Wayfair boosts catalog accuracy and support speed with OpenAI

Wayfair 将 OpenAI 模型 整合进关键的内部系统，以在大规模上改进供应商支持工作流和商品目录的数据质量。最初在 2024 年以小规模试验验证价值，随后演进为全面投入生产的系统，减少人工工作量、加快决策速度，并提升数百万件商品的数据质量。

与其把生成式 AI 当作一次性实验或点状解决方案， Wayfair 选择把 OpenAI 模型 嵌入核心运营流程。公司先集中在复杂度和规模需求最高的环节：分流与解决供应商支持请求，以及在约 3,000 万件商品的目录中，统一改进数以万计的商品属性标签。

解决大规模目录质量问题

Wayfair 的目录团队管理近千万计的商品，涵盖近一千个商品类别。像颜色、材质、尺寸或具体特征这样的属性标签要保持一致且准确，对于搜索、推荐和陈列至关重要。

“数据质量越好，就越能赢得顾客信任。这很关键，因为它能帮助消费者做出正确的购买决定，直接减少因商品描述不符而产生的昂贵退货等下游问题，” Wayfair 的目录商品管理副总监 Jessica D'Arcy 说。

在引入 OpenAI 之前，标签改进主要依赖供应商和顾客主动反馈有误之处，人工核查无法跟上海量数据的节奏。早期为单个标签定制的 AI 模型虽有效，但构建和维护成本高昂。“我们起先为各个标签做定制模型，从技术上可行，” Wayfair 的资深机器学习科学家 Carolyn Phillips 说，“但面对 47,000 个标签，这种办法根本无法扩展。”

构建可复用的 AI 架构

为突破单一模型的局限， Wayfair 构建了基于单一 OpenAI 模型 的标签无关系统。一个“定义代理”会抓取网络和内部定义，为每个标签生成上下文化的含义。“真正的瓶颈不是模型本身的性能，” Phillips 指出，“而是人力去定义并编码每个标签含义所花的时间。”这些定义连同汇聚自 Wayfair 数据生态的商品信息，输入到一个可以跨商品类别分类属性的框架中。团队现在以比一年前快 70 倍的速度为新属性扩展模型覆盖。

该系统已在超过一百万件商品上投入生产。首批属性增强的商品已上线足够长时间，可以评估数据质量改善对顾客旅程的影响。“当你提升属性完整性时，这不是抽象的；你能在 SEO 和 PLA 表现上看到——体现在顾客如何发现商品上，” Phillips 说。受控的 A/B 测试显示，试验组在曝光量、点击量和页面排名上都有显著增长。

不过， Wayfair 并未把商品数据修正的最终决定完全交给模型。“我们的目标是建立信任，让顾客对所购商品完全有信心，” Phillips 说。公司通过有结构的测试和人工抽检流程让员工实物验样以验证模型输出，并与供应商共同确认改动。现在，当基于数据的置信度足够高时，自动化系统会直接覆盖内容并通知供应商；若未达到高标准或某标签被视为高风险， Wayfair 则会先寻求供应商确认再执行更改。

借助 Wilma 重塑供应商支持流程

Wayfair 与数万家供应商合作以维护其庞大目录。此前，处理供应商支持请求时，员工需要逐一审阅来票、手动识别供应商意图，并将问题分派给相应的内部负责人——既耗时又易出错。“供应商的请求并不简单，” Wayfair 的供应商支持与运营主管 Graham Ganssle 说，“它们涵盖数百种问题类型，单个员工不可能熟练掌握全部。”

Wayfair 为名为 Wilma 的产品加入了 agentic（具代理能力的）功能，用 AI 来增强这些工作流。其中第一个上线的功能是由 OpenAI 模型 驱动的工单分流。系统读取来件、补全缺失上下文并将工单路由到合适团队。 Wilma 被设计为能快速部署；基于已有与 OpenAI API 集成的系统，它从原型到上线约用时一个月。“ Wilma 给员工带来杠杆效应，” Ganssle 说，“它读工单、识别意图、从我们的数据库补上下文、必要时回访供应商，并把问题指向正确方向。”

除分流外， Wayfair 已为若干解决团队部署了十余条具代理能力的 AI 流。例如，为 Replacement Part Operations 团队配备的协同助手会阅读复杂的案件历史、建议下一步动作并起草回复，供人工审阅。这些助手以历史数据训练，学习何为成效。“模型能在整个服务旅程中综合语境，这是单个员工难以做到的，” Ganssle 说，“这种更广的可见性提升了顾客与供应商满意度。”

Wayfair 会跟踪 AI 建议与人工最终决策的一致频率——称为“ alignment rate ”。当团队内部一致性稳定达到预设阈值后，工作流可从辅助（“协同驾驶”）逐步转为半自动（“自动驾驶”）。这种分阶段做法有助于建立信任并在推广过程中保证质量控制。

