Decagon runs 91% of a major brand's support with no humans

Decagon, the 2023-founded startup that already counts Eventbrite, Notion, and Duolingo among its customers, just dropped a statistic that should make every support team lead sit up: for one of its largest clients, 91% of all global support is now fully automated. No human in the loop. That number came directly from CTO and co-founder Ashwin Sreenivas, and it’s backed by an unusual multi-model architecture that rewrites queries with a fine-tuned GPT‑3.5 before they ever hit retrieval-augmented generation pipelines.

The company doesn’t rely on a single model to do everything. Jesse Zhang, co-founder and CEO, explained that different OpenAI models handle different slices of the customer interaction — GPT‑4 tackles complex decision-making and API requests, while that fine-tuned GPT‑3.5 variant preps incoming questions for maximum retrieval accuracy. Sreenivas admitted they tested plenty of other configurations, but fine-tuning GPT‑3.5 for the rewrite step gave them the best performance. That willingness to mix and match rather than defaulting to the biggest available model is what makes the architecture genuinely interesting.

Latency obsession runs through the engineering team’s DNA. Sreenivas put it bluntly: “Every second counts when you’re dealing with real-time customer support.” The result is a platform that can be spun up for new clients in days, not months, and that can absorb new OpenAI models almost immediately. Every time a new version drops, the team runs it through their eval suite at speed, looking for any edge in accuracy or response time they can deploy to production.

The roadmap points toward voice next. Decagon wants to bring the same fully automated treatment to phone-based support, which is a considerably harder problem than text. If they pull it off with anything close to that 91% automation rate, the economics of enterprise contact centers start looking very different.