Fyxer: 53% of Drafts Accepted, ARR at $32M

Fyxer highlights an AI executive assistant focused on email that would be accepted without modification in more than half of the cases. The company attributes these performances to a multi-model architecture, data derived from 500,000 hours of human work, and a preference learning loop. It also reports a jump to $32 million in ARR by 2025.

Adoption Indicators: 53% of Emails Sent as Is

Fyxer states that 53% of drafts generated by its AI are accepted as is, meaning they are sent without modification by users. This rate is presented as a sign that the system is able to correctly anticipate the intent and tone expected in many professional conversations. By 2025, the company reports an increase in annual recurring revenue from $1 million to $32 million.

A Stack of 30 to 50 Models to Predict Rather Than Generate in One Go

Fyxer's system relies on 30 to 50 specialized models, each responsible for a specific step in the email processing. Rather than treating email as a single text generation task, Fyxer segments it into a series of predictions. When a message arrives, one model decides whether it requires a response, a scheduling action, or simply information gathering. If a response is needed, other models analyze the intent and predict the likely outcome of the exchange. Fyxer presents this architecture as one of three lessons for designing highly contextual AI products.

Data from 500,000 Hours and a DPO Loop in Production

Before launching its AI assistant, Fyxer operated a human executive assistant service for several years, accumulating a corpus of over 500,000 hours of annotated workflow. These examples were directly used to train the system. After deployment, the solution continues to improve through user feedback: the differences between a generated draft and the final sent email serve as preference comparisons. Fyxer transforms these signals into training data via Direct Preference Optimization (DPO).

Fyxer Selects OpenAI Models to Generate Emails

OpenAI provides the models used to analyze emails and create drafts. Fyxer selected these models after conducting its own benchmarks, highlighting their ability to adapt to subjective tasks such as managing tone and intent, as well as their technical support during collaborative sessions. The assistant manages the tracking of conversations and commitments through various tools, relying on several dozen specialized models and collected data. Fyxer is working to enhance the understanding of relationships, preferences, and workflows, aiming to go beyond simple response drafting towards greater coordination. Archie Hollingsworth, co-founder, notes that some tasks that are easy for humans remain challenging for machines and describes an approach that learns from individual usage. He expresses the desire to enable clients to focus on the work they truly enjoy and to eventually entrust the management of exchanges to the assistant.