Marble raises $3.75M to teach AI agents by watching people work
What's the deal? Marble CognitionDealroom has a profile for this one. Try Dealroom →, a Munich startup founded in late 2025, has raised a $3.75 million pre-seed round co-led by Heartcore and Seedcamp. It records how teams work, maps those processes, and then builds AI agents to automate them.
Why now? Co-founder Felix Rösner told Tech Funding News the round closed in under two weeks. The angel list includes a Datadog board member, a Personio founder, and executives from McKinseyDealroom has a profile for this one. Try Dealroom →, Mistral AI, and Lovable. Marble has not disclosed a valuation.
What's the endgame? Marble cites a November 2025 McKinsey Global Institute report finding current technology could in theory automate about 57% of US work hours. The company argues technology is not the barrier — most core processes "live in people's inboxes, on spreadsheets, and in people's heads." Its desktop app records screen activity in business apps, turns it into a process map, then deploys an agent that runs the task end to end, with humans stepping in where judgement is needed.
The results so far. Marble says its agent helped a US financial firm cut case-handling time by 90%. Named customers include EGYM, Mutares, and Enam. The company also works with PE firms in Europe and PE-owned professional services firms in the US, and plans to split its customer base evenly between the two regions.
Going stateside. That US push puts Marble in a fast-forming category. Munich-based Interloom raised $16.5 million in March for a related approach, while Skan AI raised $63 million in August to expand its workflow-intelligence platform. Marble's differentiator is its end-to-end loop, from observing desktop work to deploying agents in existing tools.
What could go wrong? Watching employees work raises privacy and consent questions. Marble says it records only business apps, never private apps, tabs, or activity outside working hours, and employees can switch recording off at any time. Data is anonymised and stored in Europe — or in Germany for German customers — before any model sees it, and staff can view, delete, or block their own data.
The signal: Investors are betting that the scarce input for automation is not the models, which are commoditising, but a precise record of how work actually gets done. "Marble learns a whole company's processes, anonymises the data, derives SOPs," said Max Niederhofer, partner at Heartcore. In a crowded market, the real test will be substantiated customer outcomes, not the promise that agents can learn by watching.
Read more: Tech Funding News
Image credit: Marble Cognition
