How AI Agents Are Transforming Customer Support, with Decagon’s Jesse Zhang
Summary
Customer support is emerging as the near-term “golden use case” for AI agents because deployment can start small while savings and service quality remain directly measurable. Decagon tracks the share of conversations automated alongside customer satisfaction, NPS, and accuracy, effectively offering every customer a “personal concierge in their pocket” across languages and 24/7.
Bilt Rewards provides Decagon’s clearest proof of operating leverage: within roughly one month, it stopped scaling its support team as AI took over much of the automation. Almost a year later, Bilt had restructured the function and quantified around 65 agents’ worth of headcount saved, while customers posted that its support “doesn’t feel like any sort of AI or chatbot system we’ve ever used before.”
Decagon’s claimed differentiation is software above broadly available models, not exclusive access to GPT-4o, GPT-4, or Claude Sonnet. Its orchestration layer evaluates and combines models around customer-specific business logic; its product layer exposes data, decision steps, knowledge gaps, and conversation categories. Zhang’s framing: “Most of the alpha, or most of the special stuff that you build, is on top of models.”
For customer-service agents, instruction following matters more than the coding and mathematical reasoning gains emphasized around o1 and Sonnet. A support agent must execute an SOP “to a T,” particularly in regulated environments; better reasoning helps, but dependable adherence to workflows is the more consequential model-development vector for Decagon.
Voice expands the addressable workflow but introduces a persistent latency-versus-computation trade-off. Voice-to-voice models respond quickly, while production calls may require data retrieval and multiple model calls; speech-to-text-to-speech adds delay but permits that work. Decagon’s pragmatic bridge includes conversational cover such as, “Hey, give me a second. I’m looking up your data.”
The organizational end state is not simply fewer humans but more people “supervising and editing agents.” Zhang expects leaders to monitor, correct, and hard-code behavior across infinitely scalable systems, while screen context and computer use could let agents navigate products—not merely answer questions—though Anthropic’s demonstrated computer use was, in his view, “not production-ready yet.”
Zhang is “more bearish” on most near-term agent categories than the sector’s demos imply. Successful markets need both incremental rollout before near-perfection and easily quantified ROI: security agents struggle with nondeterminism where any small anomaly matters, while text-to-SQL often remains a monitored copilot whose economic value is difficult to price. Better models may unlock those categories later.
Deep dive
1. Customer support gives AI agents an unusually practical entry point
After selling his first company to Niantic, Zhang and Ashwin pursued AI agents without overengineering the thesis: their biggest prior lesson was that founders “can’t really overthink things too much.” Customer conversations led them to customer service as the “golden use case.”
Founded in August 2023, Decagon now serves support operations at companies including Rippling, Notion, Duolingo, Eventbrite, Vanta, Substack, and Bilt Rewards. The recurring scorecard is straightforward: how much work can AI perform, how much happier are customers, and—especially in regulated industries—how accurate is it?
2. Bilt converts exponential support demand into visible headcount leverage
Elad Gil cited Klarna, a European buy-now, pay-later service, based on a post from its CEO on X. The post reported 2.3 million chats in four weeks, customer satisfaction on par with humans, 25% fewer repeat inquiries relative to people, and two-minute resolution versus 11 minutes for a human agent. It also reported 24/7 availability across 23 markets and 35 languages. Gil said he thought Klarna had shifted 700 full-time agents to other work.
At Bilt, Zhang’s causal chain was simple: inquiries grow roughly linearly with users, so a user base growing “basically exponentially” produces similar support pressure. The opening request was blunt: “Holy crap, we’re getting overwhelmed by all this volume. Can AI help here?”
Within about a month, Bilt stopped scaling the team as Decagon automated more work. Nearly a year later, it had reorganized support and quantified around 65 agents’ worth of headcount saved—an easy-to-express ROI paired with faster service and positive social posts.
3. The defensible product sits above shared foundation models
Zhang views Decagon as a software company because everyone can access the same GPT-4o, GPT-4, and Claude Sonnet models. Differentiation comes from “the orchestration layer or the software around it,” not the underlying models.
Orchestration means evaluating which model handles each task best, combining those models, and molding the resulting system around a customer’s business logic. The surrounding software makes decisions inspectable: users can see the data consulted, steps taken, answers generated, and provide feedback.
At scale, the LLM can read a million conversations that no team could review manually, surface knowledge gaps, categorize demand, and report how things are going. Buyers can benchmark quality against human performance and metrics, then put an agent into production on 1% of volume and build up—possibly testing another option too. Zhang credits observability, explainability, and control for Decagon’s performance.
4. Instruction following and voice latency are the next constraints
Recent developments around o1 and Sonnet have focused on quantitative reasoning, coding, and mathematics, but Zhang says those are not Decagon’s largest bottlenecks. Customer support depends on instruction following: given an SOP or workflow, can the agent “follow it to a T?”
Voice is the same customer-interaction problem through another channel. Decagon customers are testing voice agents as ElevenLabs, OpenAI, and Cartesia improve realism and latency; Zhang’s broader mandate spans chat, email, SMS, and phone because “the channel really doesn’t matter” to the underlying request.
Latency remains central. Voice-to-voice is fast, but production support may require data retrieval and multiple model calls; transcribing, computing in text, and regenerating speech is slower. Teams can stream output or use conversational cover such as, “Hey, give me a second. I’m looking up your data.” Decagon is choosing these trade-offs around each customer’s priorities.
5. The human layer remains a source of operating leverage
Zhang expects a “reasonable explosion” of agents, although some use cases will take longer. Support work consequently shifts toward humans supervising and editing agents, with leaders monitoring behavior, correcting it, and retaining visibility and control.
AI supervision differs from managing people: agents are infinitely scalable, and some behavior can be hard-coded. Zhang sees substantial room to improve the interface for real-time corrections such as, “You did this thing wrong. Please do this next time.”
His founder network supplies another human advantage. Math- and coding-contest peers have moved from academia or quant trading toward startups over the last five or six years, then informally cross-invested and shared data on hiring, sales, and compensation. Contest experience is a “pretty good signal,” Zhang says, but Decagon’s hiring process remains broadly the same and does not require that background.
6. Gradual rollout and legible ROI separate durable markets from demos
Zhang says that, at the current state of the models, the vast majority of use cases will not see real commercial adoption. When Decagon started, he did not know whether real use cases would emerge over the next 12 or 24 months; in hindsight, he says the successful ones need incremental rollout and measurable ROI.
Security illustrates the reliability problem: AI may appear well suited to environments with tons of logs, but the job may require catching every small anomaly. Because generative models are inherently nondeterministic, Zhang expects enterprise adoption of agentic security systems to be “really, really, really slow.”
Text-to-SQL illustrates the economic problem. Buyers may like the output but still require someone to monitor and edit it, turning the agent into a copilot; because most teams do not have that many data scientists, it becomes hard to benchmark replacement value or justify a large contract.
The winning near-term pattern combines coding-agent-style task partitioning with support-style measurement: deploy on bounded work, provide value before perfection, and quantify the return. Zhang remains bearish on many current categories, while preserving the hedge that improving models “will unlock a lot of new use cases.”