Fyxer has developed an AI executive assistant that drafts email replies in each user’s voice by combining OpenAI models with more than 500,000 hours of annotated executive-assistant workflows and real user edits. The company divides email handling across dozens of specialized models and improves performance through supervised fine-tuning and preference learning from actual user behavior.
The problem they address
Professionals must track conversations and commitments across inboxes, meetings, messages, and apps. When context is missing, commitments can fall through the cracks, harming projects and relationships. Even seemingly simple tasks like replying to an email are difficult for AI because the same message may require completely different responses depending on the relationship, prior history, and objectives.
Fyxer’s approach
Rather than treating email as one generic text-generation task, Fyxer splits the workflow into roughly 30–50 specialized models, each responsible for a narrow prediction. One model classifies whether an incoming message needs a reply, scheduling, or just awareness. Other models analyze intent and predict likely outcomes — for example, whether the thread is moving toward scheduling, resolving a request, or continuing a longer relationship.
Memory handling is central: the system decides which details should persist across threads and which should expire after a single exchange. Retrieval models compare an incoming message with stored interactions and surface the most relevant memories for that person and thread.
Technology and model training
OpenAI models are used across Fyxer’s pipeline — for digesting email content, ranking and re-ranking contextual candidates, and generating drafts. Fyxer chose OpenAI because those models performed best on its internal benchmarks, offered robust fine-tuning options for subjective tasks like tone and intent, and provided hands-on engineering collaboration.
To create task-specific model variants while managing training cost, Fyxer used supervised fine-tuning and Low-Rank Adaptation (LoRA). Early in development they used OpenAI’s fine-tuning platform for high-accuracy tasks, and later worked with OpenAI’s managed fine-tuning team to deploy new checkpoints.
The training data
Before building the AI product, Fyxer operated a human-powered executive assistant service for years. From that service they accumulated more than 500,000 hours of annotated executive workflows that capture the small judgments behind effective responses: when to reply quickly, when to wait, which prior conversations matter, and how similar requests call for different answers from different people.
Learning from real edits
When users edit a draft before sending, the difference between the original and the final email becomes training data. Fyxer converts these draft–final pairs into learning signals using Direct Preference Optimization (DPO), so the model learns from user preferences rather than from manual labels. Every modification then goes through A/B testing; the company ships a new model version only after it produces a statistically significant improvement. With sufficient user volume, that threshold is sometimes reached within a day.
Results
Today, 53% of Fyxer’s AI-generated drafts are accepted as written, indicating the system correctly predicts intent and tone for a substantial share of real conversations. Financially, in 2025 Fyxer’s annual recurring revenue grew from $1 million to $32 million.
Co-founder Archie Hollingsworth emphasizes retention as a key signal: many Fyxer customers are not deeply technical, and for some this is the first AI system they use daily.
Next steps
Fyxer is expanding the assistant’s understanding of relationships, preferences, and ongoing work threads so it can manage a larger portion of a user’s communication and coordination workload. The long-term vision is to enable customers to focus on the work they love while trusting Fyxer to handle the rest.
Lessons for founders
Fyxer’s experience suggests three practical principles for builders of highly contextual AI products: decompose complex tasks into many smaller, specialized models; train on real, task-specific examples and continuous user feedback; and balance accuracy with latency and cost for each job type.
