# How Fyxer's AI Assistant Learns from Real Workflows to Handle Complex Tasks

Canonical URL: https://zero2vibecode.com/blog/fyxer-ai-executive-assistant-trust
Date: 2026-09-15
Tags: agents, models, tools, beginner

Fyxer built an AI executive assistant that learns from real workflows to handle complex tasks like email management, offering lessons for building contextual AI products.

For professionals juggling multiple tasks across emails, meetings, and apps, keeping track of commitments can be overwhelming. Fyxer’s AI executive assistant addresses this by integrating with tools and workflows, learning from real user interactions to provide context-aware support.

<Cover src="/blog/fyxer-ai-executive-assistant-trust.jpg" alt="A modern office desk with a laptop, a coffee cup, and a neatly organized notebook" />

## Breaking Down Complex Tasks into Smaller Models

Fyxer’s system relies on specialized models, each handling a narrow part of the workflow. Instead of treating email as a single task, Fyxer breaks it into smaller predictions:

| Task                          | Model Responsibility                          |
|-------------------------------|-----------------------------------------------|
| Email Classification          | Decides if a reply is needed                 |
| Intent Analysis               | Predicts the likely outcome of the interaction |
| Memory Retrieval              | Surfaces relevant context from past exchanges |
| Draft Generation              | Creates responses matching tone and context  |

This approach ensures accuracy and efficiency, as each model focuses on a specific aspect of the task.

## Training on Real Executive Assistant Workflows

Before launching its AI product, Fyxer operated a human-powered executive assistant service, accumulating over 500,000 hours of annotated workflows. This dataset captures the nuances of professional communication, such as:

- When to respond quickly vs. when to wait
- Which previous conversations matter
- How to tailor responses based on the recipient

Fyxer uses supervised fine-tuning and Low-Rank Adaptation (LoRA) to create task-specific model variants, ensuring the AI learns from real-world examples.

<Callout type="tip">
Training AI on real workflows helps it handle subjective tasks like tone and intent, making it more effective in professional settings.
</Callout>

## Turning User Feedback into Continuous Improvement

Fyxer’s system improves through real user feedback. When a user edits a draft, the difference between the original and final email becomes training data. Fyxer uses Direct Preference Optimization (DPO) to learn from these pairs without manual labeling.

Key metrics include:

| Metric                       | Value                                        |
|------------------------------|----------------------------------------------|
| Draft Acceptance Rate        | 53% of drafts accepted as written           |
| User Retention               | Over 90% at the 90-day mark                 |

This feedback loop allows Fyxer to refine its models continuously, ensuring they align with user preferences.

## Building a Proactive Assistant

Fyxer’s vision extends beyond drafting replies to managing more of a user’s communication and coordination workload. The goal is to create an assistant that understands relationships, preferences, and ongoing work threads, allowing users to focus on tasks they love.

<Callout type="note">
Fyxer’s approach highlights the importance of breaking complex tasks into smaller models, training on real workflows, and using user feedback for continuous improvement.
</Callout>

## Frequently Asked Questions

<Faq>
<FaqItem q="How does Fyxer handle different tones and intents in emails?">

Fyxer uses specialized models to analyze the intent of an email and predict the likely outcome of the interaction. These models consider factors like the recipient’s relationship with the sender and the context of previous conversations.

</FaqItem>

<FaqItem q="What makes Fyxer’s training data unique?">

Fyxer’s training data comes from over 500,000 hours of annotated executive workflows, capturing the small judgments behind effective communication. This ensures the AI learns from real-world examples rather than generic datasets.

</FaqItem>

<FaqItem q="How does Fyxer ensure continuous improvement?">

Fyxer uses Direct Preference Optimization (DPO) to learn from user feedback. When a user edits a draft, the difference between the original and final email becomes training data, allowing the system to refine its outputs over time.

</FaqItem>
</Faq>

## Lessons for Building Contextual AI Products

Fyxer’s approach offers three key lessons for founders building highly contextual AI products:

1. **Break complex tasks into smaller models**: Specialized models handle narrow aspects of a task, improving accuracy and efficiency.
2. **Train on real workflows**: Real-world examples ensure the AI understands subjective tasks like tone and intent.
3. **Use user feedback for continuous improvement**: Feedback loops allow the system to refine its outputs based on user preferences.

## Read next

- [How to Evaluate and Improve AI Coding Agents Effectively](/blog/evaluate-improve-ai-coding-agents)
- [How GPT-6 Astra Helps Devin Test Its Own Code](/blog/gpt6-astra-devin-testing)

Want to try all of this hands-on? Start with the free [Claude Code from Zero](/learn/claude-code) course.

<Callout type="note" title="Source">
Based on OpenAI’s announcement, "How Fyxer built an AI executive assistant people trust". Written for people learning to build with these tools.
</Callout>