like super nice jackets, and we got like Decagon Arc'teryx jackets.”
Zhang says his co-founder previously worked at Palantir, which uses forward deployed engineers who embed with customers.
“my co-founder came from Palantir, so they actually have forward deployed engineers.”
Zhang says a Palantir forward deployed engineer typically works nearly full time on a deal worth roughly 10 to 25 million dollars.
“What a forward deployed engineer at Palantir means is you're working on a ten, 25 million dollar deal.”
Jesse Zhang grew up in Boulder, Colorado.
“I grew up in Colorado, in Boulder. Boulder's a pretty academic town”
Zhang says that as a young child he played piano three to four hours a day, before his parents shifted him toward math.
“I played a lot of piano, essentially like three, four hours a day.”