Summary

• Marble Cognition has raised $3.75 million in pre-seed funding led by Seedcamp and Heartcore Capital.
• Its technology observes business processes, maps recurring tasks, and builds AI agents to perform selected activities.
• Early deployments involve EGYM, Mutares, and Enam, while the company is expanding its engineering and commercial operations.

Munich-based Marble Cognition has raised $3.75 million to develop AI agents that learn how organisations work by observing employees, aiming to automate recurring tasks without requiring companies to document every stage of their existing processes.

The pre-seed investment was co-led by Seedcamp and Heartcore Capital, with participation from angel investors linked to established technology businesses. Founded in late 2025 by Felix Rösner and Christophe Schmit, Marble is developing software that captures activities across business applications, identifies repeated processes, and creates automated workers capable of handling selected tasks.

The technology addresses an obstacle frequently encountered during automation projects: important operational knowledge often exists in employees’ routines rather than formal process documentation. Teams may understand how a task should be completed but struggle to describe every exception, approval, and workaround involved in carrying it out.

Marble begins by observing work through a desktop application that records selected interactions with business software. The resulting information is used to construct process maps that show how activities are performed across applications, after which the company can develop AI agents to execute suitable parts of those workflows.

Learning the details of everyday work

Conventional automation generally requires organisations to identify a process, document its stages, and establish the conditions under which software should act. That approach works well for stable, predictable activities, but it becomes more difficult where employees switch between systems or make frequent adjustments based on incomplete information.

Finance teams, for example, may receive documents by email, examine transactions in an accounting system, and record exceptions in spreadsheets. While individual steps can often be automated, the relationships between them may depend on knowledge accumulated by experienced employees rather than instructions that already exist in software.

Marble’s process discovery technology is designed to reveal those relationships through observed behaviour. Its system converts captured activity into structured representations of work, allowing organisations to identify repetitive tasks and determine which elements could be transferred to an automated agent.

Rösner estimates that approximately 90% of the core processes encountered in his previous work were not formally documented. That assessment reflects his experience rather than a representative industry survey, but it illustrates the problem Marble is attempting to address.

The resulting software differs from a conventional conversational assistant because it is intended to act within business applications rather than simply answer questions or produce text. An agent may retrieve information, complete a transaction, or prepare a document, although tasks requiring judgement or formal authorisation still need an appropriate approval process.

Marble says its agents can be tested against real cases before deployment and monitored after implementation. The company also describes a model in which employees remain responsible for reviewing selected actions, including decisions that fall outside the authority granted to an automated system.

Early customers provide operational examples

The business reports deployments involving EGYM, Mutares, and Enam, with applications across customer support, finance, administration, and other functions containing substantial repetitive work. Its investors have described a US financial services customer using an agent to prepare settlement offers for employees to review.

According to that account, the customer reduced handling time per case by approximately 90% and processed around 50% more cases during the first month without recruiting additional staff. The figures describe a particular customer deployment and have not been established as representative outcomes across Marble’s wider business.

The example nevertheless illustrates the potential division of work between software and employees. An agent can prepare information or complete preliminary steps, while a person retains responsibility for reviewing a financial decision before it is finalised.

Such arrangements require more than accurate execution of a recorded workflow. Organisations must establish whether the agent has access to the necessary information, which actions it may perform, and how errors or unusual circumstances should be handled. A process that changes frequently may also require regular adjustments to prevent the automation from following outdated practices.

Observation creates responsibilities

Marble’s initial process discovery relies on observing work carried out by employees, introducing data protection and workplace monitoring considerations. The company says users can decide when recording is active, while personal information is redacted before processing and analysis is performed in aggregate.

Its published controls include European hosting options and arrangements that allow personal information to be anonymised within a customer’s own environment. These features may assist organisations with data protection requirements, although employers remain responsible for determining the lawful basis and proportionality of any monitoring they introduce.

The UK’s Information Commissioner’s Office identifies screen capture and keystroke monitoring as particularly intrusive forms of worker monitoring. Organisations using such technology must explain its purpose, consider the effects on employees, and assess whether the information collected is necessary for the intended activity.

Observing a process also creates the possibility of reproducing existing weaknesses. An employee may have access permissions that extend beyond what an automated agent should receive, while informal workarounds developed over time may no longer represent an appropriate procedure.

Marble’s process maps can inform automation design, but permissions and approval requirements must be established independently of the behaviour recorded. Businesses also need mechanisms for reviewing how agents perform after deployment, particularly when applications, regulations, and internal procedures change.

Turning process discovery into a business

Marble intends to use the funding to expand its engineering team in Munich, improve its process inference technology, and support entry into the US market. Its initial customers provide experience across different operational settings, although wider deployment will require the technology to accommodate variations in software, business processes, and internal controls.

The company is entering a competitive market in which enterprise software suppliers and specialist startups are developing agents capable of acting across business applications. Its distinction is the attempt to establish which work should be automated through observation before constructing the agents responsible for carrying it out.

As those systems assume more responsibility, organisations will need evidence of completed tasks, error rates, employee oversight, and the cost of maintaining automation over time. Marble’s customer deployments will provide further opportunities to measure whether process discovery reduces the effort involved in implementing reliable business automation.