The News, Explained

In OpenAI’s Fyxer case study, the email AI assistant first decides whether a reply is needed, then handles tasks such as intent analysis, relevant-memory retrieval, and drafting. Fyxer uses 30 to 50 specialized models across these stages. Source

Classifying the message first

When a new email arrives, Fyxer first classifies whether it needs a reply, a scheduling action, or simply a notification to the user. If a reply is needed, additional models analyze the sender’s intent and the expected outcome of the interaction. Draft generation follows those steps. Source

Retrieving relevant information from earlier conversations

Fyxer decides what information from each conversation should be retained long term and searches stored interactions for memories related to the current user and conversation. The drafting stage can therefore use context from earlier exchanges alongside the current message. Source

Training and evaluating the system

Before building the AI product, Fyxer operated a human executive-assistant service and accumulated more than 500,000 hours of annotated workflow data. The data included examples of when to respond, which prior conversations were relevant, and how replies differed by user. Source

Fyxer said it uses supervised fine-tuning and LoRA to create task-specific model variants and uses OpenAI’s fine-tuning platform for tasks where accuracy is especially important. Supervised fine-tuning trains a model further on labeled examples, while LoRA adapts a model by changing a relatively small set of parameters instead of retraining the whole model. Before deployment, it evaluates drafting, classification, and prioritization with its own email task validation set and considers accuracy, latency, and cost. A validation set is held-out test data used to measure performance separately from training. Source

Learning from user edits

When a user edits an AI draft and sends the final email, Fyxer uses the difference between the original and final versions as preference data. It creates DPO training data from those comparisons and A/B tests draft changes, releasing them when the test shows a statistically significant improvement. DPO trains response preferences from pairs of preferred and less-preferred outputs, while A/B testing compares two versions under similar real-world conditions. Source

Published usage metrics

OpenAI’s case study says 53% of Fyxer’s AI-generated drafts were accepted without edits. A co-founder also said more than 90% of users were still paid customers and using the product daily 90 days after signup. These figures are the results reported in the Fyxer and OpenAI case study. Source

OYOPICK’s Take

For email AI, reply selection and retrieved context deserve attention before prose quality: fluent wording cannot rescue a reply to the wrong message. The reported 53% is draft acceptance without edits. It does not establish classification accuracy across all email or a time-saving rate for all users.

For an adoption comparison, we suggest recording reply selection, relevant-context retrieval and tone edits separately on comparable work messages. Because this case describes learning from edits, storage and use of messages and feedback are also questions to check in the contract and official guidance. We hope assistants improve at understanding the commitments behind a conversation as well as its voice.