Interview, Fireside Chat
How Decagon Runs 90% of Its Agents on Open-Source Models
- The narrative regarding Frontier labs dominating the AI startup landscape by the first half of 2030 is a past perspective, with current expectations that AGI will not eliminate software as a whole but rather create enduring needs for agents to store work, retrieve information, and reason.
- Decagon's core product is an agent designed to follow business processes rather than just customer support, with a long-term vision to become the "front door" for every business by 10 years from now, regardless of model capability growth.
- While the current market share of open source inference is declining due to the immediate ease of Frontier models for experimental projects, a shift is expected as enterprises move to production rollouts where latency, cost, and token usage make open source models the preferred choice.
- The adoption of open source post-training models is predicted to take longer than anticipated due to non-trivial fine-tuning requirements, specific data needs, inertia, model risk governance, and security requirements.
- A competitive moat for the next three years depends on infrastructure and software that ensure safety, compliance, and integration within enterprises, which may become commoditized later causing a significant shift in the competitive landscape.
- Decagon aims to compress the time between new model releases and fine-tuned, task-specific versions, building internal tooling for training tasks tightly coupled to use cases while outsourcing generic data labeling.
- Growth-stage companies prioritize agent performance and customer success over cost optimization, reserving tokenomics and cost efficiency strategies for after market dominance is achieved.
- The company employs a forward-deployed engineer strategy focused on productizing learnings from customer interactions to scale products rather than acting as a consulting team, with a goal to evolve these roles into process improvement and productization.
- Hiring remains a primary bottleneck for the company, necessitating that it build three times as much as competitors to maintain its lead despite improvements in AI coding tools, as humans are still required for high-level decisions on what to build or exclude.
- International expansion is accelerating faster than anticipated due to reduced language barriers and top-down C-suite pressure, though data residency, local competitors, and regulatory nuances remain significant hurdles.
- The long-term market outlook favors horizontal platforms over pure vertical solutions driven by the need for scale, with CRMs remaining valuable as databases of record while AI agents utilize their interfaces for interaction.
- Future AI capabilities are expected to allow agents to perform tasks previously requiring manual intervention, such as writing procedures (AOPs), creating tools, and reviewing millions of conversations to identify trends and optimize model variants.
- Career displacement from AGI is not anticipated; instead, automation is expected to eliminate specific mundane tasks while creating new roles, upskilling BPO employees to revenue-generating activities, and allowing companies to expand support to free users.
- X will remain a key channel for shaping global conversation and sentiment among influencers, while LinkedIn will serve as the primary channel for enterprise product announcements and ecosystem building.
- Business process outsourcing (BPO) employees are expected to be upskilled to perform revenue-generating tasks rather than laid off, as AI shifts company strategies from cost-cutting to revenue-generating conversational interfaces.
- Customers prefer a "glass box" approach offering control over their agents, evidenced by Decagon's ability to spin up seven new journeys in a month compared to a competitor's three years for three journeys.
- Voice-to-voice models and smarter out-of-the-box small models are key areas of development being monitored for future technical breakthroughs, alongside the need for constant model training as the shape of AI models changes.
- Forward-deployed learnings regarding process design and product configuration are used to navigate large enterprise orgs and shorten sales cycles, though the rapid market change makes rigid 12-month roadmaps difficult to define.
- Founders are creating personal AI agents to capture business context for automated reasoning on hiring and deal strategies, though current agents can brainstorm but cannot yet make high-level strategic decisions on exclusions.