「英中对照」ProptechOS分享自主代理转型经验及建筑AI代理运营方法

Most property work done at a computer could already be done by AI, but it still isn’t. Eighteen months ago that was true for ProptechOS too. People carried work between systems, context lived in people’s heads, and answers stayed with whoever gave them. Today agents on ProptechOS have completed more than 12 000 tasks so far this year, and the time between a conversation and its consequence is often measured in minutes. We did this to ourselves first. This is how, what it changed, and why agentic operations for real estate are now what we offer.大多数在电脑上完成的物业工作，AI本来就已经能够完成，但至今仍未实现。18个月前，ProptechOS的情况也是如此。人们在各个系统之间传递工作，上下文存在于人的脑海中，答案也只留在给出答案的人那里。如今，今年以来ProptechOS上的智能体已完成超过12 000项任务，从对话到产生结果之间的时间往往以分钟计。我们先在自己身上践行了这一点。以下就是实现方式、改变了什么，以及为什么房地产智能体运营如今成为我们提供的服务。

• What is agentic speed? Agentic speed is the time between a conversation and its consequence: how quickly a meeting, a decision or an alarm turns into completed work.什么是智能体速度？智能体速度是从对话到产生结果之间的时间：会议、决策或警报转化为已完成工作的速度。

• What are agentic operations for real estate? Agentic operations is operational work in real estate done by AI agents. The agents read live building and business data, act within the rights they are granted and hand results to people or other agents. Real estate professionals stay in charge.什么是房地产智能体运营？智能体运营是由AI智能体完成房地产领域的运营工作。智能体读取实时楼宇和业务数据，在被授予的权限内行动，并将结果交给人员或其他智能体。房地产专业人士仍掌握主导权。

The moment the work changed工作发生改变的那一刻

We built ProptechOS to solve one problem: property owners had no real access to their own building data, and every new integration was slow and expensive. Our answer was an operating system for buildings. It reads from and writes to operational systems, with granular rights for everything that runs on it, and it is built on RealEstateCore, the open standard for building data.我们构建ProptechOS是为了解决一个问题：业主无法真正访问自己的楼宇数据，而且每一项新集成都既缓慢又昂贵。我们的答案是一个楼宇操作系统。它从运营系统读取数据并向其写入数据，为在其上运行的所有内容提供细粒度权限，并且构建于RealEstateCore这一楼宇数据开放标准之上。

In spring 2025 we connected large language models to ProptechOS through MCP. Watching an AI model use our platform directly, we saw that much of the work done in and around buildings no longer needed a dedicated application. A domain expert with an AI assistant, or an autonomous agent, could do it on demand.2025年春季，我们通过MCP将大语言模型连接到ProptechOS。观察AI模型直接使用我们的平台时，我们发现楼宇内部及周边的许多工作已不再需要专用应用程序。领域专家借助AI助手或自主智能体，即可按需完成这些工作。

That raised a direct question for us as a company. If agents can do this work for our customers, they can do it for us. So we started with ourselves.这给我们公司提出了一个直接的问题。如果智能体能为客户完成这些工作，它们也能为我们完成。因此，我们从自己开始。

Work only moved when we moved it工作只有在我们推动时才会流转

For me, the hardest part was admitting how much of our own work only moved because someone remembered to move it. When we looked honestly at our own operations, what limited our speed was everything around the work: handovers, context and knowledge.对我而言，最难的是承认我们自己的大量工作之所以能够推进，只是因为有人记得去推动。当我们坦诚审视自身运营时，限制速度的是工作周围的一切：交接、上下文和知识。

• Handovers were manual. Someone carried each piece of work from one stage to the next, often copying information between systems that were not linked.交接依靠人工。总有人把每项工作从一个阶段带到下一个阶段，还常常在互不连通的系统之间复制信息。
• Context stayed offline. Why a meeting happened, or who a new feature was really for, lived in people’s heads.上下文一直离线。召开某次会议的原因，或者新功能究竟面向谁，都存在于人们的脑海中。
• Knowledge did not compound. Every answer to a support question stayed with the person who gave it.知识无法积累沉淀。支持问题的每个答案都只留在给出答案的人那里。

Anyone who runs buildings will recognise the pattern. Facility teams carry work between the BMS, the FM system, work orders and meters in exactly the same way.任何运营楼宇的人都会熟悉这种模式。设施团队正是以完全相同的方式，在BMS、FM系统、工单和计量表之间传递工作。

Four steps to make our work readable for agents让我们的工作可被智能体读取的四个步骤

1. Inventory the work.盘点工作。

We mapped which tasks each team does, how much time each takes and how much it matters. In onboarding, 80% of the time went to four tasks. We automated those first.我们梳理了每个团队执行哪些任务、每项任务耗时多少以及重要程度。在接入流程中，80%的时间都花在了四项任务上。我们首先将这些任务自动化。

