Decagon found customer service was an order of magnitude ahead of other AI use cases, with budgets in the low-to-mid six figures.
Oura Ring cut customers demanding a human from one in three to one in 20.
Jesse Zhang says voice is Decagon's biggest frontier because voice-to-voice hallucinates about eight times more and latency is hard.
How Decagon found the clearest return in enterprise AI
Decagon sells an AI agent that resolves customer service conversations across channels and languages. Jesse Zhang found the market by asking prospects how much they would pay rather than what they wanted, and customer service came out an order of magnitude ahead of every other use case. Voice is the next frontier and the hardest one, because voice models hallucinate more and enterprise latency is unforgiving.
Takeaways
Customer service offers AI's clearest return. Chatbots and IVRs resolve 15 to 20% of contacts, and Jesse expects 50 to 80%. Escalation to a human is built in, so enterprises start with 5% of users. At Oura Ring, customers demanding a human fell from one in three to one in 20.
Agreeing on quality is the deployment bottleneck. Decagon runs a simulation suite of 10,000 tests five times each. The long pole is aligning on what good looks like, covering tone and brand guidelines rather than question-and-answer pairs. Enterprises already hold brand guidelines from training human agents.
Voice is the biggest frontier and the hardest. Spoken language has been the human interface for about 150,000 years, typing for about 60. Voice-to-voice captures tone and cadence but generates far more tokens per sentence, so it hallucinates roughly eight times more, and latency is the hard enterprise problem.
AI eats the labour market from both ends. Engineers are the highest paid and can best leverage AI, so no company is letting them go, and coding ROI gets measured by polling them on productivity. Outsourced tier-one and tier-two support is cheaper and more replaceable.
Culture is the durable edge in hot markets. Jesse says any large, hot market is competitive, citing Databricks against Snowflake and Ramp against Brex, and that this is not specific to AI. Decagon tells recruits upfront about the intensity so they self-select.
Application companies should buy share, not margin. Jesse says exponential cost declines mean application companies should optimise for market share. Even if models stop improving, systems can be re-architected to be cheaper, and he thinks enterprise AI is further behind than people assume.
How Decagon found its market
Ideation conversations asked prospects which use cases they had, how much they would pay, who would approve the spend and how ROI would be presented.
The highest number anyone named was low to mid six figures, and customer service came out an order of magnitude ahead of every other idea.
One company revealed it had a 500-person support organization handling mid to high six figures of conversations a year.
Tap a name to go there · tap a line to see why they are linked
On this episode 2
Show 1
Decagon found that saying upfront that this is a different experience made people willing to give the agent a chance, with better outcomes.
What comes next
Near term, large companies may run different agents for teams with different budgets. Long term Jesse expects a single unified system.
Many customers want sales use cases at the end of a support flow, or the reverse.