Summary: Fyxer splits email assistance across 30 to 50 specialized models and trains them on more than 500,000 hours of annotated assistant workflows, using user edits and DPO as a feedback loop. The company reports that 53% of generated email drafts are accepted exactly as written, with 2025 ARR growing from $1 million to $32 million and over 90% of users still paying at day 90. Every edit is A/B tested before a model version ships.
Fyxer has built an AI executive assistant that people genuinely rely on: 53% of the email drafts it generates go out exactly as written. That figure carries weight because email only looks easy. The same note can call for a completely different answer, depending on the relationship, what happened before, and what each side is trying to achieve. What Fyxer’s approach demonstrates is what it takes for an AI system to come across as an assistant that already understands what matters.
Why one big model isn’t enough
Rather than treat email as a single text-generation job, Fyxer splits it across 30–50 specialized models, each handling one narrow piece. A reply-decision model works out whether an incoming message calls for a response, a scheduling action, or simply needs to be visible. Intent and outcome models forecast where the thread is heading. Retrieval models fetch the memories that matter from past interactions.
“Breaking the problem into many smaller models works much better than asking one model to write a good email,” explains Fyxer co-founder Archie Hollingsworth. It is the same pattern we’ve explored before: execution models decide how much your orchestrator can be trusted. Once every step is a narrow prediction, you can measure it and sharpen it on its own.
Training data from real assistant work
Fyxer didn’t begin from nothing. Years before it shipped the AI product, the company ran an executive assistant service staffed by people. Out of that came a dataset of more than 500,000 hours of annotated workflows—concrete records of how professional assistants manage communication.
Embedded in those records are the small judgments that make a reply good: when to answer right away, when to hold back, which earlier exchange is the one that counts. Fyxer relies on supervised fine-tuning and Low-Rank Adaptation (LoRA) to build task-specific model variants without letting training cost run away. OpenAI’s fine-tuning platform and its managed fine-tuning team helped carry that customer-specific learning into the production models.
Turning user edits into a self-training loop
Whenever someone changes a draft before sending it, the gap between the original and the final version reveals which output they wanted. Fyxer turns those comparisons into training data with Direct Preference Optimization (DPO). The model learns from pairs of outputs: the draft as it was first written, and the version the user edited.
Each drafting change is then put through an A/B test. Fyxer ships a new version only if it delivers a statistically significant improvement. Given the company’s user volume, the team can sometimes clear that bar within a single day.
What the metrics actually tell you
In 2025 Fyxer grew from $1 million to $32 million in annual recurring revenue. Hollingsworth, though, points to a different signal: “Over 90% of our users are still paying at the 90-day mark with us, and still using us every day.” Retention outranks ARR because it shows the assistant is genuinely useful rather than merely sold.
A 53% draft acceptance rate is a solid baseline, and it also means nearly half of all drafts still come back edited. That isn’t a failure—it is the feedback loop doing its job. Every edit becomes training data for the next round.
Fyxer’s vision is to bring customers to a point where they never open their computer and simply trust the assistant to manage all of it. That point is still a long way off, but the route there is clear: split the problem into specialized models, train on real workflows, and let user behavior drive continuous improvement.