1. Connect the systems of record.连接记录系统。

Agents got controlled access to project management, our source code repositories and the CRM.智能体获得了对项目管理、我们的源代码仓库和CRM的受控访问权限。

1. Capture the conversations. 捕获对话。 

Nine out of ten of our meetings are digital. Together with Slack and email, they flow into our own knowledge system, with wide but controlled access.我们90%的会议都是数字化的。这些会议连同Slack和电子邮件一起汇入我们自有的知识系统，访问范围广泛但受控。

1. Keep every answer. 保留每个答案。 

A knowledge base grows with every answer and decision. People and agents use it alike, and one agent keeps it current from our conversations.知识库随着每一次回答和决策不断增长。人员和智能体都使用它，并且有一个智能体根据我们的对话保持其内容更新。

The result is that handovers became deterministic. When the state of a piece of work changes, the next step starts.其结果是，交接变成了确定性流程。当一项工作的状态发生变化时，下一步就会开始。

Task, agent, harness, systems of record任务、智能体、运行框架、记录系统

Every piece of work we agentify is described with the same four concepts.我们将每一项工作智能体化时，都用相同的四个概念来描述。

• Task: a defined piece of work with a handover in and a handover out. Onboarding devices is one example: modelling the digital twin and connecting the integration.任务：一项定义明确、具有输入交接和输出交接的工作。设备接入就是一个例子：对数字孪生进行建模并连接集成。
• Agent: performs the task and hands the result to the next step.智能体：执行任务并将结果移交到下一步。
• Harness: gives the agent the tools and skills to use outside systems and data, and limits what it may do.运行框架：为智能体提供使用外部系统和数据的工具与技能，并限制其可执行的操作。
• Systems of record: where the data of the work lives. The agent reads and writes there only through the harness.记录系统：工作数据存放的位置。智能体仅通过运行框架在其中读取和写入数据。

From there the method is repeatable, task by task. Identify and prioritise the task, trace it to its systems of record, choose or build the harness, then iterate on the task and its handover until it runs reliably. The same model applies directly to building operations, where ProptechOS is the harness and the building’s operational systems are the systems of record.从这四个概念出发，该方法可以逐任务重复。识别任务并确定优先级，追溯其记录系统，选择或构建运行框架，然后迭代任务及其交接，直到可靠运行。同一模型可直接适用于建筑运营：ProptechOS 是运行框架，建筑的运营系统是记录系统。

Agentic speed: from conversation to consequence智能体速度：从对话到成果

Every meeting we hold is transcribed. The transcript plus the right context lets agents deliver the next step: a proposal within 15 to 75 minutes of an early sales meeting, a customer-specific prototype and quotation the next day, and often a new product version before a customer working meeting has ended.我们召开的每次会议都会被转录。转录内容加上适当的上下文，使智能体能够交付下一步：在早期销售会议后 15 至 75 分钟内给出提案，第二天提供客户专属原型和报价，并且常常在客户工作会议结束前推出新产品版本。

The effect on sales alone: in the twelve months after we changed how we work, we sent more than four times as many proposals as in the twelve months before, with a smaller team.仅对销售的影响：在改变工作方式后的 12 个月里，我们发送的提案数量是此前 12 个月的四倍多，而团队规模更小。

What changed变化

The clearest evidence is in the work our customers see: buildings onboarded, systems connected and tasks completed.最明显的证据体现在客户看到的工作中：建筑已接入、系统已连接、任务已完成。

The Q3 figure includes one new customer’s full portfolio of about 6 000 buildings, onboarded by agents. Connectors can now be provisioned, and even built, by people who are not developers. Each one is also reusable for the next customer with the same system.Q3 数据包括一位新客户约 6 000 栋建筑的完整物业组合，由智能体完成接入。现在，非开发人员也可以配置甚至构建连接器。每个连接器还可复用于下一个采用相同系统的客户。

Agents at work in real buildings智能体在真实建筑中工作

The same platform now runs agents in our customers’ live operations. Across the platform, agents have completed more than 12 000 tasks so far in 2026, from hourly plant checks to daily analyses and enrichment of building data. More than 300 agents are in production. At a conservative 15 minutes of expert work per task, that is more than 3 000 hours of expert work, time that operations teams can spend on decisions.同一平台现在在客户的实时运营中运行智能体。2026 年截至目前，全平台智能体已完成超过 12 000 项任务，从每小时设备巡检到建筑数据的每日分析与充实。超过 300 个智能体已投入生产。按每项任务需 15 分钟专家工作的保守估计，这相当于超过 3 000 小时的专家工作，运营团队可将这些时间用于决策。

