供资
大理石 Cognition 融资 3.75M 预播构建AI同事通过观察学习
在 Google 上的首选来源中添加 Unite.AI
自动化办公工作最困难的部分往往发生在AI agent采取第一行动前：找出员工的实际行动。 一个工作流程可能跨越一个收件箱、一个电子表格和几个业务应用程序，但小组理解的例外情况除外，但从未记录在正式的规格中。 Marble Cognition拥有375万美元的种子前资金，将业务知识转化为AI同事。
Heartcore和Seedcamp共同领导了该轮比赛。 天使参与者包括Datadog董事会成员，Personio创始人和来自麦肯锡，Mistral AI和Lovable的经理，根据公告。 慕尼黑公司将利用融资规模推论，扩充其工程团队和美国的发布。
Marble的方法始于观察工作。 它捕捉了流程的展开方式，将这些观测结果转化为结构化的地图，并部署Marble Coworks在员工已经使用的工具中执行由此产生的工作流程。 在需要判决或批准的地方，人们仍然参与其中。
将未写的工作流程输入自动化蓝图
能够导航软件的agent仍然需要知道哪些记录可以查询，什么是完成的任务，什么时候例外需要升级。 如果公司的实际操作程序是习惯、非正式的交割和分散在其系统内的知识，就很难提供这些指示。
大理石打赌，随着软件开发和agent建设变得更加容易，这种上下文变得更有价值。 它提出的优点是准确理解正在自动化的工作：行动顺序、所使用的信息和将一个步骤与另一个步骤联系起来的决定。
公司称其桌面应用记录屏幕，点击和按键在业务应用内部。 它利用这些观测来构建结构化的流程图，AI系统可以解释，然后创建合作者来执行流程结束。 这把工作流程发现与部署联系起来，而不是让客户将一个流程分析转换成单独的自动化项目。
“我们谈论的每一个公司都希望大赦国际接管真正的工作。 但没有人能告诉人工智能该如何完成这项工作，“所以我们首先要解决这个问题。”
AI 操作现有商业工具的同事们
Marble现有产品网站使得执行模式更加具体。 它描述了使用本地计算机和浏览器的同工，允许团队在其已经使用的环境中授权工作。 它的退税图案结合了客户请求，退税政策和商业记录：合作者对请求进行分类，检查政策，起草答复，并向一个人发送例外。
这个例子说明了背景为何重要。 提出一个可信的答复只是一步。 一个有用的同事必须将请求与相关规则和客户信息联系起来，然后区分日常工作与不应自动进行的案件。
在业务方面，Marble正在设法缩小理解工作流程与运行工作流程之间的差距。 流程图为工作提供了规格；合作者将其带入业务应用；人文审查保留了任务需要的决定点。 能否成功部署取决于该规格如何能很好地反映例外情况以及共同路径。
早期的客户和金融服务部署
Marble 称公司与技术独角兽公司、私人股权支持公司和AI公司合作，在其客户中将EGYM、Mutares和Enam命名为EGYM公司。 这些组织为一种产品提供了一个自然的试验场，目的是进行重复的、工具繁多的行政操作。
在一家无名的私人股权拥有的美国金融服务公司的一次部署中，一名同事Marble 称将每个案件的处理时间减少了90%，帮助团队在第一个月里处理了50%以上的案件，而没有额外的聘请。 合作者在通过观察小组了解工作流程后，为一名专家准备和解提议。
这些数字是连报的具体部署结果，而不是一般业绩基准测试。 他们描述了两个不同的结果 -- -- 一个个案花费的时间和小组的总吞吐量 -- -- 专家的批准仍然是这一进程的一部分。
这个例子也指出了公司的商业重点。 具有可重复结构和明确批准边界的工作可以提供一个部署agent的初始位置，而更模棱两可的决定仍由小组作出。 扩大这一方法要求系统在一个显然熟悉的案件不属于学到的工作流程时予以识别。
隐私控制是观察工作的核心
从雇员活动中学习到的产品需要围绕收集和获取的界限明确。 大理石称其桌面应用程序不包括私人应用程序，私人标签和工作时间以外的活动。 雇员可以暂停记录，分析是综合进行的，而不是用来评估个人。
该公司还在某人或AI模型看到之前，包括在客户自己云层内的部署中，将称捕获的材料匿名。 这些是公司所说的保障措施；它们对于系统如何运作很重要，而不是产品附带的特点。
就企业部署而言，实际问题包括：哪些应用属于范围，如何删除敏感信息，以及谁可以获取由此产生的程序信息。 学习足够的工作流程复制，同时限制不必要的收集，是这一模式的核心工程和治理挑战。
用于推论、工程和美国扩张的种子前资本
Rösner和共同创始人Christophe Schmit,Marble的CTO在慕尼黑的数字技术与管理中心会面，并于2025年末创立了公司。 Rösner此前曾在AI销售-agent公司11x工作。 前麻省理工学院研究员施密特(Schmit)，此前根据发行情况曾在埃纳姆工作。
这笔资金将支持将观测到的工作流程转变为可部署的合作者所需的基础设施和工程，以及Marble计划的美国发布。 Seedcamp的参股也将公司与一个早期支持UiPath开发的投资者联系起来，将Marble置于一个更广泛的企业中，试图通过软件减少重复的办公工作。
Marble的主张是，对企业自动化的下一个限制是，agent收到的工艺知识的质量越来越高。 如果工作流程需要，将如何很好地保护人类的判断。
本文由 AI图鉴自动抓取并翻译，版权归原作者所有；如需查看原文请点下方链接。
Funding
Marble Cognition Raises $3.75M Pre-Seed to Build AI Coworkers That Learn by Observing
Add Unite.AI to your preferred sources on Google
The hardest part of automating office work often happens before an AI agent takes its first action: figuring out what employees actually do. A workflow might cross an inbox, a spreadsheet and several business applications, with its exceptions understood by the team but never captured in a formal specification. Marble Cognition has raised $3.75 million in pre-seed funding to turn that operational knowledge into AI coworkers.
