对于许多专业人士而言，工作意味着要在收件箱、会议、消息和应用中跟踪对话与承诺。如果没有坚实可靠的上下文，这些承诺就可能被遗漏，从而损害项目和人际关系。
Fyxer 打造了一款 AI 行政助理，当工作在不同工具之间流转时，它能持续跟进整个脉络。它将最新的 OpenAI 模型与超过 500,000 小时的行政助理工作流程相结合，把工作分配给数十个专用模型，并通过真实用户反馈不断改进。
电子邮件是观察其实际效果最直观的场景之一。两个人可能收到同一封邮件，却需要根据彼此关系、过往发生的事以及各自想要完成的目标，给出完全不同的回复。这使得一项看似简单的任务对 AI 来说异常困难。
“有一种现象叫莫拉维克悖论，”Fyxer 联合创始人 Archie Hollingsworth 解释道。“人类觉得容易的事情，对计算机来说很难；而计算机觉得容易的事情，对人类来说很难。”Fyxer 通过学习每位用户的工作方式来处理这种复杂性，因此它能像一位早已知道什么重要的助理那样作出回应。
前沿 OpenAI 模型支撑着这一体验的关键环节，从理解电子邮件、找到正确的上下文，到生成草稿。Fyxer 选择 OpenAI，是因为其模型在 Fyxer 的内部基准测试中表现最佳，为语气和意图等主观任务提供了强大的微调能力，并通过白板讨论会和技术协作提供了实操性的工程支持。
“我们选择 OpenAI，是因为他们拥有最好的模型，而且给了我们真正的访问权限和紧密的合作关系。我可以在 Slack 里丢一个问题过去，很快就能得到答复；当我们遇到问题时，团队会来我们办公室，和我们一起把它解决。他们真的会到场。”
——Archie Hollingsworth，Fyxer 联合创始人
Fyxer 的方法为构建高度依赖上下文的 AI 产品的创始人提供了三条经验：
1. 将电子邮件拆解为更小的任务
Fyxer 围绕 30–50 个专用模型构建其系统，每个模型负责电子邮件工作流中一个很窄的部分。Fyxer 没有把电子邮件当作单一的文本生成任务，而是把问题拆解为一个预测系统，例如判断一封邮件是否需要回复，或起草与用户语气和上下文相匹配的回复。
“把问题拆成许多更小的模型，比让一个模型写出一封好邮件要有效得多，”Hollingsworth 解释道。
当新邮件到达时，一个回复决策模型会对邮件进行分类：这是需要回复的内容、需要安排日程的动作，还是用户只需了解的信息？
如果需要回复，其他模型会分析邮件的意图，并预测这次互动的可能结果。这些模型会判断诸如对话是否正朝着安排会议、解决请求或延续更长期关系脉络的方向发展等模式。
记忆是系统最重要的部分之一。Fyxer 需要决定哪些细节应在多次对话中持续保留，哪些应在一次交流后消失。当新邮件到达时，检索模型会将其与已存储的互动进行比较，并浮现出与该联系人和该邮件线程最相关的记忆。
OpenAI 模型为 Fyxer 系统中的多个步骤提供支持。“我们使用 OpenAI 模型来完成从消化邮件、理解它究竟在讲什么，到拉取并重新排序我们想纳入的上下文，再到实际生成邮件的一切工作，”Shantsila 说。
2. 基于优秀助理的真实工作方式进行训练
在推出 AI 产品之前，Fyxer 花了多年时间运营一项由人工提供支持的行政助理服务。随着时间推移，团队积累了一个由超过 500,000 小时标注过的行政工作流程构成的数据集，记录了真实助理如何管理职业沟通。
这些示例为 Fyxer 提供了源自工作本身的训练数据。它们捕捉了一封好的回复背后的细小判断：何时快速回复，何时等待，哪段早先的对话更重要，以及同一个请求为何对一个人和对另一个人需要不同回应。
Fyxer 在其更广泛的系统中使用监督微调和低秩适配（LoRA），以创建任务特定的模型变体，同时控制训练成本。在产品开发早期，团队使用 OpenAI 的微调平台来处理需要高准确率的任务。最近，Fyxer 与 OpenAI 的托管微调团队合作，将一个新的检查点投入生产。
“OpenAI 对我们至关重要，帮助我们迁移关于客户的学习成果，并成功将其融入模型的工作方式中。”
——Joey Dwonczyk，Fyxer AI/ML 产品工程师
在任何模型部署之前，Fyxer 都会围绕其自身电子邮件任务构建的验证集对其进行评估，包括起草、分类和优先级排序。团队会同时权衡准确率、响应时间和成本，因为最佳选择可能因任务而异。
3. 将用户反馈转化为自我训练循环
部署之后，Fyxer 的系统会通过真实用户反馈持续改进。当有人在发送前编辑草稿时，原始版本与最终邮件之间的差异就表明了用户更偏好哪种输出。
Fyxer 使用直接偏好优化（DPO）将这些比较转化为训练数据。模型不是手动标注每一个示例，而是从成对的输出中学习：原始草稿和用户编辑后的版本。
随后，每一次起草变更都会经过 A/B 测试。只有当新版本带来统计上显著的改进时，Fyxer 才会发布它。其用户规模意味着团队有时可以在一天内达到这一阈值。
如今，Fyxer 生成的 AI 草稿中有 53% 按原文被接受。这意味着该系统在很大一部分真实对话中正确预测了意图和语气。仅在 2025 年，Fyxer 的年度经常性收入就从 100 万美元增长到 3200 万美元。
对 Hollingsworth 来说，更强的信号是留存率。许多 Fyxer 客户并不是技术深度用户，对其中一些人来说，Fyxer 是他们每天使用的第一个 AI 系统。
“大家都在谈 ARR，但我认为留存率才是真正的实力体现。超过 90% 的用户在 90 天时仍在付费，并且仍然每天使用我们。”
——Archie Hollingsworth，Fyxer 联合创始人
从草稿到主动型助理
展望未来，Fyxer 正在构建对关系、偏好和持续进行的工作线程更丰富的理解。他们的助理正持续演进，不再局限于起草回复，而是朝着更广泛的 AI 助理发展，能够管理用户更多的沟通与协调工作负载。
“我们的愿景是让客户尽可能多地去做他们真正热爱的工作，”Hollingsworth 说。“我们想把他们带到一个甚至不必打开电脑的状态，并让他们相信 Fyxer 能管理所有这些事情。”
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 by learning how each user works, so it can respond like an assistant who already knows what matters.
