This AI Learns Your Job by Watching You Do It

The AI that studies you before replacing you

Marble Cognition, a Munich-based startup founded in late 2025, raised $3.75 million in pre-seed funding for a product that sits somewhere between workplace assistant and workplace anthropologist. Its AI agents don't wait for instructions. They watch employees work — every click, screen transition, and workaround — then reconstruct those steps into process maps they can execute themselves.

The funding round closed in under two weeks, co-led by Heartcore and Seedcamp. What investors are backing is not just another automation tool. It's a system premised on the idea that most companies can't actually explain how work gets done in enough detail to automate it. So Marble's answer is: don't ask. Just observe.

Co-founders Felix Rösner and Christophe Schmit describe their "Coworkers" as AI agents designed to carry out entire white-collar processes end to end — inside the same business applications employees already use. The agents handle routine steps and escalate decisions that require human judgment. In one reported case, Marble's system helped a U.S. financial-services firm reduce case-handling time by 90%.

But to get there, the system has to study humans first.

How workplace knowledge becomes training data

Marble's desktop application records screen activity, derives standard operating procedures from what it sees, anonymizes the data, and deploys agents against the reconstructed workflow. It's designed to capture not just the official process, but the unofficial one — the workarounds, exceptions, and improvised judgments that experienced employees make without thinking.

That's the part that doesn't show up in manuals. A claims processor knows which fields can be skipped. A customer service rep knows which complaints need escalation. These details live in people's heads, accumulated through repetition and trial. Marble is betting that observation can extract them more accurately than interviews or consultants ever could.

The problem, as Rösner himself acknowledged, is that this requires employee buy-in. And asking workers to be monitored in detail — so that a machine can eventually perform their job — is inherently more contentious than rolling out ordinary productivity software. Rösner reportedly said change management, not the underlying AI, is the hardest part.

That's not surprising. Employees are being asked to participate in a system that simultaneously recognizes their expertise and makes their routines reproducible. The better someone is at a repetitive process, the more valuable their behavior becomes as training material. It's documentation and surveillance at once.

The quiet shift in enterprise automation

Most enterprise AI starts with a software interface: a chatbot, copilot, workflow builder. Marble starts with observation. That represents a broader shift in how automation enters the workplace.

The next generation of AI agents may not replace legacy software systems. They may simply imitate the humans who know how to navigate them. Instead of persuading companies to rip out old tools and adopt new platforms, agents can operate across the existing stack by learning the sequences people already use. The technology adapts to the organization, not the other way around.

That makes Marble compelling for reasons beyond the technology itself. It signals the moment when workplace knowledge stops being held primarily in employees' heads and starts being harvested from their screens — one routine, exception, and judgment call at a time.

The uncomfortable question is what happens when observation becomes the default. When the person training you is also learning how to do your job. When expertise is no longer just experienced — it's extracted.

Marble didn't invent that tension. But it did build a business model around it. And investors wrote a check in under two weeks.