How Scout24 is building the next generation of real-estate s…

由 Scout24 运营的德国最大房地产平台，将寻房者、房主、房东和中介汇集在同一个生态系统中。多年来， AI 已在反欺诈、营销效率和房产估价等方面发挥作用，而更强大的大型语言模型的出现，为平台向客户推出一种新的东西创造了机会：一个智能的对话式房产助手。

我们采访了 Scout24 的首席技术官 Gertrud Kolb ，听她讲述团队如何用 GPT‑5 打造对话式搜索体验、他们对“智能交互”的理解，以及在上线前如何在创新与质量、信任之间取得平衡。

“保持好奇、保持开放、开始去做。多交流、互相学习——并且享受其中。”

—— Gertrud Kolb ，首席技术官， Scout24

成果一览

• 由 GPT‑5 驱动的 Scout24 对话式搜索助手 HeyImmo
• 自适应回答格式：简明摘要、要点列举，或带图预览的直接房源列表
• 基于 OpenAI Evals 框架自建的评估系统，用以定义并衡量回答质量
• 全公司范围的“ swarm testing ”，用于校准预期并压测极端情况
• 面向速度与可靠性的系统架构，支持快速迭代改进
• 与 OpenAI Solutions Architects 的紧密协作，提升安全性、质量与用户体验

推广内部

对 Scout24 来说，首要任务是搜索——这是平台的核心功能。但在与 OpenAI 共同做原型时，团队很快发现用户需要的不只是更好的搜索结果，而是指引。“我们很快意识到，这不仅仅是搜索问题。我们需要一个房地产专家式的助手——一个能在整个平台体验中陪伴并引导你的对话伙伴，” Gertrud 说。

基于这一认识，团队开发出一个能提问澄清问题、呈现相关房源、总结可选项并根据用户需要调整回答格式的对话式助手。有的人偏好要点式，有的人要完整摘要，还有的人希望直接看到房源。弄清“什么才是真正的智能交互”成了最大的挑战之一。

在技术架构上，团队也做出了取舍。早期方案倾向于复杂的多代理框架，但团队有意简化，采用了 function calling 和高度聚焦的组件，以便快速交付并从用户反馈中学习。延迟、相关性、性能以及答案结构成为关注的核心。

另一个关键教训是如何衡量质量。为此， Scout24 基于 OpenAI Evals 框架打造了自有评估模型——帮助团队明确“达到可接受标准”究竟意味着什么，并在多种场景下进行衡量。他们还配合大范围内部测试，几乎邀请了全体员工试用产品、反馈问题并压测边缘案例。公司为确保体验达到既定标准甚至推迟过一次上线。

在开发过程中， Scout24 与 OpenAI 保持密切合作。 Gertrud 称这段协作“极为有帮助”，有助于提升回答质量、塑造响应结构，并统一对 AI 驱动搜索体验中“优良”标准的认识。

正是这种迭代工程、严格评估与广泛测试的结合，为用户带来了既智能又值得信赖的体验奠定了基础。

领导力要点（来自 Scout24 ）

• 从 AI 能改造产品核心的环节入手，并在此基础上持续迭代。
• 质量至关重要：定义“达标”的标准并建立衡量工具。
• 大规模的内部测试能揭示真实的用户期望。
• 早发觉、快学习——但也要知道何时放慢脚步。
• 产品团队与技术专家的紧密合作能加速进展。

下一步

Scout24 现在的重点是互联互通：把这位房产助手扩展到寻房者、房东、房主和中介之间。从帮助中介生成经过验证的平面图，到为寻房者提供更个性化的指导，团队看到了众多深化体验、增强平台网络效应的机会。

“针对每类客户群，我们都有大量基于 AI 的功能设想。我们才刚刚起步而已。”

—— Gertrud Kolb ，首席技术官， Scout24

加入前沿

全球已有超过一百万家企业借助 OpenAI 达成实质性成果。欲了解详情，请联系： https://openai.com/contact-sales/

Scout24 operates Germany’s largest real-estate platform, connecting seekers, homeowners, landlords and agents in one ecosystem. AI has supported areas like fraud detection, marketing efficiency, and property valuation for years, and  the rise of powerful large language models created an opportunity to build something new for customers: an intelligent, conversational real-estate assistant.

We sat down with Gertrud Kolb, Chief Technology Officer at Scout24, to hear how her team built a GPT‑5 powered search experience, what they learned about “intelligent interaction,” and how they balanced innovation with quality and trust before launch.

“Be curious, be open, start doing things. Talk with each other, learn from each other—and have fun.”

• GPT‑5 powering Scout24’s conversational search assistant HeyImmo
• Adaptive answer formats: concise summaries, bullet points, or direct listings with image previews
• Custom eval system built on the OpenAI Evals framework to define and measure answer quality
• Company-wide “swarm testing” to refine expectations and stress-test edge cases
• Architecture built for speed and reliability, and rapid improvements
• Tight collaboration with OpenAI Solutions Architects to improve safety, quality, and user experience

Scout24’s first priority was search—the core function of its platform. But as the team began prototyping with OpenAI they realised that customers wanted more than improved results. They wanted guidance. “We saw very quickly it’s not only about search. We needed a real-estate expert assistant—a sparring partner—which guides you through the whole platform experience” said Gertrud.

This insight led to a conversational assistant that can ask clarifying questions, surface relevant listings, summarize options, and adapt the answer format to the user’s needs. Some people want bullet points; others want full summaries; others want listings immediately. Understanding “what intelligent interaction really means” became one of the biggest challenges.

Scout24 also had to find the right technical architecture. Early ideas leaned towards a complex multi-agent framework, but the team intentionally simplified: using function calling and focused components so they could ship quickly and learn from users. Latency, relevance, performance, and the structure of answers became central themes.

One of the most important lessons was understanding quality. To solve this, Scout24 built its own evaluation model inspired by the OpenAI evals framework—helping the team define what “good enough” actually means, and measure it across many scenarios. They combined this with broad internal testing—almost all employees were invited to try the product, provide feedback, and stress-test edge cases. The company delayed launch once to ensure the experience met the bar they had defined.

Throughout development, Scout24 worked closely with the OpenAI team. Gertrud described the collaboration as “immensely helpful” in raising answer quality, shaping response structures, and aligning on what ‘good’ looks like for an AI-powered search experience.

This combination of iterative engineering, rigorous evaluation, and broad testing set the foundation for a customer experience that feels both intelligent and trustworthy.

• Start where AI can transform the core of your product—and iterate from there.
• Quality matters: define “good enough” and build the tools to measure it.
• Internal testing at scale uncovers real user expectations.
• Launch early, learn fast—but know when to slow down.
• A tight partnership between product teams and technical experts accelerates progress.

Scout24’s focus now is on interconnectivity: expanding the real-estate assistant across seekers, landlords, homeowners, and agents. From helping agents produce verified floor plans to giving seekers more personalized guidance, the team sees many opportunities to deepen the experience and strengthen network effects across the marketplace.

“For each customer group we have so many ideas for AI-driven features. We’re only just getting started”

Join the frontier

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

Source