Frontier OpenAI models support key parts of the experience, from understanding the email and finding the right context to generating the draft. Fyxer chose OpenAI because its models performed best on Fyxer’s internal benchmarks, offered strong fine-tuning capabilities for subjective tasks like tone and intent, and provided hands-on engineering support through whiteboarding sessions and technical collaboration.
“We chose OpenAI because they have the best models, and they’ve given us real access and a close working relationship. I can drop a question in Slack and get an answer quickly, and when we face a problem, the team comes to our office and works through it with us. They show up.”
—Archie Hollingsworth, Co-founder, Fyxer
Fyxer’s approach offers three lessons for founders building highly contextual AI products:
1. Break email into smaller jobs
Fyxer built its system around 30–50 specialized models, each responsible for a narrow part of the email workflow. Instead of treating email as a single text-generation task, Fyxer breaks the problem into a system of predictions, like deciding whether a message requires a reply or drafting responses that match a user’s tone and context.
“Breaking the problem into many smaller models works much better than asking one model to write a good email,” Hollingsworth explains.
When a new email arrives, a reply decision model classifies the message: is this something that needs a response, a scheduling action, or simply information the user should see?
If a response is needed, additional models analyze the intent of the email and predict the likely outcome of the interaction. These models determine patterns such as whether the conversation is moving toward scheduling a meeting, resolving a request, or continuing a longer relationship thread.
Memory is one of the most important parts of the system. Fyxer needs to decide which details should persist across conversations and which should disappear after a single exchange. When a new email arrives, retrieval models compare it with stored interactions and surface the memories most relevant to that person and thread.
OpenAI models power steps across Fyxer’s system. “We use OpenAI models for everything from digesting the email, so we can understand what it is actually about, to pulling in and re-ranking the context we want to include, to the actual email generation,” Shantsila says.
2. Train on how great assistants actually work
Before launching its AI product, Fyxer spent years operating a human-powered executive assistant service. Over time, the team accumulated a dataset built from more than 500,000 hours of annotated executive workflows, capturing how real assistants manage professional communication.
Those examples gave Fyxer training data drawn from the job itself. They captured the small judgments behind a good response: when to answer quickly, when to wait, which earlier conversation matters, and how the same request can call for a different response from one person to another.
Fyxer uses supervised fine-tuning and Low-Rank Adaptation (LoRA) across its broader system to create task-specific model variants while controlling training cost. Early in the product’s development, the team used OpenAI’s fine-tuning platform for tasks that needed high accuracy. More recently, Fyxer worked with OpenAI’s managed fine-tuning team to put a new checkpoint into production.
“OpenAI has been pivotal for us in helping us transfer the learning that we have about our customer and successfully incorporate it into how the models work.”
—Joey Dwonczyk, AI/ML Product Engineer, Fyxer
Before any model is deployed, Fyxer evaluates it on validation sets built around its own email tasks, including drafting, classification, and prioritization. The team weighs accuracy alongside response time and cost, since the best choice can vary by job.
3. Turn user feedback into a self-training loop
Once deployed, Fyxer’s system continues improving through real user feedback. When someone edits a draft before sending it, the difference between the original and final email shows which output they preferred.
Fyxer converts those comparisons into training data using Direct Preference Optimization (DPO). Instead of manually labeling every example, the model learns from pairs of outputs: the original draft and the user-edited version.
Every drafting change then goes through an A/B test. Fyxer ships the new version only when it produces a statistically significant improvement. Its user volume means the team can sometimes reach that threshold within a day.
Today, 53% of Fyxer’s AI-generated drafts are accepted as written. That means the system is correctly predicting intent and tone for a large share of real conversations. In 2025 alone, Fyxer grew from $1 million to $32 million in annual recurring revenue.
For Hollingsworth, the stronger signal is retention. Many Fyxer customers are not deeply technical, and for some, Fyxer is the first AI system they use every day.
“Everyone talks about ARR, but I think retention is the real flex. Over 90% of our users are still paying at the 90-day mark with us, and still using us every day.”
—Archie Hollingsworth, Co-founder, Fyxer
From drafts to a proactive assistant
Looking ahead, Fyxer is building a richer understanding of relationships, preferences, and ongoing work threads. Their assistant continues to evolve beyond drafting replies toward a broader AI assistant that can manage more of a user’s communication and coordination workload.
“Our vision is to get our customers doing as much of the work they absolutely love,” Hollingsworth says. “We want to get them to a place where they never have to open their computer and can trust Fyxer to manage all of that.”
本文内容采集自官方网站，排版和翻译可能与原页面存在差异。
阅读官方全文