- Marble Cognition has raised a $3.75 million pre-seed round co-led by Heartcore and Seedcamp.
- The Munich startup tracks how teams work and then builds AI agents to automate those processes.
- Marble says its AI agent helped a US financial firm cut case-handling time by 90%.
For an AI agent to take over a job, the process has to be well defined. Most organisations don’t have that. Marble Cognition, founded in Munich in late 2025, raised $3.75M to address this by observing people at work.
Heartcore and Seedcamp co-led the pre-seed round. The angel investor list includes a Datadog board member, a Personio founder and execs from McKinsey, Mistral AI and Lovable. Marble has not shared a valuation. Co-founder Felix Rösner told Tech Funding News the round closed in under two weeks.
“Every company we talk to wants AI to take over real work. But nobody can tell the AI how that work is actually done. So we solve that first. Our Coworkers watch how the work happens before they take it over. They do it the way your best people do,” said Rösner.
Two founders, one undocumented problem
Marble cites a November 2025 McKinsey Global Institute report stating that, in theory, current technology could automate about 57% of US work hours. Marble argues that technology is not the primary barrier. In most of the companies they work with, 90% of core processes live in people’s inboxes, on spreadsheets, and in people’s heads.
Rösner believes this is why so many enterprise AI initiatives get off to a bad start: “The problem is that everyone is starting with a solution. The vendors will typically come in and say, ‘ This is the use case; this will save you X number of hours a month. The reality is they have no idea whether that use case is a real problem, how it works, or what the potential savings are.”
Marble takes a different approach. Marble starts with the current state and only builds where there’s a genuine opportunity. Founded in late 2025 by Rösner and Christophe Schmit, the startup uses a desktop app to record screen activity in business apps and turn it into a process map AI can understand.
The company then creates a Coworker that can execute the process from beginning to end in the team’s existing tools, with humans able to step in where judgement is required. Marble’s approach takes a traditional automation model, which adapts a standard product to the business process, and flips it on its head to focus on deeply understanding the process itself.
Marble’s named customers are EGYM, Mutares and Enam. Rösner added that Marble works with big tech unicorns in Munich and Germany, PE firms in Europe, and PE-owned professional services firms in the US. The ambition is to split the customer base evenly between the US and Europe.
The privacy question
Marble is very clear about the boundaries of what it will record in terms of employee workflows. All data is anonymised before anyone, including any AI model, can access it, even within the customer’s own cloud. It records only business apps, never private apps, never tabs, and nothing outside working hours. Employees can turn recording off at any time, and Marble looks at the data in the aggregate rather than on an individual basis.
Rösner tells TFN Marble’s approach has two parts. The first is compliance and security. Marble strips private data from the customer’s local network before it reaches the LLM and stores it in Europe, or Germany for German customers. The second is adoption. Employees at each customer can see the analysed data and outputs; they can delete their own data and block certain apps. Security and certifications matter, but so does earning the customer’s trust through transparency and giving them full control of their data.
Max Niederhofer, partner at Heartcore, said: “The scarce input for automation is not the models, which are commoditising. It is building a precise record of how work gets done. Marble learns a whole company’s processes, anonymises the data, derives SOPs [standard operating procedures] and deploys agentic AI against them … What Felix and Christophe are building is a key technology for the AI-enabled workplace.”
Marble’s home market will test those promises. German law allows works councils to have a say in any technical systems that are monitoring employee behaviour or performance. The EU AI Act also classifies AI systems which are designed to monitor and assess employee performance and behaviour as high risk.
A crowded field, and what the money is for
Rösner sees three types of competitors. The first are consultancies that do what Marble does by hand: they talk to everyone in the organisation, find the problems, and then come up with fixes. The second are the vertical AI companies who send their own engineers in to do the same thing and then build the solution. The third are the software companies that do process mapping.
Others are after the same budgets. Skan AI in Menlo Park raised $63 million in a Series C, co-led by Cathay Innovation and Dell Technologies Capital, and launched AI agents at the same time. Interloom in Munich raised $16.5 million in a seed round in March 2026, led by DN Capital, to build a memory layer for AI agents from operational data. TFN has also covered Duvo, founded by Rohlik’s founder, which raised $32M from Balderton to help enterprise AI get beyond the pilot phase, and Tekst in Belgium, which raised €11.5M to automate back-office workflows.
Reshma Sohoni, general partner at Seedcamp, added, “Seedcamp has a long tradition of backing founders who change how we work, from UiPath and Synthesia to Marble today. Most companies want to automate more, but the hard part has always been mapping out how the work gets done. Marble fills that gap, giving companies vital context for AI transformations that could define their future. We’re proud to back them from the start.”
The money will fund scaling inference, growing the engineering team in Munich, and launching operations in the US. Team building is the top priority, followed by speed.
The real challenge is getting people and AI workers to work together. Marble’s approach is predicated on employees agreeing to be observed. Rösner says change management is the trickiest part of the process, as monitoring staff will always be more contentious than most tools. The first works council’s reaction to these steps will tell you whether they are sufficient.