Decagon Shifts to Open-Source for Enterprise AI Scale

Decagon initially utilized frontier models to establish early functionality for its AI applications.

However, as the company began working with larger enterprise clients and developed voice agents, the critical need for lower latency became a significant factor.

Jesse Zhang notes that frontier lab models often lack the granular control necessary to align precisely with specific business workflows and requirements.

Currently, 90% of Decagon's workflows are built upon fine-tuned open-source models, demonstrating a strategic pivot towards more controllable and efficient solutions.

Optimization allows for achieving lower costs, lower latency, and better task-specific accuracy simultaneously, challenging the binary intelligence narrative often found on social media.

He argues that fine-tuning smaller models for specific tasks leads to higher performance compared to massive, state-of-the-art models designed for broader applications.

This specialized optimization allows Decagon to achieve lower costs, reduced latency, and enhanced accuracy simultaneously for particular tasks.

The company prioritizes tailoring models to achieve superior results in targeted enterprise applications, rather than relying on the general capabilities of larger, less focused models.

The AI model landscape is evolving too rapidly for static, 'build once' solutions, necessitating continuous adaptation.

Jesse Zhang describes Decagon Labs as a 'model factory' designed to compress the time between the release of new models and their useful deployment in task-specific applications.

This continuous training and deprecation of models are essential for maintaining a competitive edge as frontier AI capabilities advance.

Decagon's approach ensures that its AI agents remain at the forefront of performance and relevance for enterprise clients.

Fine-tuning models is primarily for optimizing specialized use cases, rather than for individual customer customization.

Customizing AI behavior for specific customer procedures is more effectively managed through in-context adjustments within the application.

Business logic, such as the steps involved in complex flight cancellations, operates independently of the core AI model itself.

Successful enterprise AI applications are built on a robust software stack that accurately captures and executes these specific operational logics.

Early-stage AI companies rely on forward-deployed engineers to define and explore workflows that have not yet been formalized.

These engineers embed with customers to understand their unique needs and 'pave the road' for new operational workflows not yet covered by the product.

Long-term reliance on these staff without productizing their discoveries can transform an AI startup into a 'glorified consulting firm'.

True forward-deployed work at Decagon aims to 'eat pain and excrete product,' meaning their contributions should benefit all future customers by enhancing the core product offering.

Decagon distinguishes its approach from companies like Palantir that rely heavily on service-based models.

Over-reliance on forward-deployed engineers risks creating a non-scalable consulting business rather than a product-led company.

To achieve scalability, a company must transition from offering extensive services to focusing on a core product.

Prioritizing product development over merely chasing various use cases is crucial for long-term growth and market position.

This system automates the creation of Agent Operating Procedures (AOPs) and generates comprehensive test simulations for these agents.

The shift from manual coding to plain text instructions for AOPs represents a significant leap in efficiency for agent development.

Duet enables proactive identification of errors and the drafting of improvements, streamlining the iterative process of agent refinement.

Ashwin Sreenivas identifies the current competitive advantage in enterprise AI as the ability to build robust infrastructure that allows companies to safely deploy models.

Enterprises require comprehensive tools for governance, rigorous testing against regulatory standards, and collaborative frameworks for AI implementation.

Integrating AI with existing legacy systems presents a major hurdle that cannot be solved by model performance alone.

Long-term viability in enterprise AI depends on facilitating control and ensuring human oversight in AI-driven workflows.

Initial sales success for Decagon was driven by a 'hot' market and strong early sales talent, leveraging individuals with Ivy League and non-traditional backgrounds.

Jesse Zhang dedicates approximately 80% of his time to sales, emphasizing direct founder involvement in calls.

This deep engagement aims to accelerate the sales process and identify opportunities to productize specific enterprise processes.

The strategy focuses on securing early wins by targeting one or two top use cases within each client organization.

Early limitations in model capability and client needs initially constrained Decagon's agents to customer support.

The company has evolved its agents to handle inbound sales and proactive operational workflows.

Ashwin Sreenivas states that the core differentiator is an agent designed to follow general business processes, not just perform specific support tasks.

Improved instruction following capabilities now allow for more open-ended discovery and execution of complex operational requirements.

Ashwin Sreenivas notes the difficulty of maintaining a rigid 12-month product roadmap in an AI market that changes daily.

Decagon prioritizes building features immediately in response to strong customer demand signals.

While the long-term vision of becoming the 'front door' of a brand remains, the company's execution strategy is highly adaptive.

This flexibility allows Decagon to rapidly integrate new model capabilities and address emerging client needs as they arise.

Jesse Zhang clarifies that office attendance is voluntary, serving primarily to maximize communication and collaboration among team members.

The company fosters a 'team sport' environment where engineers and sales staff work together directly to solve problems.

This collaborative atmosphere focuses on collective achievement and tackling significant tasks rather than individual 'grind'.

Decagon builds its AI platform horizontally, believing that broad scalability and deep product robustness offer greater value than niche, vertical solutions.

Jesse Zhang cites historical precedents such as Salesforce and Zendesk, which achieved success through horizontal platforms.

The company anticipates future consolidation in the AI market, with platforms offering greater depth and versatility emerging as winners.

This strategy positions Decagon to serve a wide range of enterprise needs rather than being confined to a specific industry segment.

Jesse Zhang has developed internal AI agents to track hires, deals, and ongoing operational challenges within Decagon.

These agents serve as a repository of 'business context,' providing tailored advice for complex decisions.

The AI model assists in identifying complementary skills among job candidates, preventing the repetition of existing team patterns.

The ultimate goal is to outsource decision-making support by comprehensively capturing and leveraging institutional context across the organization.

Decagon treats LinkedIn as a channel for tactical updates and fundraising, whereas X serves as a primary timeline for shaping the tech ecosystem's broader narrative.

Jesse Zhang explains that X (formerly Twitter) is utilized to influence the broader industry narrative and shape the 'mindshare' of the tech ecosystem.

Content shared on X by influencers and podcasts creates secondary effects, ultimately reaching enterprise decision-makers.

The founders clarify that they personally write their posts, only using AI for brainstorming potential topics rather than generating content.

Jesse Zhang observes that customers often reinvest the 30% cost savings from AI automation into expanding service access rather than reducing staff.

For instance, a company might transition from a hidden support panel to a prominent, universally accessible support system.

The founders frame this phenomenon as an application of Jevons paradox, where increased efficiency leads to greater overall consumption of the service.

They assert that AI will transform jobs by eliminating mundane, repeatable tasks, rather than causing widespread career elimination.