How Enterprise AI Really Gets Deployed

Episode Summary

Executive Summary: Decagon co-founders Jesse Zhang and Ashwin Srinivas argue enterprise AI’s real value comes from product design, workflow integration, and model tuning—not just frontier labs. They explain why Decagon now runs 90% of its workflow on open-source models for latency, control, and task-specific performance, while still using frontier models for exploratory work. The episode reframes AI apps as durable infrastructure-plus-product businesses, not thin wrappers.

Main Topics: Open source vs. frontier models in enterprise AI (Priority: 5/5): The founders describe Decagon’s shift from OpenAI/Anthropic to mostly open source because smaller, fine-tuned models deliver better latency, lower cost, and higher performance on narrow tasks. Frontier models remain useful for broad, exploratory, or rapidly changing problems. Agents as business-process executors, not just chatbots (Priority: 5/5): Decagon emphasizes that its core value is building agents that follow business processes reliably across customer support, sales, and operations, rather than merely answering questions well. Enterprise deployment requires productized infrastructure (Priority: 5/5): The discussion stresses that enterprises need controls, testing, governance, integrations, and visibility around AI systems, so the moat lies in the software stack surrounding the model. Forward-deployed work as a path to productization (Priority: 4/5): They argue FDEs and agent PMs should not become consulting layers; instead, they should surface workflows and pain points that are then turned into reusable core product features. Why applications and SaaS are not disappearing (Priority: 4/5): The speakers reject the idea that AGI will eliminate software, arguing that even highly capable agents still need systems of record, storage, coordination, and reasoning layers. Go-to-market, speed, and founder-led enterprise sales (Priority: 4/5): Decagon says its growth is driven by fast enterprise sales, deep founder involvement, and a product-led but sales-informed motion tailored to large regulated customers. Jobs, careers, and demand expansion (Priority: 3/5): They contend AI will automate tasks and some jobs, but not careers; cheaper support and better service often expands demand, leading companies to redeploy people rather than simply lay them off.

Key Arguments: Smaller open-source models can outperform frontier models on narrowly defined enterprise tasks when fine-tuned and evaluated against customer-specific outcomes. Latency is a decisive enterprise requirement, especially for voice and conversational agents, making smaller controllable models preferable in production. Decagon’s moat is not just model access; it is the ability to deploy AI safely inside enterprises with controls, testing, workflows, and legacy-system integrations. Applications remain necessary because agents need business logic, operational workflows, and systems of record even in an AGI world. Forward-deployed work is valuable only if it feeds reusable product; otherwise the company risks becoming a services firm or consulting shop. The best enterprise AI companies will be product-driven, but sales-led learning is still essential because workflows and customer needs are still being discovered. AI adoption can increase demand instead of destroying it, as cheaper support and higher quality service let companies offer more help and expand usage.

Data Points: Share of Decagon workflow on open source: 90% - Jesse says most of Decagon’s workflow now runs on open-source models. Share of workflow still on closed/frontier models: 10% - Frontier models are still used for new projects, auxiliary tasks, and exploratory systems like Autopilot. Model evaluation dimensions: 3 - The founders describe models as balancing cost, intelligence, and latency. Timeframe for open-source adoption: About a year plus ago - They began shifting toward open source after latency and controllability became more important at scale. Founders’ time spent on sales: 80% - Jesse says he spends most of his time on sales and enterprise deal progression. Original customer support ticket volume example: 50,000 per month - A customer example used to show how lowering support costs can increase demand. From one to seven journeys: 7 new journeys in a month - A customer moved from building three journeys in a year to seven on Decagon in roughly a month. Old pace at a prior vendor: 3 journeys in a year - Used as comparison to illustrate Decagon’s faster iteration and more productized approach.

Pivotal Quotes: "An AI agent should just be the front door of your business." — Podcast intro / framing quote: Introduces Decagon’s vision that AI should handle both reactive and proactive customer interactions. "The thing that we built was not an agent that does customer support well, but rather an agent that follows business process well." — Jesse Zhang: Explains the core product thesis behind Decagon’s expansion beyond support into broader concierge-like workflows. "I think the point is that at a certain point, it's strictly better to use open source models because when your use case is solidified and you're in production at scale... there’s no reason not to use open source." — Jesse Zhang: Summarizes why Decagon shifted most of its stack to open source once production requirements stabilized.

Implications: Enterprise AI winners will likely be product companies that own workflow, governance, and deployment—not just model APIs. Open source, fine-tuning, and FDE-led discovery will matter more as AI moves from demos to production.