Jesse Zhang on Decagon, AI Customer Support, and Building a System of Intelligence
In this episode of the Founders in Arms podcast, we sit down with Jesse Zhang, co-founder and CEO of Decagon, to discuss AI customer support agents, enterprise adoption, and what it takes to build real software moats in the age of large language models.
Jesse explains how Decagon helps companies automate customer support across chat, email, and voice, while improving customer experience and reducing operational burden. The conversation also dives into enterprise AI adoption, model competition, fundraising discipline, and why the best AI companies may become a company’s system of intelligence.
This conversation dives deep into:
AI customer support agents
Replacing or augmenting support teams
Enterprise AI adoption
Systems of intelligence
Moats in AI application companies
Enterprise sales tactics
Fundraising discipline in AI
In this episode, we cover:
(00:00) What AI customer support agents actually replace
Jesse explains that AI support agents are not simply about replacing humans.
In practice, companies use them in different ways:
Improve customer experience
Increase support efficiency
Avoid hiring linearly with growth
Reduce reliance on BPOs or outsourced support
For some companies the goal is cost reduction. For others, it is better customer satisfaction and faster response times.
(02:49) Why AI support can improve customer experience
Traditional support automation is often frustrating because it is brittle and rigid.
Older chatbots and phone trees tend to:
Misclassify issues
Trap users in fixed flows
Escalate poorly
Lose context
Generative AI makes support systems much more flexible, which can meaningfully improve customer experience metrics like NPS.
(04:41) AI vs BPOs for scaling support
The conversation compares AI support agents with outsourced call centers and BPOs.
Jesse argues that if the alternative is scaling low-complexity support through BPOs, a strong AI agent can be the more obvious choice because it offers:
Faster responses
Lower ongoing training costs
Better consistency
Easier monitoring and iteration
Availability across time zones
(07:39) Why the current generation of AI support is different
Jesse notes that support automation has existed for years, but older systems relied on brittle workflows like:
IVR trees
Rule-based chatbots
Intent routing systems
Hard-coded decision logic
The breakthrough with modern AI is that these systems can now operate much more naturally and flexibly, even though the category is still early.
(09:17) Teaching AI agents like humans
One of Decagon’s core ideas is that AI agents should be trained the same way companies train human support agents.
Instead of forcing business logic into rigid frameworks, Jesse describes Decagon’s approach as using agent operating procedures written in natural language.
This allows companies to:
Define support behavior more naturally
Iterate faster
Maintain systems more cheaply
Reach higher quality more quickly
(11:05) How real AI company growth should be evaluated
Jesse shares a skeptical view of many AI growth claims circulating online.
His basic framework:
Prosumer growth can be real and explosive
Enterprise growth can be real if contracts are large and ROI is clear
Mid-market or SMB AI growth can be shakier due to churn
The takeaway is that not all “fast growth” in AI is equally durable.
(14:23) Why AI adoption is moving faster inside enterprises
Enterprises are under strong pressure from leadership, boards, and investors to adopt AI.
That pressure shortens sales cycles and creates strong tailwinds for companies selling enterprise AI products.
Jesse predicts meaningful support AI penetration in the near term, with the category potentially reaching 10–20% adoption within a year or two.
(19:22) The moat question in AI startups
The conversation turns to one of the biggest questions in AI startups:
If models are easy to access and many companies can build wrappers, what creates real defensibility?
Jesse’s answer is twofold:
The best teams execute faster and build better products
Great AI applications can become a company’s system of record or system of intelligence
(21:33) What a “system of intelligence” means
Jesse introduces the idea that AI support agents can evolve beyond simple ticket resolution.
If an AI agent handles most support volume, it begins to own not just ticket data, but also the business logic around how support should work.
That includes:
Workflow logic
Escalation rules
Tone and language policies
Action rules across backend systems
Company-specific support knowledge
Over time, this makes the AI layer much more embedded and harder to replace.
(22:53) AI agents that take action, not just answer questions
A real support agent does more than provide information.
For example, in financial services, a support agent may need to:
Verify identity
Lock an old card
Reissue a new one
Update backend systems
Follow company-specific workflows
This is where the “system of intelligence” becomes valuable: not just storing tickets, but encoding how the business actually operates.
(25:33) Learning enterprise sales as a startup founder
Jesse discusses the learning curve of enterprise sales, especially for technical founders.
Compared with consumer startups, he finds enterprise sales more systematic because the signals are clearer:
There is a buyer
There is a budget
There is a concrete deal to win
Still, success depends heavily on:
Warm introductions
Trusted networks
Investor access
Understanding enterprise incentives
(29:42) How VCs can actually help enterprise startups
The episode highlights a rare area where venture investors can provide meaningful operational help: enterprise introductions.
For Decagon, investors and backers have been especially useful in:
Opening doors at large enterprises
Accelerating sales cycles
Reaching buyers through board and executive networks
This is especially valuable in AI, where many companies now have internal mandates to explore adoption.
(32:20) The model wars: OpenAI, Anthropic, Gemini, and switching costs
Jesse explains that for Decagon, evaluating new models is fast and highly operationalized.
He says the top models are still clearly led by OpenAI and Anthropic, with Gemini improving as well.
Because the company has strong internal evaluation systems, new models can be tested very quickly and swapped in as needed.
(35:25) What’s real vs hype in AI startup valuations
The discussion closes with a candid take on AI fundraising and inflated valuations.
Jesse argues that founders and VCs are not always perfectly aligned:
VCs can tolerate many losses if a few companies become massive winners
Founders can get trapped by valuations that are too high
Employees may suffer if option pricing gets distorted
Big rounds can create pressure to spend irresponsibly
The core advice: founders should be careful about taking valuations that look great in the short term but may create long-term problems.
(41:17) Should founders take secondary liquidity?
The hosts and Jesse discuss when founders should take money off the table.
The general view is that once valuations become meaningful, founder secondary can be rational because it:
Reduces personal financial pressure
Better aligns founders with long-term company building
Helps founders take the bigger swing more comfortably
The real danger is not necessarily raising a lot — it is raising too much at the wrong price and then spending carelessly.
Key Takeaways for Founders
AI support is about more than cost cutting
The best AI agents improve both efficiency and customer experience.
Modern AI works because it is less brittle
Generative systems are a major improvement over rigid chatbots and IVR flows.
Teaching AI like a human is a powerful design principle
Natural-language operating procedures are often more flexible than hard-coded workflows.
Real AI moats come from embedded workflow intelligence
The strongest products may become a company’s system of intelligence.
Enterprise sales is a learnable advantage
Great products matter, but enterprise sales also depends on relationships, trust, and process.
Fundraising discipline matters in AI
High valuations can be tempting, but founders need to think carefully about long-term consequences.
About the Guest
About Jesse Zhang
Jesse Zhang is the co-founder and CEO of Decagon, an AI company building customer support agents for enterprises across chat, email, and voice.
Before Decagon, Jesse studied AI at Harvard and started a company that was later acquired by Niantic. At Decagon, he is focused on building AI systems that improve customer experience while helping large businesses operate more efficiently.
Listen to Founders in Arms
Founders in Arms is a podcast for ambitious builders—covering startup strategy, AI, fintech, enterprise software, and the realities of scaling companies.
🎙 Subscribe to Founders in Arms on your favorite platform.
💬 Join the conversation at TribeChat.com.
🚀 Discover more insights from founders and operators shaping the future of tech.
