Many companies want to use artificial intelligence to reduce manual labor, but they encounter the same obstacle even before choosing a model or platform: they cannot precisely describe how the tasks they wish to automate are actually carried out. Layered procedures built up over time, exceptions handled via email, spreadsheets updated by different people, and knowledge reliant on individual experience make it difficult to turn daily work into instructions executable by software.
Marble Cognition, a Munich-based startup that raised $3.75 million—equivalent to approximately €3.35 million—in a pre-seed round, is tackling this issue. The deal was co-led by Heartcore and Seedcamp; angel investors also participated in the funding, including a Datadog board member, a Personio founder, and executives from McKinsey, Mistral AI, and Lovable.
The company, which operates under the name Marble, plans to use the funds to expand inference capacity, grow its engineering team in Munich, and prepare for a launch in the United States. It is an early stage for a company founded in late 2025, but the funding signals which segment of the AI agent market is drawing attention: not just building assistants capable of answering questions or generating content, but capturing the operational context required for them to execute concrete tasks.
First understand the work, then delegate it to an agent
Marble was founded by Felix Rösner and Christophe Schmit, who met in Munich's CDTM program. The startup's value proposition begins with observing activities carried out across enterprise software: its desktop application records signals such as on-screen activity, clicks, and keystrokes across workplace applications. From these inputs, Marble claims to reconstruct structured process maps that AI systems can read.
On top of this mapping, the so-called Marble Coworkers are created—agents designed to execute end-to-end processes using the tools a company already owns. The architecture does not necessarily eliminate human oversight: for steps requiring appraisal, approvals, or sensitive decisions, the planned model keeps a human in the loop.
The distinction is significant. Most enterprise automation projects start from formalized rules, app integrations, and workflows that someone has to design. Marble, on the other hand, is trying to productize the preceding phase: identifying the actual way work gets done, including the habits and steps that rarely make it into an operating manual. Rösner summarizes the approach by arguing that agents should first observe the work and only then take over a part of it.
According to Marble, in the organizations it works with, around 90% of key processes are not formally documented. The figure is a company estimate rather than a general market measure, but it accurately describes the practical difficulty that accompanies AI adoption across administrative, commercial, and operational functions. McKinsey has estimated that available technology could theoretically automate about 57% of hours worked in the United States; the gap between theoretical potential and actual implementation often depends on the quality of available procedures, data, and controls.
Gathering context is also a matter of trust
Software that observes screens, clicks, and keystrokes inevitably opens a sensitive chapter on privacy, security, and employee relations. Marble states that the information gathered is anonymized before it becomes visible to humans or AI models, even when processing takes place in the client's cloud. The startup also maintains that the app is restricted to enterprise applications, excluding private apps and tabs, as well as any activity outside working hours.
Users should be able to pause the platform at any time, while the analysis is reportedly performed on aggregated data and not with the goal of evaluating individual workers. These are important safeguards, but their practical effectiveness will depend on the configuration at each client, company policies, and applicable regulatory constraints. In European organizations, in particular, the use of tools capable of monitoring digital activity requires careful attention to the principles of data minimization, transparency, proportionality, and, where required, consultation with worker representatives.
For Marble, then, simply demonstrating that an agent can replicate a procedure is not enough. It will have to convince enterprises that the observation required to learn it can take place without turning into invasive surveillance and without exposing confidential information contained within workflows. It is a balance that involves passwords, customer data, financial documents, internal conversations, and all the content that can appear on a screen even during a seemingly simple task.
A crowded market, with a different entry point
The enterprise AI agent sector is increasingly crowded with model providers, orchestration platforms, robotic process automation software, and vertical products targeted at specific functions. In this scenario, Marble seeks to differentiate itself not by offering standard automation to be tailored to customer processes, but by deriving the process from observing the customer itself.
The premise has an obvious advantage: if the reconstruction is accurate, the agent can better align with the tools, exceptions, and practices already in use. However, there is also a structural limitation. Business processes change, people adopt different shortcuts, and a procedure replicated because it is frequent is not automatically the correct, efficient, or compliant one. The map produced by the software must therefore be validated by domain experts, while the transition from observation to execution requires adequate controls over the actions the agent is authorized to perform.
Marble cites EGYM, Mutares, and Enam among its clients, alongside other technology companies, private equity-backed organizations, and fund-backed AI roll-ups. For a young startup, these early deployments are above all a proving ground: they will have to show whether workflow observation truly cuts process analysis times and whether the agents can deliver reliable results once integrated into daily operations.
The arrival of Heartcore and Seedcamp adds two European investors with experience in early-stage software companies to the project. Commenting on the transaction, Seedcamp pointed to past investments in companies that transformed workplace tools, including UiPath and Synthesia. The comparison should be taken for what it is—an investor's take on a pre-seed bet, rather than a forecast of Marble's trajectory. Yet it clarifies the ambition: bridging the gap separating demonstrative AI from automation capable of embedding itself into enterprise processes.
The next test will be execution
With the fresh capital, Marble will need to build infrastructure, hire engineers, and tackle expansion into the United States. On a technical level, scaling inference is essential if the company aims to analyze and deploy agents across complex workflows for multiple clients. Commercially, the critical hurdle will be proving that the time and cost required to observe, map, and govern a procedure remain lower than the value delivered by the resulting automation.
The funding round also reflects a shift in how the market views AI agents. The ability to generate text or invoke an application is no longer enough to penetrate enterprise operating systems. It demands context, procedural memory, integrations, permissions, and error-handling mechanisms. Marble chooses to start precisely from that context, which is often dispersed across employees' daily routines. If it manages to turn that into verifiable automation without compromising confidentiality and human oversight, it will have unlocked a viable path into one of the most challenging frontiers of enterprise AI transformation.
