Fyxer's email assistant hit $32M ARR in 2025

OpenAI has published a customer case study on Fyxer, the AI executive assistant startup, disclosing that the company grew from $1 million to $32 million in annual recurring revenue during 2025, retains 90% of users after 90 days, and sees 53% of its AI-generated email drafts sent as written.

The technical detail is the more useful part for builders. Rather than treating email as one generation task, Fyxer runs 30–50 specialized models, each handling a narrow step: a reply-decision classifier that sorts incoming mail into response, scheduling action, or information; intent and outcome-prediction models; and retrieval models that decide which stored interactions persist across threads and which are surfaced for a given sender. "Breaking the problem into many smaller models works much better than asking one model to write a good email," co-founder Archie Hollingsworth says.

Training data comes from the years Fyxer spent running a human executive assistant service — more than 500,000 hours of annotated workflows. The company uses supervised fine-tuning with LoRA to produce task-specific variants, and converts user edits into preference pairs for Direct Preference Optimization. Drafting changes ship only after a statistically significant A/B result, which Hollingsworth's team can sometimes reach within a day at current volume.

The piece is vendor-published and unaudited. It does not disclose customer counts, gross margins, which OpenAI models are used at each step, or how the 53% acceptance rate was measured.