Decagon’s Playbook for Building Enterprise AI Applications

Jesse Zhang and Ashwin Srinivas, co-founders of Decagon, discuss their journey building AI agents for enterprise customer support and operations. Initially using frontier models from OpenAI and Anthropic, they transitioned to 90% open-source models approximately one year prior to this conversation, driven primarily by latency requirements for their voice agents rather than cost considerations. They discovered that smaller, fine-tuned open-source models can outperform larger state-of-the-art models on specific tasks, achieving superior performance, lower latency, and reduced costs simultaneously.

The founders address the common misconception that open-source represents a downgrade. They argue this is a false tradeoff—when fine-tuned for specific tasks, smaller models actually outperform frontier models while being faster and cheaper. They maintain a research team specifically for model fine-tuning and have built internal tooling tightly coupled to their use cases, including custom evaluation frameworks that measure end-to-end system performance rather than just isolated task performance.

Regarding the app versus infrastructure company dichotomy, they argue that application companies will continue thriving because they solve real enterprise deployment challenges beyond model capability. These include encoding business logic, handling integrations, managing compliance and QA, and providing tools for continuous improvement. They note that even AGI would require databases, CRMs, and software infrastructure for agents to function effectively.

On forward deployment and agent PMs, they emphasize that these roles should "eat pain and excrete product"—converting customer learnings into scalable core product improvements rather than providing consulting services. Every forward-deployed engineer contributes to core product, and this approach differentiates them from competitors offering black-box solutions where customers depend entirely on FDEs for changes. They call their approach a "glass box" model providing transparency and control.

The founders discuss their evolution from customer support to an "AI concierge" vision—positioning AI agents as the front door handling reactive and proactive customer interactions across support, sales qualification, and operational workflows. This expansion became possible as models improved at following nuanced instructions rather than requiring tight constraints. They built their agent to follow business processes generically rather than customer support specifically.

On go-to-market strategy, they emphasize being sales-led while remaining product-driven. They spend significant founder time on sales (80% for one founder) and have built strong early sales teams from non-traditional backgrounds including Ivy League athletes. They navigate complex enterprise deals by helping customers understand the entire deployment journey, from model risk governance through testing and rollout. They recently converted a Sierra customer who preferred their glass-box, productized approach over Sierra's FDE-dependent black-box model.

They note that international expansion has accelerated due to widespread AI awareness and improved language capabilities, though they remain selective about markets and ensure adequate resource investment. They expect consolidation in both geographic markets and verticals, with horizontal platforms winning rather than pure vertical solutions.

On the jobs debate, they argue AI kills jobs but not careers. Their customers typically expand support availability when costs drop rather than reducing headcount, creating latent demand. They've observed BPO employees transitioning to other roles rather than being laid off. They anticipate future value will come from AI-enabled revenue-generating use cases rather than just cost-cutting.

On hiring, they note that despite AI productivity tools, they continue aggressive hiring because competitors have access to the same tools. Everyone uses AI to build more, requiring matched investment to maintain relative velocity. They haven't seen the "one-person unicorn" narrative play out in practice.

The founders discuss using AI personally—Jesse built context-capturing agents that monitor business developments (hires, deals, challenges) to provide informed decision-making support. They note frontier models are good at brainstorming but struggle without proper context. On X versus LinkedIn presence, they found X valuable for timeline-setting and broader ecosystem influence despite being primarily a customer acquisition channel through secondary effects like media coverage.