Heartcore and Seedcamp co-led the round. Angel participants include a Datadog board member, a Personio founder and executives from McKinsey, Mistral AI and Lovable, according to the announcement. The Munich company will use the financing to scale inference, expand its engineering team and launch in the United States.
Marble’s approach starts with observing work. It captures how processes unfold, turns those observations into structured maps and deploys Marble Coworkers to carry out the resulting workflows in the tools employees already use. People remain involved where judgment or approval is required.
Turning Unwritten Workflows Into an Automation Blueprint
An agent that can navigate software still needs to know which records to consult, what counts as a completed task and when an exception requires escalation. Those instructions are difficult to supply when a company’s real operating procedures exist as habits, informal handoffs and knowledge scattered across its systems.
Marble is betting that this context becomes more valuable as software development and agent-building become easier. Its proposed advantage is an accurate understanding of the work being automated: the sequence of actions, the information used and the decisions that connect one step to another.
The company says its desktop application records screens, clicks and keystrokes within business applications. It uses those observations to build structured process maps that AI systems can interpret, then creates coworkers that execute the process end to end. This connects workflow discovery with deployment, instead of leaving the customer to translate a process analysis into a separate automation project.
“Every company we talk to wants AI to take over real work. But nobody can tell the AI how that work is actually done,” said Felix Rösner, Marble’s co-founder and CEO, in the announcement. “So we solve that first.”
AI Coworkers That Operate Across Existing Business Tools
Marble’s existing product website makes the execution model more concrete. It describes coworkers with native computer and browser use, allowing teams to delegate work in the environments they already use. Its refund-triage illustration combines a customer request, a refund policy and business records: the coworker classifies the request, checks the policy, drafts a response and sends exceptions back to a person.
That example illustrates why context matters. Producing a plausible reply is only one step. A useful coworker must connect the request with the relevant rules and customer information, then distinguish routine work from a case that should not proceed automatically.
In operational terms, Marble is seeking to close the gap between understanding a workflow and running it. The process map supplies a specification for the work; the coworker carries it into the business applications; and human review preserves a decision point where the task requires it. Successful deployment depends on how well that specification captures exceptions as well as the common path.
Early Customers and a Financial-Services Deployment
Marble says it works with technology unicorns, private-equity-backed companies and AI roll-ups, naming EGYM, Mutares and Enam among its customers. These organisations offer a natural testing ground for a product aimed at repetitive, tool-heavy administrative operations.
In one deployment at an unnamed, private-equity-owned US financial-services company, Marble says a coworker reduced handling time per case by 90% and helped the team handle 50% more cases during its first month without additional hires. The coworker prepares settlement offers for a specialist to approve, having learned the workflow by observing the team.
Those figures are company-reported results from a specific deployment, rather than a general performance benchmark. They describe two different outcomes—time spent on an individual case and the team’s total throughput—and the specialist’s approval remains part of the process.
The example also points to the company’s commercial focus. Work with a repeatable structure and a clear approval boundary can provide an initial place to deploy an agent, while more ambiguous decisions remain with the team. Scaling that approach requires the system to recognise when an apparently familiar case falls outside the learned workflow.
Privacy Controls Are Central to Observing Work
A product that learns from employee activity needs clear boundaries around collection and access. Marble says its desktop application excludes private applications, private tabs and activity outside working hours. Employees can pause recording, and analysis is performed in aggregate rather than used to assess individuals.
The company also says captured material is anonymised before a person or AI model sees it, including in deployments within a customer’s own cloud. These are the company’s stated safeguards; they are material to how the system is meant to operate, rather than incidental features of the product.
For an enterprise deployment, the practical questions include which applications are in scope, how sensitive information is removed and who can access the resulting process information. Learning enough to reproduce a workflow while limiting unnecessary collection is a core engineering and governance challenge for this model.
Pre-Seed Capital for Inference, Engineering and US Expansion
Rösner and co-founder Christophe Schmit, Marble’s CTO, met at Munich’s Center for Digital Technology and Management and founded the company in late 2025. Rösner previously worked at AI sales-agent company 11x. Schmit, a former MIT researcher, previously worked at Enam, according to the release.
The financing will support the infrastructure and engineering needed to turn observed workflows into deployable coworkers, alongside Marble’s planned US launch. Seedcamp’s participation also connects the company with an investor that backed UiPath early in its development, placing Marble in a broader lineage of businesses trying to reduce repetitive office work through software.
Marble’s proposition is that the next constraint on enterprise automation is increasingly the quality of the process knowledge an agent receives. Its AI coworkers will be judged on how reliably they turn that knowledge into completed work—and how well they preserve human judgment when the workflow calls for it.
Auto-fetched and translated by AI Atlas. All rights belong to the original author.