Pomegra
Munich-based Marble Cognition raised a $3.75M pre-seed led by Heartcore and Seedcamp to build AI agents that learn workflows by watching employees work.
Key Takeaways
- Marble Cognition closed a $3.75M pre-seed on November 2, 2026, co-led by Heartcore and Seedcamp; valuation is undisclosed.
- The company records screen activity to map how work is done, then deploys AI agents to run those processes.
- Customers include EGYM, Mutares and Enam, with a US launch and a larger Munich engineering team planned.
Lead
Marble Cognition, a Munich startup founded in late 2025, announced a $3.75M pre-seed round on November 2, 2026. Heartcore and Seedcamp co-led the round. Angels include a Datadog board member, the founder of Personio, and executives from McKinsey, Mistral AI and Lovable. The company did not disclose a valuation.
Marble sells what it calls "Coworkers": AI agents that take over business processes after first observing how humans perform them. The pitch targets a gap that stalls many enterprise AI projects, which is that companies rarely have an accurate written record of how their work actually gets done.
What Does Marble Actually Build?
Marble builds a desktop application that records employee activity inside business software, including screens, clicks and keystrokes. The system anonymizes that data before processing, then converts it into structured process maps and standard operating procedures. Agents then run those processes end to end across the tools a company already uses.
Human approval points stay in the loop. When a task requires judgment, the agent escalates to a person instead of guessing. That design choice matters for regulated customers, where an unreviewed automated decision carries legal exposure.
CEO Felix Rösner, previously at 11x AI, framed the problem in the announcement: every company wants AI to take over real work, but nobody can tell the AI how that work is done. CTO Christophe Schmit is a former MIT researcher who worked at Enam. The two met at Munich's Center for Digital Technology and Management.
Who Is Paying for It?
Marble names EGYM, Mutares and Enam as customers. It also works with large technology companies in Germany, private equity firms in Europe, and PE-owned professional services firms in the US. In one deployment at a US financial services firm, Marble reports a 90% drop in case-handling time and a 50% rise in team throughput within the first month.
Those figures come from a single deployment and the company itself. Pre-seed metrics of this kind show a working pilot, not repeatable performance across industries. Whether the results hold in less structured back-office work is the open question.
Why Did Heartcore and Seedcamp Back a Pre-Seed This Size?
A $3.75M pre-seed is large by European standards, and the size reflects the cost of the product as much as investor appetite. Marble plans to spend on inference capacity, since running agents against live workflows consumes compute continuously. The rest goes to engineering hires in Munich and a launch in the United States.
The investor list also tells a story. Angels from Mistral AI and Lovable bring model and product-building credentials, while the Personio and Datadog connections point to European software go-to-market experience. For a company planning to sell to PE portfolio firms on two continents, those introductions carry practical value beyond the cheque.
How Crowded Is the Workflow-Mapping Market?
The category is filling quickly. Skan AI has raised a $63M Series C on process discovery. Interloom raised a $16.5M seed in March 2026, Duvo a $32M Series A, and Tekst €11.5M. Marble enters with a fraction of that capital and a company history measured in months.
Its differentiator is combining observation and execution in one product. Process-mining vendors have long mapped workflows without running them, while robotic process automation incumbents such as UiPath automate tasks that someone scripted by hand. Marble argues that the map should write the automation. That is a credible claim, but incumbents with large customer bases can add similar observation features, and the window to establish a lead is likely short.
What Are the Risks?
Recording employee screens raises privacy questions, particularly in Germany, where works councils hold real influence over workplace monitoring. Marble's anonymization step addresses part of this, but each enterprise deployment will still need employee consent and internal sign-off. That friction could slow sales cycles in exactly the markets where Marble is based.
Accuracy is the second constraint. An agent that learns from observed behavior also learns bad habits and workarounds. Customers will want evidence that Marble separates the intended process from the one people improvise.
Outlook
Marble Cognition has secured funding, a named customer base and a clear product thesis, but also a crowded field and a monitoring model that invites scrutiny. The next test is whether the claimed gains survive outside a handful of deployments. A US launch and a larger Munich team will show over the coming quarters whether observation-first automation can scale beyond early adopters.
