Marble Cognition launched with $3.75 million in pre-seed funding to build AI agent workers that first watch how employees work and then take over repetitive tasks. The round was led by Heartcore and Seedcamp, with angel investors including executives from McKinsey, Mistral AI and a board member of Datadog. The Munich startup positions observation before automation as the missing step for enterprise AI adoption. That framing matters because agents have become common, yet few companies can describe their own processes precisely enough to automate them.
How Marble maps work before automating it
The company was co-founded by German AI engineers Felix Rosner and Christophe Schmit, who argue that agents are already a commodity used by almost every enterprise in some form. Most current use centers on software writing, while Marble targets granular operational context that shows what should be built and automated. The startup says more than 90% of business processes live in messy email threads, scattered spreadsheets and employee knowledge rather than neat blueprints. A McKinsey study cited in the announcement estimates that 57% of US work hours could theoretically be automated, a level most firms have not approached.
Marble captures workflows at the source through a dedicated desktop application that observes employees across tools and processes. It takes screenshots and records clicks and keystrokes, then converts that activity into a structured process map that AI agents can read and execute in a code-like form. The company says the data is anonymized and can be reviewed before it reaches AI models. Processing can run inside a customer private cloud or on-premises environment. The system is also designed to recognize non-business computer use and exclude it, such as browsing Facebook during lunch.
The approach responds to a long-standing bottleneck in automation projects, where discovery and process documentation consume most of the effort. Seedcamp General Partner Reshma Sohoni, an early backer of UiPath, described mapping how work gets done as the hard part of automation. Marble places its product in that gap, supplying context for AI transformation programs. Its founders contrast this with deployments that start from assumed workflows and then require repeated correction. The result is positioned as a foundation layer: first record real execution, then assign autonomous work.
What this means for automation buyers
After mapping, Marble deploys systems called Coworkers that run selected repetitive processes end to end across existing software stacks. Humans stay in the loop where important decisions are required, with agents preparing work for review rather than acting alone. Early adopters named by the company include EGYM, Mutares and Enam. For smaller operations this model reduces the need for internal process analysts and bot builders. For larger firms it offers a way to standardize execution across teams without replacing current applications.
One cited deployment involved a private equity-backed US financial services firm where Coworkers prepared settlement offers for human review. Case-handling time fell by 90%, while staff processed around 50% more claims without additional hiring. Those figures describe a single customer case, not a guaranteed outcome across industries. Buyers should verify which process volume, data quality and exception rates produced the result. Key questions include how review workloads scale, how errors are detected, and what integration work each legacy system requires.
The test will be whether Marble converts pilots into repeatable deployments beyond the first named customers and one finance case. Watch for additional disclosed clients, measured time savings, and retention of Coworkers in production after six to twelve months. Expansion of supported applications and documented security reviews for private-cloud and on-premises setups would also signal maturity. If those markers appear, observation-first automation could become a practical procurement category for operations leaders.
