The Shift from Frontier Models to Specialized Open-Source

Decagon’s journey illustrates a common evolution in enterprise AI: starting with frontier models to prove value, then transitioning to open-source models to achieve production-grade performance. Jesse Zhang and Ashwin Sreenivas argue that while frontier models are essential for exploratory tasks, they are often overkill for specific, repetitive enterprise workflows. By fine-tuning smaller, open-source models, Decagon achieves a "triple win": higher accuracy on specific tasks, lower latency, and reduced costs. They emphasize that this is not a trade-off where one sacrifices intelligence; rather, it is a specialization that outperforms general-purpose models in a controlled environment.

The "Model Factory" Approach

The founders describe Decagon Labs as a "model factory" that continuously iterates on model architecture. Because the AI landscape shifts rapidly, they do not build a static set of models. Instead, they constantly train new models and deprecate old ones as open-source capabilities advance. This requires a sophisticated internal infrastructure for evaluation. They stress that public benchmarks are insufficient; enterprises must build proprietary evals tied directly to customer outcomes to ensure the system functions correctly in production.

The False Dichotomy of App vs. Infrastructure

Addressing the debate over whether application companies are merely "thin wrappers" over foundation models, the founders argue that the real value lies in capturing business logic. They note that fine-tuning is often used to optimize for a category of work (like customer support), while business-specific procedures are handled via in-context learning and robust software stacks. Enterprises partner with companies like Decagon because it is inefficient for them to dedicate their own research resources to tuning models for generic business behaviors. The "moat" is not just the model, but the software stack that manages the agent's behavior and integrates it into existing enterprise workflows.

Managing Enterprise Complexity

Decagon highlights that enterprise AI is less about "tokenomics" and more about reliability and performance. For growth-stage companies, the cost per conversation is secondary to the quality of the output. They observe that as their agents become more sophisticated, token usage per conversation actually increases because they perform more internal checks and parallel processes to ensure accuracy. The goal is to build a "glass box" system where the business logic is transparent and manageable, rather than relying on black-box frontier models that are difficult to steer.