Munich-based Marble Cognition has raised €3.35 million ($3.75 million) in pre-seed funding to build AI coworkers that learn how employees work before automating their tasks. Heartcore Capital and Seedcamp co-led the round, announced on November 2, 2026. The approach targets a practical obstacle for enterprise AI: companies cannot easily automate processes they have never properly documented. EU-Startups reports the funding and company announcement.
Founded in late 2025 by Felix Rösner and Christophe Schmit, Marble plans to spend the money on scaling model inference, expanding its Munich engineering team and supporting a US launch. Its product starts with a desktop application that observes activity across business software, then turns that activity into structured process maps.
For companies evaluating agentic workflow tools, the interesting question is what that observation can teach an AI system, and where employees still need to supply context or approval.
Why Marble Starts With Observation
Marble’s premise is that employees leave important details out when explaining their work. A written procedure might describe the normal sequence while missing the exceptions, workarounds and judgment calls that determine what actually happens.
In Seedcamp’s November 2 announcement, Rösner estimates that around 90% of core processes at companies he has worked with exist in inboxes and people’s heads. That is a founder’s assessment of his engagements, not a representative survey of businesses.
The announcement coverage also cites a McKinsey estimate that current technology could theoretically automate about 57% of US working hours. These figures describe different things: one concerns documentation within Marble’s experience; the other concerns theoretical technical potential across working hours. They cannot be combined into a market forecast or a prediction of Marble’s results.
The useful connection is narrower. Even when technology can perform individual tasks, a company still needs to specify their sequence, inputs and exceptions. Marble is trying to make that discovery work part of the product.
As Rösner puts it in Seedcamp’s announcement:
“Our Coworkers watch how the work happens before they take it over.”
My view is that process discovery is the more consequential part of this pitch. Generating an action is only useful once the system understands when that action belongs in a workflow.
From Desktop Activity to a Process Map
Marble says its desktop application captures three types of activity inside business applications:
- Screen activity, providing visible context.
- Clicks, recording interactions with application controls.
- Keystrokes, capturing keyboard input.
The company says it converts observed activity into structured maps that AI systems can interpret. It then builds Marble Coworkers to execute those processes using the tools teams already use, with humans involved where judgment or approval is needed. Tech.eu describes this observation-to-execution sequence.
The distinction from a workflow that starts with written instructions is where the specification comes from. Marble proposes deriving it from actual activity. That could reduce the initial burden on employees who would otherwise have to explain every step.
However, capturing a click does not automatically explain its purpose. An employee might open a spreadsheet because it is an official source of truth, because another system is temporarily broken, or simply out of habit. Those possibilities imply different automation choices.
That makes the process map an important evaluation artifact. Buyers should check whether it captures meaningful branches and exceptions, and whether employees can correct its interpretation. The published descriptions do not detail the map format, editing controls or underlying model architecture.
I’d want to inspect a map before approving its resulting automation. A convincing replay of a common sequence is weaker evidence than a clear account of what happens when required information is missing.
Privacy Controls Cover Different Stages
Marble’s observation model requires close attention to what its desktop application can see. Screens and keyboard input can contain sensitive information even inside legitimate business applications.
According to Seedcamp’s announcement, Marble says captured information is anonymized before any person or AI model accesses it, including in a customer’s own cloud. It also describes exclusions, employee controls and aggregate analysis. These privacy and security claims have not been independently verified in the opened coverage.
| Company-described control | What buyers should establish |
|---|---|
| No recording of private apps, private browser tabs or activity outside working hours | How those boundaries are identified, configured and tested |
| Employees can pause recording at any time | What the pause stops and how employees confirm its status |
| Anonymization before human or model access | What information is removed, where processing happens and how failures are handled |
| Analysis in aggregate, not by individual | What administrators can retrieve, export or associate with an employee |
These controls address different stages. Excluding an application limits collection. Anonymization changes information after capture. Aggregate analysis limits how information is interpreted or presented. A company needs to examine each stage separately.
The anonymization claim makes the boundary around initial capture especially important. Buyers should ask whether raw captures are stored temporarily, how long records remain available and who can access them. The published material does not answer those questions.
A sensible deployment review would also test mixed-content situations, such as a personal message appearing inside a work application. An app-level exclusion policy needs to be assessed against what appears on screen.
Human Approval Matters More Than the “Coworker” Label
Marble describes humans staying involved where judgment matters. Its clearest published example is a private-equity-owned US financial services company, where a Marble Coworker prepares settlement offers for a specialist to approve.
That is a meaningful division of responsibility: the AI prepares the work, while a person authorizes the consequential decision. The example supports an approval-based workflow, although the public descriptions do not specify a universal approval requirement for every action or automation.
Frequently Asked Questions
3 questions
1Does Marble Need Written Workflow Documentation?
Marble says it learns workflows by observing business-app activity. Employees still need to validate whether the resulting process map accurately reflects the work.
