Fyxer’s OpenAI Case Study Turns a Decade of Assistant Work Into an AI Moat

OpenAI published a new customer case study on September 14, 2026, and the most interesting part of it has nothing to do with the model. It is about where the training data came from.

The startup, Fyxer, builds an AI assistant that reads a user’s email, decides whether it needs a reply, drafts one in that person’s own voice, and takes meeting notes along the way. According to OpenAI’s write-up, 90 percent of Fyxer’s users are still active 90 days after signing up, and 53 percent of its AI-generated drafts get sent exactly as written, with no edits. Those numbers are not the product of a single clever prompt. They are the product of a data-collection strategy that started years before Fyxer ever touched a large language model.

Thirty Small Models Instead of One Big One

Fyxer’s system does not ask one model to read an email and write a good reply. Per OpenAI, it is built around 30 to 50 specialized models, each responsible for a narrow slice of the workflow: a reply-decision model first classifies whether a message needs a reply, is a scheduling action, or is just information to note, other models then analyze intent and predict where the conversation is headed (toward a scheduled meeting, a resolved request, or an ongoing thread), and a separate set of retrieval models decide which stored details from past interactions are relevant to the message in front of them.

“Breaking the problem into many smaller models works much better than asking one model to write a good email,” Fyxer co-founder Archie Hollingsworth said in the case study. He also pointed to the reason email is deceptively hard for AI in the first place: “There’s something called Moravec’s paradox. Things that humans find easy are hard for computers, and things that computers find easy are hard for humans.” Two people can receive an identical email and need completely different replies depending on their relationship with the sender and what has already happened between them, and that context is exactly what a single, stateless prompt struggles to hold onto.

Another Fyxer team member, named in OpenAI’s post only as Shantsila, described the models’ role more plainly: OpenAI systems are used “for everything from digesting the email, so we can understand what it is actually about, to pulling in and re-ranking the context we want to include, to the actual email generation.”

The Fine-Tuning Stack Underneath

Fyxer trains its specialized models with supervised fine-tuning and Low-Rank Adaptation (LoRA), and it recently worked with OpenAI’s managed fine-tuning team to move a new checkpoint into production. Once a model is live, the system keeps improving through a feedback loop: when a user edits a draft before sending it, the difference between the original and the final version becomes a labeled preference pair, which Fyxer feeds into Direct Preference Optimization (DPO) rather than manual labeling. Every change then goes through an A/B test, and Fyxer only ships it once the result is statistically significant, something the company’s user volume lets it reach in as little as a day.

The Real Moat Is a Decade of Grunt Work

The specialized-model architecture is not what makes Fyxer hard to copy. Access to OpenAI’s models, its fine-tuning tools, and even the same LoRA and DPO techniques Fyxer uses is available to any competitor with an API key. What is not available to competitors is Fyxer’s training data: more than 500,000 hours of annotated executive-assistant work, according to OpenAI.

That data did not come from scraping the web. It came from a business the Hollingsworth brothers built specifically to generate it. In a podcast interview with Madrona managing director Karan Mehandru, Fyxer CEO Richard Hollingsworth said the plan predates the current AI boom by years: “The intention behind starting the EA agency, which we did in 2016, was to use it as a platform to build the AI solution, but we recognized the technology wasn’t there at the time.” The agency staffed real human executive assistants for real clients, and every judgment those assistants made (when to answer quickly, when to wait, which past conversation mattered) became a labeled example. When GPT-3 arrived, Hollingsworth said, the founders judged the technology was finally ready, brought in a third co-founder (referred to in the interview only as Matt) as CTO, and built Fyxer on top of that archive.

Fyxer AI, the software company, was only founded in 2024, according to EU-Startups. But by the time it needed training data, it already had years of it sitting in an archive, collected under conditions no scraped dataset can replicate: real professionals, real inboxes, and a human assistant already trying to get the judgment calls right.

Retention, Not ARR, Is the Number the Founders Point To

Fyxer grew fast by any measure. OpenAI’s case study says the company went from $1 million to $32 million in annual recurring revenue over the course of 2025. But Hollingsworth used the case study to steer attention toward a different number: “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.”

That framing is worth taking seriously, not just as a talking point. Enterprise AI has a well-documented adoption problem, where pilots generate revenue commitments or usage spikes that evaporate once the novelty wears off. A tool people are still opening every day three months in is a much harder thing to fake than a growth chart. EU-Startups independently reported the same 90 percent figure back in September 2025, when Fyxer said it had grown from €1 million to €17 million in ARR and passed 180,000 users in just seven months, generating more than 15 million draft replies and half a million meeting notes along the way. An interim €17 million mid-year figure and a full-year $32 million figure from two different sources are consistent with continued growth through the rest of 2025, not a contradiction.

The Funding Behind the Growth

Fyxer raised a €25.5 million (about $30 million) Series B in September 2025, led by Madrona, with participation from Lakestar Capital. Madrona’s Karan Mehandru joined Fyxer’s board as part of the round. EU-Startups lists Fyxer’s previous investors as 20VC, 20Growth, and Salesforce founder Marc Benioff, whose personal stake sits in a company that automates inbox-and-calendar admin work, the same category Salesforce’s own Agentforce push targets.

London-based Fyxer has used funding rounds to help expand into the U.S. market, on top of the UK and European base OpenAI’s own case study still lists as the company’s home region.

A Case Study OpenAI Has Every Reason to Publish

OpenAI tags the Fyxer piece “Startup,” and it is one of a growing list of similar customer write-ups the company has published this year, each one doubling as a demonstration of its managed fine-tuning, supervised fine-tuning, and preference-optimization tooling for a specific, well-documented use case. That is not a knock on Fyxer’s results, which independent reporting backs up in multiple places. But a vendor’s own case study is also a sales document, published in a year when OpenAI, Anthropic, and Google are all competing hard to keep high-usage, fast-growing startups building on their own APIs. Google published its own training-data-strategy post, ToolGrad, a day earlier.

The lesson for other teams building narrow, high-stakes AI products has less to do with which foundation model to pick than with what Fyxer did before it touched one: collect real, labeled examples of the judgment calls the product would eventually need to make. Teams that want to try a smaller version of the “many specialized fine-tuned models” pattern Fyxer describes can start with this site’s own guide to fine-tuning a local LLM with QLoRA, or with the tool-calling fundamentals in building a local tool-calling agent. The harder part, the one most competitors will not have the patience or the pre-AI business model to replicate, is building a reason to collect the data in the first place.