Two examples from the Nordics:来自北欧的两个例子：

• Vasakronan built SpotOn Control, an AI agent, with ProptechOS. It has been nominated for Tech Excellence Innovation at Dagens Industri’s Tech Excellence Awards.Vasakronan 借助 ProptechOS 构建了 AI 智能体 SpotOn Control。它已获 Dagens Industri 的 Tech Excellence Awards 之 Tech Excellence Innovation 提名。
• KLP reduced energy use by 31% at its Eufemia property with ProptechOS.KLP 借助 ProptechOS 将其 Eufemia 物业的能耗降低了 31%。

Agents can work across building systems and business systems alike: BMS, meters and IoT on one side, FM, work orders, leases, tenant data and BIM on the other. That lets them do cross-system property work, such as turning a deviation in a plant into a correctly routed work order. ProptechOS is also open to any AI model. Customers’ own AI tools and partner agents can operate on the same harmonised data through MCP, within the same rights.智能体可跨建筑系统与业务系统工作：一端是 BMS、仪表和 IoT，另一端是 FM、工单、租约、租户数据和 BIM。这使它们能够执行跨系统的物业工作，例如将设备偏差转化为正确路由的工单。ProptechOS 也开放给任何 AI 模型。客户自有的 AI 工具和合作伙伴智能体可以通过 MCP，在相同权限范围内基于同一套统一协调的数据运行。

Speed with guardrails有护栏的速度

Our agents act on real buildings, so speed only counts when it is governed. The rights model we originally built for applications now governs every agent.我们的智能体作用于真实建筑，因此速度只有在受控的前提下才有意义。我们最初为应用构建的权限模型，现在管理着每个智能体。

• Permission policy per agent: each agent gets only the access and rights it is granted, down to the individual data point.每个智能体的权限策略：每个智能体仅获得被授予的访问权限，细化到单个数据点。
• Audit trail: every action an agent takes is recorded and traceable.审计追踪：智能体的每个操作都会被记录且可追溯。
• Observability: we and our customers see what agents do, as they do it.可观测性：我们和客户可以在智能体执行操作时查看其行为。
• Security and compliance: ISO 27001 and NIS2, for operations that count as critical infrastructure.安全与合规：面向被认定为关键基础设施的运营，遵循 ISO 27001 和 NIS2。

Business-critical means someone has assured it. An agent that ran once is a demo.业务关键意味着有人为此提供了保证。只运行过一次的智能体只是演示。

Agentic Fridays

Tools did not change our habits by themselves. Building away your own manual work while doing that work takes deliberate time.工具本身并没有改变我们的习惯。一边做着手工工作，一边把自己的手工工作构建掉，需要专门投入时间。

Every third Friday, everyone at ProptechOS agentifies part of their own team’s processes. It gives us coordinated time for the change, so it does not get squeezed in between everything else. It has become our most loved institution and our fastest learning loop. Alongside it, more frequent all-hands strategy reviews keep everyone aligned on where we are heading and why.每个月第三个周五，ProptechOS 的每位员工都会将自己团队流程的一部分进行智能体化。这为我们提供了协同一致的时间来推动变革，使其不会被挤在其他所有事务之间。它已成为我们最喜爱的制度，也是最快的学习闭环。与此同时，更频繁的全员战略评审让每个人都对我们的前进方向及原因保持一致。

Where to start in your portfolio从资产组合的何处着手

The four steps that worked for us work for building operations too.对我们行之有效的四个步骤，同样适用于建筑运营。

1. Inventory the operational work. List the recurring tasks: alarm handling, meter readings and anomalies, work order triage, energy follow-up, tenant requests. A few tasks usually take most of the time. Start there.盘点运营工作。列出重复性任务：告警处理、仪表抄表与异常、工单分诊、能耗跟进、租户请求。少数任务通常占用大部分时间。就从这些任务开始。
2. Connect the systems of record once. BMS, SCADA, meters and IoT, plus FM, work order, lease and tenant systems. Connectors read data and write control back.一次性连接各记录系统。BMS、SCADA、仪表和 IoT，以及 FM、工单、租约和租户系统。连接器读取数据并回写控制。
3. Give agents context. Harmonise the data on an open standard such as RealEstateCore, so any agent understands what a sensor, a room or a lease is, and how they relate.赋予智能体上下文。基于 RealEstateCore 等开放标准统一数据，让任何智能体都能理解传感器、房间或租约是什么，以及它们之间的关系。
4. Start with one task and its handover. Grant the agent only the rights it needs, keep the audit trail on, and expand once it runs reliably.从一项任务及其交接开始。仅授予智能体所需权限，保持开启审计追踪，并在其稳定运行后扩展。

