Fyxer built its AI executive assistant on 30-50 specialized OpenAI models

Fyxer, a startup that builds an AI executive assistant for email and scheduling, says its system now accepts 53% of AI-generated email drafts as written and has grown annual recurring revenue from $1 million to $32 million in 2025 — a case study OpenAI is using to show how a startup built durable trust into an AI product handling sensitive daily communication.

What's new

Fyxer's assistant combines OpenAI's frontier models with more than 500,000 hours of annotated human executive-assistant workflow data, splitting the work across 30-50 specialized models rather than one general model handling email end to end. Cofounder Archie Hollingsworth frames the design choice around Moravec's paradox — the observation that tasks easy for humans, like judging tone and relationship context, are hard for computers, while tasks hard for humans are easy for computers.

The pipeline OpenAI describes: a reply-decision model first classifies whether an incoming email needs a response, a scheduling action, or is just informational; if a reply is needed, additional models predict the likely direction of the exchange (scheduling, resolving a request, an ongoing relationship thread); and retrieval models pull the specific prior context relevant to that sender. Fyxer uses supervised fine-tuning and LoRA to produce these task-specific model variants, and has moved from OpenAI's self-serve fine-tuning platform to working directly with OpenAI's managed fine-tuning team to ship new checkpoints.

Feedback closes the loop: when a user edits a draft before sending, the diff between the model's draft and the final email becomes training data via Direct Preference Optimization, and every change to the drafting model ships only after it clears a statistically significant A/B test — a bar Fyxer says its volume lets it clear within a day. On retention, Hollingsworth says: "Over 90% of our users are still paying at the 90-day mark with us, and still using us every day."

Context

This is the second OpenAI-published customer story in short succession describing a company building durable, semi-autonomous AI systems on top of OpenAI models rather than a single chat wrapper — OpenAI separately detailed how Perplexity now trusts GPT-6 Astra with full end-to-end production systems. Both stories emphasize reduced human oversight and system-level trust rather than raw model benchmarks.

Why it matters

Fyxer's numbers double as an answer to a common objection to AI assistants handling business-critical communication: draft-acceptance rate and 90-day retention are harder to fake than a benchmark score, and they show a way to earn trust with continuous, mundane use rather than a single impressive demo. For OpenAI, publishing detailed architecture case studies like this is part of how it markets its models and fine-tuning tooling to technical buyers building vertical AI products, not just to end users of ChatGPT.