How Fyxer built an AI executive assistant people trust

Fyxer combines OpenAI models with more than 500,000 hours of executive assistant workflows and real user feedback to draft email replies tailored to each person’s voice and context. The system tracks conversations across tools to ensure commitments are met, addressing the challenge of maintaining context in professional communication. By dividing tasks among dozens of specialized models, Fyxer improves accuracy and adaptability, handling nuances like tone and intent that vary by user and relationship.

Fyxer’s system uses OpenAI models for core functions such as understanding emails, retrieving relevant context, and generating drafts. The company selected OpenAI after internal benchmarks showed superior performance in fine-tuning for subjective tasks like tone and intent, alongside engineering support. The system relies on 30–50 specialized models, each handling a specific part of the email workflow, such as classifying messages or predicting interaction outcomes.

Before launching its AI product, Fyxer operated a human-powered executive assistant service, accumulating a dataset from over 500,000 hours of annotated workflows. This data informed how real assistants manage communication, capturing decisions like when to respond quickly or wait. Fyxer employs supervised fine-tuning and Low-Rank Adaptation (LoRA) to create task-specific model variants while controlling costs, using OpenAI’s fine-tuning platform for high-accuracy tasks.

Fyxer evaluates models on its own email tasks, including drafting, classification, and prioritization, balancing accuracy, response time, and cost. The system improves through real user feedback, converting edits into training data via Direct Preference Optimization (DPO). In 2025, Fyxer’s annual recurring revenue grew from $1 million to $32 million, with 53% of AI-generated drafts accepted as written, indicating strong alignment with user intent and tone.