Most of our customers start with a shared view of their data in dashboards, then add agents for tasks that are repetitive or not done at all today. Your team stays the expert and in control. Agents add speed and scale.我们的大多数客户先从仪表盘中的数据共享视图开始，然后为重复性或如今完全无人执行的任务添加智能体。您的团队始终是专家并掌控全局。智能体带来速度和规模化能力。

What this means for real estate这对房地产意味着什么

We believe the operational work of real estate will be done by AI agents, led by real estate professionals. Our role is to make building and business data usable for any AI model, so that agents can run real estate operations. That is what we mean by agentic operations for real estate professionals.我们相信，房地产的运营工作将由 AI 智能体完成，并由房地产专业人士主导。我们的职责是让建筑和业务数据可供任何 AI 模型使用，从而使智能体能够运行房地产运营。这就是我们所说的面向房地产专业人士的智能体运营。

Real estate companies use AI in three ways. Autonomous expert agents take on the repeatable operational work around the clock. Domain experts build the tools they need on their own data. Experts with AI assistants get answers and analysis on demand. All three need the same foundation from the building: live data, harmonised so agents understand it, the ability to write back to operational systems, and rights that control what each agent may do. That foundation is ProptechOS.房地产公司通过三种方式使用 AI。自主专家代理全天候承担可重复的运营工作。领域专家基于自有数据构建所需工具。配备 AI 助手的专家可按需获取答案和分析。这三者都需要楼宇提供相同的基础：经过统一协调、让代理能够理解的实时数据，回写运营系统的能力，以及控制每个代理可执行操作的权限。这个基础就是 ProptechOS。

The potential is large. Morgan Stanley Research estimates that 37% of tasks in real estate can be automated. Our own study from early 2026 puts the figure for building operations at 48%. Our target is best-practice operations in energy, tenant experience and operational quality at about half the work.潜力巨大。Morgan Stanley Research估计，房地产行业 37% 的任务可以实现自动化。我们 2026 年初的研究显示，楼宇运营领域的这一比例为 48%。我们的目标是，在能源、租户体验和运营质量方面，以约一半的工作量实现最佳实践运营。

We have moved from systems people use to agents that work for people, with agents in production and our own numbers to show for it. I believe every real estate team can make the same move, one task at a time. See how agentic operations work, or book a demo to explore what agents could take on in your portfolio.我们已经从人们使用的系统，转变为为人们工作的代理；代理已投入生产，我们也有实际数据为证。我相信每个房地产团队都能实现同样的转变，一次推进一项任务。了解代理式运营如何运作，或预约演示，探索代理可以为您的投资组合承担哪些工作。

Frequently asked questions常见问题

What are agentic operations for real estate? Operational work in real estate done by AI agents that read live building and business data and act within the rights they are granted. Real estate professionals lead the work and stay in control.什么是房地产代理式运营？由 AI 代理完成房地产运营工作，这些代理读取实时楼宇和业务数据，并在授予的权限内采取行动。房地产专业人员主导工作并始终保持控制。

Where should a property owner start with AI agents? Inventory the recurring operational tasks, connect the systems that hold their data, and start with one task and its handover. Most owners begin with a shared view of their data, then add agents for repetitive work.物业业主应从何处着手使用 AI 代理？梳理重复性运营任务，连接保存其数据的系统，并从一项任务及其交接开始。大多数业主首先建立共享数据视图，然后为重复性工作添加代理。

How are AI agents governed in ProptechOS? Each agent has its own permission policy down to the individual data point, every action is recorded in an audit trail, and operations are observable as they happen. Built to the requirements of ISO 27001 and NIS2, for operations that count as critical infrastructure.ProptechOS 中如何治理 AI 代理？每个代理都有独立的权限策略，精细到单个数据点；每个操作都记录在审计追踪中，运营过程可在发生时进行观测。系统按照 ISO 27001 和 NIS2 的要求构建，适用于属于关键基础设施的运营。

Do agents replace facility teams? Agents take on repeatable monitoring, analysis and routine actions. The team stays the expert, decides what agents may do and spends more time on decisions.代理会取代设施团队吗？代理承担可重复的监控、分析和常规操作。团队仍是专家，决定代理可以做什么，并将更多时间用于决策。

Which AI models can work with ProptechOS? Any. ProptechOS harmonises building哪些 AI 模型可与 ProptechOS 配合使用？任何模型。ProptechOS 统一楼宇

The post How we became an agentic company appeared first on ProptechOS.文章 我们如何成为一家代理式公司最早发布于 ProptechOS。