Marble Cognition raises $3.75M to build AI coworkers

By Amara Osei

Marble Cognition has raised $3.75 million in pre-seed funding to build AI coworkers that learn workplace workflows by observing employees use business software. The startup is based in Munich.

The Munich-based company said Heartcore and Seedcamp co-led the round, with angel investors including a Datadog board member, a Personio founder and executives from McKinsey, Mistral AI and Lovable. Marble said the capital will scale inference, expand its engineering team and support a launch in the United States.

Marble targets the step before an AI agent acts. The company says an agent must know which records to consult, what counts as a completed task and when an exception requires escalation. A workflow may cross an inbox, a spreadsheet and several business applications, while its exceptions remain understood informally but never captured in a formal specification, according to Marble.

Those rules often live in habits, informal handoffs and knowledge scattered across systems, according to Marble. The company says this context becomes more valuable as software development and agent-building become easier.

In the announcement, Marble Cognition’s co-founder and CEO Felix Rösner said, “Every company we talk to wants AI to take over real work. But nobody can tell the AI how that work is actually done,” and “So we solve that first.” Marble’s approach starts with observing how processes unfold.

Marble says its proposed advantage is an accurate understanding of the sequence of actions, information used and decisions connecting one step to another. Its desktop application records screens, clicks and keystrokes inside business applications and turns them into structured process maps that AI systems can interpret.

Marble says it then deploys Marble Coworkers to carry out the mapped work in tools employees already use. The company says this connects workflow discovery with deployment and avoids making customers translate a process analysis into a separate automation project.

People remain involved where judgment or approval is required, according to Marble. The company says the resulting specification lets a coworker carry a workflow into business applications while human review preserves a required decision point.

Marble’s product website says its coworkers have native computer and browser use, letting teams delegate work in environments they already use. Its refund-triage illustration combines a customer request, a refund policy and business records. The coworker classifies the request, checks the policy, drafts a response and sends exceptions back to a person, according to the website.

Marble says generating a plausible reply is only one step; a coworker must connect the request with relevant rules and customer information. It separates routine work from cases that should not proceed automatically, while people approve decisions that require judgment. Successful deployment depends on capturing exceptions as well as the common path, according to Marble.

Marble says it works with technology unicorns, private-equity-backed companies and AI roll-ups, naming EGYM, Mutares and Enam as customers. Those sectors provide what Marble describes as natural testing grounds for repetitive, tool-heavy administrative work.

In one deployment at an unnamed private-equity-owned U.S. financial-services company, Marble says a coworker cut handling time per case by 90%. During its first month, the customer handled 50% more cases without additional hires, according to the company. The coworker learned the workflow by observing the team, prepares settlement offers and sends them to a specialist for approval.

Marble describes those figures as results from one deployment, not a general performance benchmark. The 90% figure measures time per individual case, while the 50% figure measures total throughput; specialist approval remains part of the workflow.

The case points to Marble’s commercial focus: repeatable work with a clear approval boundary can provide an initial place to deploy an agent, while more ambiguous decisions remain with the team. Scaling will require the system to recognize an apparently familiar case outside the learned workflow, according to Marble.

Marble says a product that learns from employee activity needs defined limits on collection and access. Its desktop application excludes private applications, private tabs and activity outside working hours.

Employees can pause recording, and the system performs analysis in aggregate rather than using activity to assess individuals. The company also says captured material is anonymised before a person or AI model sees it, including in deployments inside a customer’s own cloud.

Marble presents those limits as core safeguards rather than incidental product features. The company says enterprise deployment still requires decisions on which applications are in scope, how sensitive information is removed and who can access the resulting process information.

Those questions sit at the center of the model, according to Marble, because the system must learn enough to reproduce work while limiting unnecessary collection.

Co-founder and CEO Felix Rösner and co-founder and CTO Christophe Schmit met at Munich’s Center for Digital Technology and Management before founding Marble in late 2025. According to the announcement, Rösner previously worked at AI sales-agent company 11x, while Schmit was a former MIT researcher and previously worked at Enam.

Marble said the financing will support infrastructure and engineering needed to turn observed workflows into deployable coworkers, alongside the United States launch. The announcement also notes that Seedcamp backed UiPath early in its development, linking Marble to a wider group of companies pursuing office-task automation.

Marble’s thesis is that enterprise automation’s next constraint will increasingly be the quality of process knowledge given to agents. Marble says its AI coworkers will be judged on whether they complete work reliably and preserve human judgment where workflows require it.

Marble’s planned United States launch will test whether that model can reproduce the reported deployment gains while maintaining the stated privacy boundaries.