Fyxer’s AI assistant is built from half a million hours of human workflow

Fyxer has built an AI executive assistant around more than 500,000 hours of annotated human-assistant workflows, using OpenAI models alongside dozens of smaller, specialised systems. The useful idea is not simply “AI writes your email”, but a pipeline that decides what needs a reply, retrieves the right memories and matches the user’s intent and tone.

OpenAI Watch analysis

What happened

In an OpenAI customer story, Fyxer says its system uses roughly 30–50 specialised models for tasks such as reply classification, intent prediction, context retrieval and drafting. It fine-tunes task-specific variants with supervised learning and LoRA, then turns user edits into preference data through Direct Preference Optimisation.

Fyxer says 53% of its AI-generated drafts are accepted as written, while more than 90% of users remain paying customers after 90 days. OpenAI’s account also says Fyxer grew from $1m to $32m in annual recurring revenue during 2025. Those are company-reported figures, not an independent evaluation.

Key findings

• Many small jobs
  Fyxer splits email work across specialised models instead of asking one model to do everything.
• Human workflow data
  More than 500,000 hours of assistant work provide examples of timing, tone, prioritisation and context.
• Edits become training data
  Changes users make to drafts create preferred-versus-original examples for further optimisation.
• Context is the product
  The system stores and retrieves relationship and thread details, not just the latest message.
• A measurable acceptance signal
  Fyxer says 53% of generated drafts are sent without edits.

Why it matters

The interesting lesson is architectural. Reliable workplace assistance is less about producing eloquent sentences than deciding whether a message needs action, which past details matter and what the user would actually mean.

That also explains why a seemingly modest email assistant can require a small committee of models. The system is trying to reproduce the quiet judgement of an experienced assistant, where “reply now”, “schedule this” and “leave it alone” are very different outcomes.

Our read

Fyxer’s approach is more convincing than the usual one-model-does-everything pitch because it maps the work into testable decisions and lets real edits feed back into development. The catch is that the most important numbers come from Fyxer itself, and acceptance rate is not the same as correctness. A beautifully toned reply can still be the wrong reply.

For builders, the practical takeaway is clear: collect task-specific examples, separate the workflow into decisions you can evaluate and measure whether users trust the result enough to leave it alone.

What to watch

• Whether Fyxer publishes independent evaluations of accuracy, privacy and unwanted automation.
• How its memory system handles outdated, conflicting or sensitive context.
• Whether acceptance rates hold as the assistant takes on scheduling and coordination tasks.
• Whether the specialised-model approach remains cheaper and faster than a larger general model.

Discussion spark: Would you trust an email assistant more if it used many narrow, inspectable models, or would the added complexity make it harder to audit?

OpenAI Watch is independently operated by WittyWires. It is not affiliated with, endorsed by, or operated by OpenAI.