53% Accepted With Zero Edits! Inside Fyxer's "AI Assistant People Trust"

Hi, I'm Shiichan! Today I want to share a case study OpenAI just published about an AI executive assistant startup. It got me wondering: how much can you really trust an AI to write your emails for you? Let's find out!

OpenAI News openai.com

What was announced?

This is a case study from OpenAI's News blog about Fyxer, a startup building an AI executive assistant that follows your work as it moves across email, meetings, and messages. Fyxer combines the latest OpenAI models with more than 500,000 hours of executive-assistant workflow data to draft replies and decide what needs a response on your behalf.

Fyxer co-founder Archie Hollingsworth put it this way:

"Things that humans find easy are hard for computers, and things that computers find easy are hard for humans."

Two people can get the exact same email and need completely different replies, depending on the relationship and what's happened before. To get past that Moravec's-paradox problem, Fyxer is built to learn how each individual user works.

Why it matters

Fyxer didn't start as an AI company. It began by running a human-powered executive assistant service, and the real-world workflow data it collected along the way became the foundation for its AI product. Instead of just asking a large model to "write a good email," Fyxer trains its system on how genuinely skilled assistants make decisions, which is why its judgment about context lands so well.

That pattern, pairing real operational data with AI, is a useful blueprint for anyone building an assistant for a different line of work, too.

What changes

Fyxer users get their inbox triaged and reply drafts written for them. Deciding whether a reply is needed, reading tone and intent, and pulling in relevant context from past conversations all happen as one pipeline inside Fyxer's system.

The numbers back it up: 53% of Fyxer's AI-generated email drafts are accepted without any edits. On the business side, Fyxer grew its annual recurring revenue from $1 million to $32 million over the course of 2025.

Dive Deep

What's interesting about Fyxer's architecture is that it doesn't hand the whole email task to one giant model. It splits the work across 30 to 50 specialized smaller models, each responsible for one narrow job, like deciding whether a reply is needed, predicting intent, or generating the actual text.

On the training side, Fyxer combines supervised fine-tuning with LoRA (Low-Rank Adaptation) to produce task-specific model variants while keeping training costs down. Early on, the team used OpenAI's fine-tuning platform directly; more recently, they've worked with OpenAI's managed fine-tuning team to push new checkpoints into production.

The continuous-improvement loop is worth calling out too:

• When a user edits a draft before sending it, that diff becomes training data showing which output was preferred
• Fyxer applies DPO (Direct Preference Optimization), learning directly from pairs of the original draft and the user-edited version
• Every change goes through an A/B test, and only ships once it shows a statistically significant improvement

Thanks to Fyxer's user volume, that significance threshold can sometimes be reached in a single day.

Wrap-up

• Fyxer is an AI executive assistant built on OpenAI models plus more than 500,000 hours of assistant workflow data
• It splits email handling across 30-50 specialized models, improving continuously through LoRA fine-tuning and DPO
• 53% of its AI-generated emails are accepted with no edits, and its 2025 ARR grew from $1 million to $32 million
• A great read for developers and product folks curious about building on real operational data, and about turning user edits into training signal!