“如果一开始就没有把问题正确分流，后续所有环节都会变慢。分流是基础。”—— Graham Ganssle， Wayfair 供应商支持运营

跨团队的可衡量影响

自将 OpenAI 模型 整合进内部系统以来， Wayfair 报告了可量化的改进。

在目录端，公司已修正约 250 万个商品属性标签，覆盖了超过一百万件在 Wayfair 目录中最显眼且最常被购买的商品。他们预计未来六个月内这一影响将扩大四倍。

在供应商支持方面，分流、协同驾驶与自动驾驶系统每月自动处理约 41,000 张工单（在部分工作流中提升率达 70%），并通过剔除常规人工工作大幅缩短周转时间。这显著降低了多条工作流的解决时长，提升供应商满意度，并减少这些流程中的工单重开率。

模型为工单和供应商意图提供的更广视野——超出单个员工在屏幕上能看到的范围——也助推了满意度的提升。

各团队在运营上报告的收益包括：

• 更快地分流与解决复杂供应商工单
• 提高供应商满意度
• 减少人工数据录入与分类工作
• 在无需掌握数百个专题的前提下覆盖更广的问题类型
• 在发布前对目录属性有更高把握

此外， Wayfair 已在约 1.2 万名员工规模的组织中部署了超过 1,200 个 ChatGPT Enterprise 帐号，用于支持临时任务、内部问题解决以及对生成式模型的试验。

与 OpenAI 的长期合作

Wayfair 长期投入机器学习并与多个 AI 平台和大模型提供商合作，以推进业务发展。如今，前沿模型，尤其是多模态系统的进步，正在扩展团队能构建的能力。在家居零售领域，商品多为视觉性、风格性且常伴随主观判断，这点尤为重要。

“我们对如今能解决的问题范围感到振奋，” Carolyn Phillips 说，“传统算法需要严格定义的数据集，这些模型让我们能以曾不可扩展的方式处理模糊性与语境。”

从早期试点到生产服务， Wayfair 与 OpenAI 的关系已发展为战略伙伴，涵盖模型选择、部署最佳实践及更广泛的内部采纳。 Wayfair 的首席技术官 Fiona Tan 说：“最有价值的不是单纯获得模型的访问权，而是与他们共同探讨新用例并能迅速推进。”

展望未来，员工对 ChatGPT Enterprise 的需求强劲。 Wayfair 的团队把它视为能提升工作效率的实用工具。

顾客预期也在快速变化。越来越多消费者在日常生活中接受并使用 AI，他们也开始期待在浏览、比较与在线购物时获得类似能力支持。

“在家居场景中，顾客往往很难用确切的词来描述他们想要什么，” Fiona Tan 说，“自然语言和多模态系统有助于弥补这个差距。”

对 Wayfair 的领导层来说，目标仍是在人类专长之上进行增强，同时扩大内部能力。“我们正在为一个 AI 成为购物旅程一部分的世界做准备——无论是在网站上、通过支持渠道，还是通过对话界面，” Fiona Tan 总结道。

Wayfair, one of the world’s largest home goods retailers, has integrated OpenAI models into critical internal systems to improve supplier support workflows and product catalog quality at scale. What began as value-testing small scale releases in 2024 has evolved into a full production system that reduces manual effort, accelerates decision-making and improves data quality across millions of products.

Rather than treat generative AI as an experiment or point solution, Wayfair embedded OpenAI models into core operational workflows. The company focused first where complexity and need for scale were highest: routing and resolving supplier support requests and improving tens of thousands of product attributes consistently across a catalog of roughly 30 million items.

Solving catalog quality at scale

Wayfair’s catalog team manages tens of millions of products across nearly a thousand different product classes. Consistent and accurate product attribute tags — such as color, material, size or specific features — are essential for search, recommendations and merchandising.    

"The better our data quality, the more trust we build with the customer. It's essential because it empowers shoppers to make the right buying decisions, directly reducing costly downstream issues like returns from misrepresented products," said Jessica D'Arcy, Associate Director of Catalog Merchandising at Wayfair. 

Before OpenAI, tagging improvements primarily relied on suppliers and customers to tell Wayfair that something looked wrong. Manual effort could not keep up with the volume.  Early custom AI models for individual tags were effective, but proved expensive to build and maintain. “We started by building bespoke models for individual tags, and technically that worked,” said Carolyn Phillips, Wayfair’s staff machine learning scientist. “But when you’re looking at 47,000 tags, that approach just doesn’t scale.”

Building a reusable AI architecture

To get beyond one-off models, Wayfair created a tag-agnostic system built on a single OpenAI model. A “definition agent” ingests the web and internal definitions to produce contextual meaning for each tag. “The real bottleneck wasn’t the model performance,” said Phillips. “It was the human time required to define and encode what each tag actually meant.”   This context, along with product data aggregated from across Wayfair’s data ecosystem, feeds into a framework that can classify attributes across product classes. The team is now expanding model coverage to new attributes at 70x the rate they were just a year ago.

The system has now run in production on more than 1 million products.   And the first wave of products with enhanced attributes has now been live long enough to measure the impact of improving data quality on the customer journey.   “When you improve attribute completeness, it’s not abstract. You see it show up in SEO and PLA performance - in how customers discover products”, said Phillips.  A controlled A/B test showed a substantial and significantly significant increase in impressions, clicks , and page rank in the treatment group.

