OpenAI Publishes Case Study Video on Fyxer's AI Executive Assistant

The three-minute-and-sixteen-second video METAL reviewed opens with one customer's story. Mike, a commercial real estate manager who deals with hundreds of buildings and hundreds of people, drives all day and cannot text, so he spends an hour or two every evening catching up on email instead of being with his family. Co-founder Archie Hollingsworth says in the video that a real customer told the team this year, "I think you've saved me from getting a divorce." The goal the company describes is for Mike to open his laptop and find everything already read and sorted, with only the items that need him sitting at the top.

The company starts by explaining why email replies are a hard problem. Two people can receive the same message and need completely different answers depending on the relationship, what came before, and what each is trying to get done. Hollingsworth cites Moravec's paradox, the observation that "things that humans find easy are hard for computers, and things that computers find easy are hard for humans," and admits the task is far harder than it looks. So instead of asking one model to write a good email, Fyxer split the work into 30 to 50 small specialized models. When a new message arrives, a reply-decision model first classifies it: does it need a response, is it a scheduling action, or is it just information the user should see. If a reply is needed, other models read the intent and predict whether the conversation is heading toward booking a meeting or resolving a request.

An executive office is the easiest way to picture it. Rather than one capable assistant doing everything, there is someone at reception, someone handling the calendar, and someone keeping track of what happened before. The company says the memory role matters most. It has to decide which details carry over to the next conversation and which are used once and dropped, and when a new email comes in, retrieval models compare it with stored interactions and surface only the memories relevant to that person and that thread. "Breaking the problem into many smaller models works much better than asking one model to write a good email," Hollingsworth says in the case study.

The 500,000-hour figure comes from a business that existed before the AI product. For several years Fyxer ran a human-powered executive assistant service, and along the way it accumulated annotated work records capturing the small judgments behind a good response: when to answer right away and when to wait, and which earlier conversation matters now. The company says it used supervised fine-tuning and LoRA, a method that adjusts only a small part of a model at low cost, to build task-specific model variants, relying on OpenAI's fine-tuning platform early on and more recently working with OpenAI's managed fine-tuning team to put a new checkpoint into production. "OpenAI has been pivotal for us in helping us transfer the learning that we have about our customer and successfully incorporate it into how the models work," says Joey Dwonczyk, an AI/ML product engineer at Fyxer.

Learning continues after deployment. When a user edits a draft before sending it, the difference between the original and the final version becomes a signal of which one they preferred, and the company converts those pairs into training data for Direct Preference Optimization (DPO). It works like a new assistant learning from a boss's red-pen corrections, except nobody has to label every example by hand. Every change to draft generation goes through an A/B test, and a new version ships only when it shows a statistically significant improvement. With its user volume, the company says, that threshold is sometimes reached within a day.

The results are clear even on the company's own published numbers. According to the case study, annual recurring revenue grew from $1 million to $32 million in 2025, and in the video Hollingsworth says the company went from zero to $30 million in its first year. He argues that retention, not revenue, is the real metric. "Everyone talks about ARR, but I think retention is the real flex. Over 90% of our users are still paying at the 90-day mark with us, and still using us every day," he says in the case study. The company adds that many of its customers are not deeply technical, and for some, Fyxer is the first AI system they use every day.

Read the other way, 47% of drafts still pass through human hands, and the company does not hide that email is a tricky job for AI. Scenes in which people describe the OpenAI team as feeling "like a five star hotel" are a reminder that this is OpenAI introducing its own customer. Even so, the methodology is worth studying for other companies. METAL has reported on the private AI assistant Instinct raising its valuation fivefold in three weeks and on the controversy over that assistant's access to email, and as more products ask for the whole inbox, how a company earns trust becomes its competitive edge.

The ingredients of trust that Fyxer shows come down to three things: a model structure that splits the work into small pieces, 500,000 hours of records of people actually doing the job, and a loop that relearns every day from what users corrected. "We want to get them to a place where they never have to open their computer and can trust Fyxer to manage all of that," Hollingsworth says in the case study. The next contest among AI assistants will be decided not by bigger models but by how quickly a company can shrink that 47% of drafts users still rewrite.