Demand for AI that performs genuine work is near universal among large organisations. What almost none can supply is an account of how that work gets done today.
Marble Cognition, founded in Munich in late 2025 by Felix Rosner and Christophe Schmit, raised $3.75M in a pre-seed round co-led by Heartcore and Seedcamp. Angels include a Datadog board member, a Personio founder and executives from McKinsey, Mistral AI and Lovable.
The company starts from an uncomfortable statistic it cites from its own customer base: about 90% of core business processes are not formally documented. Someone knows them, and that someone eventually leaves.
Marble’s desktop application observes how employees move through business software, capturing screens, clicks and keystrokes, and turns that into a structured process map. Agents called Marble Coworkers then execute the process with the tools already in place, with humans approving anything that needs judgement.
Privacy is the obvious objection, and the design answers it in specifics. Data is anonymised, analysed in aggregate rather than per person, skips private apps and private browser tabs, stops outside working hours, and employees can pause monitoring themselves.
Customers already running it include EGYM, Mutares and Enam, plus other technology companies and private-equity-backed businesses. McKinsey reckons existing technology could automate about 57% of US working hours, which is the prize if the documentation problem gets solved.
The funding goes to scaling inference capacity, expanding the Munich engineering team and launching in the United States.
The bet underneath is that process discovery, not model quality, is the real bottleneck in enterprise automation, and that whoever owns the map owns the agents built on top of it.
