Models & tools · 14 Sep 2026 · 14:00 CEST
How Fyxer built an AI executive assistant people trust
Story brief · OZZZER analysis pending editorial review.
Fyxer pairs OpenAI models with 500,000+ hours of EA workflows and real user feedback to draft replies in each person’s voice. For many professionals, work means keeping track of conversations and commitments across inboxes, meetings, messages, and apps. Without rock-solid context, those commitments can fall through the cracks, damaging projects and relationships. Fyxer built an AI executive assistant that follows the thread as work moves across tools.
It combines the latest OpenAI models with more than 500,000 hours of executive assistant workflows, dividing the work among dozens of specialized models that improve through real user feedback. Email is one of the clearest places to see it in action. Two people can receive the same email and need completely different replies depending on the relationship, what has happened before, and what each person is trying to get done.
That makes a seemingly simple task deceptively hard for AI. “There’s something called Moravec’s paradox,” explains Fyxer Co-founder Archie Hollingsworth. “Things that humans find easy are hard for computers, and things that computers find easy are hard for humans.” Fyxer handles that complexity…
Excerpt supplied by the publisher.
ADDITIONAL STORY CONTEXT
The story in brief
Fyxer is building an assistant for the routine communication that can quietly consume a working day: deciding which emails need attention, recalling the right context and drafting a reply that sounds like its user. The product draws on OpenAI models, a record of human executive assistant work and the corrections customers make to suggested drafts.
OpenAI presents the company as an example of a more specialized approach to AI assistance.
A series of decisions, not one prompt
An incoming message first has to be classified. It may call for a reply, a scheduling step or no action at all. If a response is needed, other parts of Fyxer’s system work out the message’s intent, retrieve relevant past interactions and shape a draft. The company says it uses roughly 30 to 50 specialized models across this workflow.
This design matters because a useful reply depends on the relationship and previous commitments, not just the words in the newest email.
Training from real assistant work
Before its AI product, Fyxer operated a service staffed by human executive assistants. That work produced more than 500,000 hours of annotated workflows, according to OpenAI. Fyxer uses examples from those workflows to teach narrow tasks, including prioritization and drafting. It also evaluates models against its own email cases and weighs quality against latency and cost.
The article describes fine-tuning, including low-rank adaptation, as part of this task-specific approach.
The user remains the feedback loop
When a person changes a suggested reply, Fyxer can compare the draft with the version that was sent. Those differences become preference data for future improvements. The company says it tests changes before release and currently sees 53% of generated drafts accepted without edits. OpenAI also reports that more than 90% of users are still paying after 90 days.
These are company figures in a publisher case study, rather than independent measures of performance.
The broader direction is an assistant that can keep track of relationships and unfinished work across more than an inbox. For teams considering similar tools, the useful lesson is the workflow: separate the decisions, preserve relevant context and measure whether real users accept the result. The full OpenAI article gives the technical and business detail behind Fyxer’s account.
Source
OpenAI · 14 Sep 2026 · 14:00 CEST
Open the original at OpenAI ↗