Fyxer uses OpenAI models to build a contextual executive assistant

Fyxer reports strong acceptance for AI email drafts

Fyxer, a European and UK startup, says 53% of its AI-generated email drafts are accepted by users without changes, while more than 90% of users are still paying and using the service after 90 days. The company is building an AI executive assistant with OpenAI models, more than 500,000 hours of executive-assistant workflows, and feedback from real users.

The product is intended to follow work across inboxes, meetings, messages and applications. In email, that requires more than producing fluent text: a useful reply depends on the relationship, prior exchanges and the user’s intended outcome. Fyxer says it learns how an individual works so that its drafts reflect the context that matters to that person.

Specialised models divide the email workflow

Rather than treating email as one generation task, Fyxer uses a system of 30 to 50 specialised models. Each model handles a narrow prediction or operation within the workflow, including deciding whether a message needs a reply, scheduling action or simple visibility.

When a response is needed, further models assess the email’s intent and the likely direction of the interaction. They can identify whether a thread is heading towards a meeting, resolving a request or continuing a longer relationship. Retrieval models compare new messages with stored interactions and surface the memories considered most relevant to the person and conversation.

OpenAI models are used across the pipeline, from digesting an email and understanding its subject to retrieving and re-ranking context and generating the final draft. Fyxer selected OpenAI after internal benchmarking, citing model performance, fine-tuning capabilities for subjective tasks such as tone and intent, and engineering collaboration.

Workflow data and feedback shape model variants

Before launching its AI product, Fyxer operated a human-powered executive-assistant service. It says this produced a dataset from more than 500,000 hours of annotated workflows, capturing decisions such as when to respond, when to wait, what history is relevant and how responses differ between users.

The company uses supervised fine-tuning and Low-Rank Adaptation, or LoRA, to create task-specific variants while managing training cost. It initially used OpenAI’s fine-tuning platform for high-accuracy tasks and later worked with OpenAI’s managed fine-tuning team to put a new checkpoint into production. Models are evaluated on Fyxer validation sets for drafting, classification and prioritisation, with accuracy, latency and cost all considered.

Edited drafts become preference signals

Once deployed, the system uses differences between an original draft and a user-edited email as feedback. Fyxer converts these pairs into training data through Direct Preference Optimization, allowing the system to learn which version users preferred without manually labelling every example.

Every drafting modification is A/B tested, and Fyxer says a version is released only after it delivers a statistically significant improvement. The company grew annual recurring revenue from $1 million to $32 million during 2025, while positioning retention as a more meaningful measure of daily product value.

For businesses deploying contextual AI, Fyxer’s approach underscores the practical importance of decomposing complex work, testing against the organisation’s own tasks, and treating user corrections as evidence for controlled iteration.