On September 14, OpenAI published a customer story on its official site about Fyxer, a startup building an AI executive assistant[1]. Fyxer combines more than 500,000 hours of human executive-assistant workflow data with OpenAI's models to draft replies in each user's own voice[1]. In 2025, Fyxer grew its annual recurring revenue (ARR) 32x, from 1 million USD (about 150 million yen) to 32 million USD (about 4.9 billion yen), making it a notable example of AI agents in real-world use[1].
Splitting Email Work Across Nearly 50 Specialized Models
Fyxer is an AI executive assistant that follows conversations across inboxes, meetings, and multiple apps to draft replies automatically[1]. Two people can receive the same email yet need very different replies depending on the relationship and prior context, which makes the task deceptively hard for AI, according to Fyxer co-founder Archie Hollingsworth[1].
Instead of relying on one large model to handle everything, Fyxer solved this by dividing the work among 30 to 50 specialized, narrowly focused models[1]. When a new email arrives, a reply-decision model first determines whether a response is needed; if so, additional models analyze intent and retrieve the past context that should be remembered[1]. "Breaking the problem into many smaller models works much better than asking one model to write a good email," Hollingsworth says[1].
500,000 Hours of Human Assistant Work Became the AI's Training Ground
Before building the AI product, Fyxer's founding team ran a human-powered executive assistant service that generated up to 5 million USD (about 770 million yen) in revenue[2]. The more than 500,000 hours of workflow data collected there now form the foundation of the AI's training[1][2]. Fyxer uses OpenAI's fine-tuning platform for supervised fine-tuning and Low-Rank Adaptation (LoRA) to create task-specific model variants, and it feeds the edits users make to AI drafts into Direct Preference Optimization (DPO) training data to keep improving in production[1]. New models go live only after an A/B test shows a statistically significant improvement[1].
As a result, 53% of Fyxer's AI-generated drafts are accepted as written, and 90% of users are still paying customers 90 days after signing up, the company says[1].
ARR Up 32x in a Year, With Backing From High-Profile Investors
Fyxer grew its ARR from 1 million USD (about 150 million yen) to 32 million USD (about 4.9 billion yen) over the course of 2025[1]. In just the eight months leading up to its Series B round, ARR climbed from 1 million USD (about 150 million yen) to 17 million USD (about 2.6 billion yen), pointing to an accelerating growth curve[2].
In September 2025, Fyxer raised a 30 million USD (about 4.6 billion yen) Series B round led by Madrona, with participation from Lakestar Capital and Salesforce founder Marc Benioff[2][3]. The funding is earmarked for expanding into the US market[3]. The company was founded by brothers Richard and Archie Hollingsworth, alongside CTO Matt Ffrench, and is based in London[2]. Customers include real estate firms Knight Frank and eXp Realty, and the company holds ISO 27001 and SOC 2 Type II certifications[2].
※1 USD = 154.1 JPY (as of September 2026)
Summary
OpenAI's case study highlights how, rather than relying on a general-purpose large model as-is, Fyxer achieved production-grade accuracy through continuous fine-tuning on real workflow data and a division of labor among smaller specialized models[1]. With ARR up 32x in 2025 and a Series B round backed by high-profile investors, Fyxer stands out as a concrete example of AI agents at work, and its next moves are worth watching[1][2][3].
Source: https://openai.com/index/fyxer
Source: https://sifted.eu/articles/ai-exec-assistant-startup-fyxer-raises-30m-to-expand-to-the-us
