CHAPTERS
- 0:00 – 1:02
Decagon’s mission: enterprise-grade AI agents for customer support
Elad introduces Jesse Zhang and Decagon, framing the company as a fast-moving startup building generative AI for enterprise customer support. Jesse sets the context for why customer interactions are a “golden” agent use case and what makes Decagon distinctive.
- •Decagon builds AI agents focused on customer service/customer experience
- •Rapid adoption across large enterprises and fast-growing startups
- •Customer support seen as a natural early fit for LLM-powered agents
- •Company emphasis on being production-ready, not just demos
- 1:02 – 2:32
Founder origin story and finding the initial wedge through customer discovery
Jesse shares his background (math contests, Harvard CS, first startup acquired) and how he met co-founder Ashwin. They started from broad interest in agents, talked to many potential customers, and converged on customer service as the most compelling initial application.
- •Jesse’s path: math contests → CS at Harvard → startup acquired by Niantic
- •Early approach: don’t overthink, talk to lots of customers
- •Discovery process led to customer service as the clearest wedge
- •Focus on making early design decisions based on real customer needs
- 2:32 – 3:15
What Decagon ships: agents plus transparency, observability, and control
Jesse explains Decagon’s product positioning: AI agents in production for large support organizations, built to avoid the “black box” problem. Transparency and feedback loops are highlighted as critical for enterprise trust and ongoing improvement.
- •AI agents deployed for customer service and customer experience teams
- •Transparency: show what data was used and how answers were formed
- •Tooling for feedback, oversight, and iterative improvement
- •Target customers: any company with sizable support operations
- 3:15 – 6:15
Business impact metrics: cost savings, deflection, and customer satisfaction
Elad references Klarna’s widely shared results to illustrate how dramatic the upside can be. Jesse generalizes what customers care about most: share of conversations automated and customer happiness, with accuracy as a gating requirement in regulated contexts.
- •Primary KPIs: % of workload handled and CSAT/NPS outcomes
- •Secondary constraints: accuracy/compliance needs by industry
- •Benefits include 24/7 coverage, faster resolution, and language scaling
- •Agents can shift humans to higher-value work rather than pure replacement
- 6:15 – 8:00
Case study: BILT Rewards—scaling support without scaling headcount
Jesse walks through a concrete example of ROI and operational change at BILT Rewards. Rapidly increasing user volume created exponentially rising support demand; Decagon helped stabilize team growth and improve responsiveness.
- •Support volume grows with user base; hypergrowth overwhelms teams
- •Within ~1 month: slowed/paused support team scaling via automation
- •After ~1 year: restructuring support operations around agent capability
- •Reported impact: ~65 support-agent headcount worth of savings; better user experience
- 8:00 – 10:17
Under the hood: orchestration layer + classic software to make models usable
Jesse describes Decagon as primarily a software company leveraging shared foundation models. Differentiation comes from orchestration (combining models, evals, business logic) and the surrounding product (analytics, transparency, conversation understanding at scale).
- •Foundation models are commoditized inputs; value accrues in software layers
- •Orchestration: multi-model routing, eval-driven decisions, business-logic shaping
- •Visibility tooling: data provenance, step tracing, conversation analysis
- •LLMs summarize trends, detect knowledge gaps, and categorize support themes
- 10:17 – 12:06
What’s missing for next-level agents: instruction-following over pure reasoning
Elad asks what technology gaps remain for agentic systems. Jesse contrasts recent progress in quantitative reasoning (math/coding) with what matters most in customer support: strict instruction-following aligned to SOPs and workflows.
- •Different “intelligence types” matter for different applications
- •Customer support agents need reliable instruction-following and policy adherence
- •Non-plateau view: progress may be shifting from reasoning to other competencies
- •Expectation of continued lab research improving instruction compliance
- 12:06 – 14:28
Voice agents arrive: multi-channel support beyond chat and email
The conversation shifts to voice-based customer support. Jesse explains why channel flexibility matters and why voice is increasingly requested once companies see text agents working in production.
- •Customers want the same agent capability across voice, chat, email, SMS
- •Enterprises are now piloting voice agents after validating text agents
- •Voice quality and realism have improved via providers like ElevenLabs/OpenAI/Cartesia
- •Decagon works closely with model providers to make voice work at scale
- 14:28 – 16:24
Latency trade-offs in voice: voice-to-voice vs speech-to-text pipelines
Elad highlights latency as a core UX blocker for voice interactions. Jesse outlines architectural options—voice-to-voice models for responsiveness versus text pipelines for richer orchestration—and how product techniques can mask unavoidable delays.
- •Latency sources: transcription, LLM processing, tool/data fetches, TTS generation
- •Voice-to-voice can reduce latency but may limit complex multi-step workflows
- •Speech-to-text → compute → TTS offers flexibility but adds delay
- •UX mitigations: streaming, partial responses, and “one moment while I look that up” patterns
- 16:24 – 18:27
Why math/competition communities show up in AI startups
Elad asks about the strong presence of Math Olympiad/IOI-style founders in AI. Jesse attributes it to an existing tight-knit talent pool and a broader shift of top technical talent toward startups as that path became mainstream.
- •Math/coding contest communities created dense networks early
- •Many founders are same-age peers who’ve known each other for years
- •Shift from academia/quant trading toward startups over the past ~5–6 years
- •Community visibility accelerates learning and company formation momentum
- 18:27 – 19:50
Informal founder networks: advice, angel investing, and shared learning loops
Jesse describes a largely informal support system among peer founders—mutual angel investing, frequent hangouts, and exchanging hard-won operating lessons. The benefit is breadth: company-building has massive surface area beyond pure engineering.
- •Mutual angel investing between peer founders/companies
- •Peer support helps across hiring, sales, compensation, and operations
- •Social bonds (games, regular meetups) enable continuous knowledge transfer
- •Founder communities tend to rotate every 5–7 years with new cohorts
- 19:50 – 21:09
Hiring signals: contests help, but building teams requires broader lenses
Elad asks whether Jesse’s background changes how Decagon hires. Jesse notes contest credentials can be a useful signal and community catalyst, but emphasizes that talent is widespread and the hiring process can’t rely on pedigree alone.
- •Contest backgrounds can be a strong signal when shared context exists
- •Ashwin’s similar contest history reinforces the lens but doesn’t dominate it
- •Decagon still hires broadly; many great candidates lack contest pedigrees
- •Community events/puzzles are effective for meeting strong engineers
- 21:09 – 25:07
The next 12–24 months: multimodal context, computer use, and supervising agents
Jesse outlines what he’s excited about: better models across modalities and deeper contextual understanding (e.g., screen/context awareness). He also predicts job roles will shift toward supervising and editing fleets of agents, creating new product surface area for control and monitoring.
- •Multimodal improvements: richer context like “entire screen” understanding
- •Agentic UI automation (computer-use demos) is promising but not yet production-ready
- •Work evolves toward supervising/editing agents rather than doing tasks end-to-end
- •Decagon focuses on visibility/control systems analogous to managing humans
- 25:07 – 30:09
Where agents win (and where they won’t yet): rollout path + measurable ROI
Closing discussion covers the near-term limits and advantages of agent deployments. Jesse argues many domains will adopt slowly due to non-determinism and trust requirements, while successful early categories allow incremental rollout and offer clearly quantifiable ROI—like customer support.
- •Many use cases require near-perfection immediately (e.g., security), slowing adoption
- •Non-deterministic behavior makes trust and risk management harder for buyers
- •Other areas (e.g., text-to-SQL) struggle with ROI clarity and become “copilots”
- •Best near-term use cases: incremental deployment + easy-to-measure business impact
