Fyxer's AI assistant hits 90% retention using OpenAI models
OpenAI published a customer case study on Fyxer, an AI executive-assistant startup co-founded by Archie Hollingsworth, describing how the company built an email assistant that drafts replies in each user's own voice. Fyxer combines OpenAI's models with more than 500,000 hours of human executive-assistant workflow data collected before the company launched its AI product, when it ran a human-powered EA service. That dataset captures judgment calls a human assistant makes: when to answer quickly, when to wait, which earlier conversation matters, and how the same request can call for different replies to different people.
Rather than asking one model to write a whole email, Fyxer splits the job across 30 to 50 specialized models, each handling a narrow part of the workflow. When a new email arrives, a reply decision model first classifies it: does it need a response, a scheduling action, or is it purely informational? If a reply is needed, further models analyze intent and predict the likely outcome, such as whether the thread is heading toward scheduling a meeting or resolving a request. Separate retrieval models compare the new email against stored interactions to surface the memories relevant to that person and thread, since the system also has to decide which details should persist across conversations and which should not.
Fyxer uses supervised fine-tuning and Low-Rank Adaptation (LoRA) to produce task-specific model variants while controlling training cost, and has worked with OpenAI's managed fine-tuning team to put new checkpoints into production. Once deployed, the system keeps improving from real user feedback: when someone edits a draft before sending it, the difference between the original and the edited version becomes training data through Direct Preference Optimization (DPO), so the model learns from pairs of outputs rather than manually labeled examples. Every drafting change then goes through an A/B test, and Fyxer ships a new version only when it produces a statistically significant improvement; the company's user volume means that threshold can sometimes be reached within a day.
The case study reports that 53% of Fyxer's AI-generated drafts are now accepted as written, which OpenAI's writeup frames as evidence the system is correctly predicting intent and tone for a large share of real conversations. Fyxer's annual recurring revenue grew from $1 million to $32 million over 2025. Hollingsworth points to a different number as the stronger signal: over 90% of Fyxer's users are still paying customers at the 90-day mark and still using the product daily. Looking ahead, Fyxer says it is building a richer understanding of relationships, preferences, and ongoing work threads, aiming to extend the assistant beyond drafting email replies toward managing more of a user's communication and coordination workload.
Key facts
- Over 90% of Fyxer's users are still paying customers and using the product daily at the 90-day mark.
- 53% of Fyxer's AI-generated email drafts are accepted as written by users.
- The system is built from 30 to 50 specialized models, each handling a narrow part of the email workflow, trained on 500,000+ hours of human executive-assistant data.
- Fyxer's annual recurring revenue grew from $1 million to $32 million over the course of 2025.
- User edits to drafts feed Direct Preference Optimization (DPO), and every change goes through an A/B test before shipping.
Why it matters
Fyxer is a concrete example of a highly contextual AI product built not as one large model call but as a pipeline of dozens of narrow, specialized models plus a large proprietary dataset of human task execution. Co-founder Archie Hollingsworth frames the design choice around Moravec's paradox: tasks that are easy for humans, like judging what a specific relationship calls for in a reply, are hard for a single general model, so Fyxer decomposes the problem into many smaller predictions instead.
Who it affects
The case study is aimed at founders building AI products that depend on deep, personal context rather than generic text generation, and at OpenAI's enterprise customers evaluating fine-tuning and managed model support. It also describes a product used directly by professionals who rely on an AI assistant to triage and draft email on their behalf.
How to use it
Fyxer's pipeline runs a reply-decision model first to classify whether an incoming email needs a response, a scheduling action, or is purely informational. If a reply is needed, further models analyze intent and predict the likely outcome, retrieval models pull in the memories relevant to that sender and thread, and a drafting step produces the reply. Users can then edit the draft before sending, and those edits both improve the assistant and, through DPO, become training data for the next version.
How solid is it
The evidence in the case study is self-reported: the 53% draft-acceptance rate, the 90% 90-day retention figure, and the ARR growth from $1 million to $32 million all come from Fyxer, relayed through an OpenAI-published customer story rather than independently verified numbers. The stated methodology, supervised fine-tuning plus LoRA plus DPO with A/B-gated rollouts, is a coherent and fairly standard approach, but the piece does not name which OpenAI model version or versions power the system, referring only to 'the latest OpenAI models' and 'frontier OpenAI models'.
Risks and caveats
This is a promotional case study published on OpenAI's own site to showcase a customer, not an independent audit, so the figures should be read as Fyxer's own claims about its own product. The text gives no information on Fyxer's pricing, funding, headcount, or total user count, and does not date when the company was founded or when this case study was produced.
“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.”
— Archie Hollingsworth, Fyxer co-founder