However, Wayfair didn't simply hand off decisions on correcting product data to the model. “Our objective is to build trust so that customers are completely confident in what they are purchasing,” said Phillips.   The company developed structured testing using a hands-on audit process in which associates physically inspect samples to validate model output, and worked with suppliers to validate changes.    Now, when data-based confidence is high, automated systems will overwrite the content directly and notify the supplier of the change.   And, when a high standard is not met or the tag is deemed high risk, Wayfair first seeks supplier confirmation before making the change.

Rethinking supplier support workflows with Wilma

Wayfair works with tens of thousands of suppliers to support their comprehensive catalog. To manage supplier support requests, Wayfair associates historically reviewed every incoming ticket, manually identified what suppliers were trying to accomplish, and routed issues to the correct internal owner—a time-consuming and error-prone process. “Supplier requests aren’t simple,” said Graham Ganssle, supplier support and operations at Wayfair. “They span hundreds of issue types, and no single associate can realistically master all of them.”

Wayfair added agentic features to a product named Wilma to augment these workflows with AI. One of the first features in production is ticket triage powered by an OpenAI model. The system reads incoming requests, fills in missing context and routes tickets to the appropriate team. Wilma was designed to be deployable fast; built on a system already integrated with OpenAI APIs, it moved from prototype to live in approximately one month. “Wilma gives associates leverage,” said Ganssle. “It reads the ticket, identifies intent, fills in context from our databases, reaches back out to suppliers if necessary, and points the issue in the right direction.”

Beyond routing, Wayfair has deployed a dozen agentic AI flows for specific resolution teams. For example, a co-pilot for the Replacement Part Operations team reads complex case history, proposes next steps and suggests draft responses that human associates review. These assistants are trained on historical data so they learn what success looks like in context. “The models can synthesize context across the entire journey in a way that’s hard for a single associate to do,” said Ganssle. “That broader visibility contributes to higher customer and supplier satisfaction.”

Wayfair tracks how often the AI’s recommendations match the human agent’s final decision—a metric called “alignment rate.” Within each team, when alignment consistently reaches a predetermined threshold, workflows can shift from assistive (“co-pilot”) to semi-autonomous (“autopilot”) modes. This staged approach builds trust and ensures quality controls during rollout.

“If you don’t route the issue correctly at the start, everything downstream slows down. Triage is foundational.”

Measurable impact across teams

Wayfair reports measurable improvements since integrating OpenAI models into internal systems.

On the catalog side, the company reduced the number of wrong or missing product attribute tags a customer might see—having corrected 2.5M product tags across over a million of the most visible and purchased products in the Wayfair catalog. They expect to quadruple this impact in the next six months.

In supplier support, triage, co-pilot, and auto-pilot systems have increased throughput by automating 41,000 tickets per month (that’s up to 70% in some workflows) and reduced turnaround times by removing routine manual work from associate workloads.  This dramatically cuts time to resolution for multiple workflows, significantly lifts supplier satisfaction, and reduces ticket re-opens  in those workflows.

The broader visibility that models provide into tickets and supplier intent—beyond what a single associate can see on a screen—has contributed to that increase in satisfaction.

Operationally, teams report:

• Faster routing and resolution of complex supplier tickets
• Increased supplier satisfaction
• Reduced manual data entry and classification work
• Broader issue coverage without requiring expertise across hundreds of topics
• Higher confidence in catalog attributes before publication.

Wayfair has also deployed more than 1,200 ChatGPT Enterprise seats across its approximately 12,000-person workforce to support ad hoc tasks, internal problem solving and experimentation with generative models.

A long-term partnership with OpenAI

Wayfair has a long history of investing in machine learning and partnering with AI platforms and LLM providers to advance their business. Now, advances in frontier models, particularly multimodal systems, are expanding what its teams can build. That matters in home retail, where products are visual, stylistic and often subjective.

“We’re excited about the scope of problems we can now tackle,” said Carolyn Phillips. “Traditional algorithms require tightly defined datasets. These models allow us to work through ambiguity and context in a way that wasn’t previously scalable.”

From early pilots to production services, the relationship between Wayfair and OpenAI has grown into a strategic partnership that spans model selection, best practices for deployment and broader internal adoption. “What’s been most valuable is the thought partnership,” said Fiona Tan, Chief Technology Officer, Wayfair. “It’s not just access to the models. It’s working through new use cases together and being able to move quickly.”

Looking forward, the employee demand for ChatGPT Enterprise has been strong. Teams at Wayfair see it as a practical tool that helps them move faster.

Customer expectations are also shifting quickly. More shoppers are becoming comfortable using AI in their daily lives, and they are beginning to expect similar capabilities when they browse, compare and buy online.

“At home, customers often don’t have the exact words for what they’re looking for,” said Fiona Tan. “Natural language and multimodal systems help bridge that gap.”

For Wayfair leaders, the goal remains to augment human expertise while scaling internal capability. “We’re building for a world where AI is part of the shopping journey—whether that’s on our site, through support, or through conversational interfaces,” concluded Fiona Tan.

Join the new era of work

Generated by RSStT. The copyright belongs to the original author.

Source