Marble Cognition Secures $3.75M to Train Enterprise AI Agents via Workflow Observation
German AI startup Marble Cognition raises $3.75 million in pre-seed funding to build desktop-monitoring AI agents that observe daily enterprise workflows and convert them into executable processes.
Enterprise interest in autonomous artificial intelligence agents has surged, but many organizations still struggle to apply these tools to their core business operations. Addressing this operational gap, Munich-based startup Marble Cognition has launched with $3.75 million in pre-seed financing. The investment round was led by Heartcore and Seedcamp, with participation from angel investors including executives at McKinsey & Co., Mistral AI, and a board member from Datadog Inc. Co-founded by German AI engineers Felix Rösner and Christophe Schmit, Marble aims to solve the missing context problem that prevents AI agents from executing complex enterprise tasks.
Mapping messy business workflows at the desktop level
While software development has emerged as a primary use case for AI agents, broader enterprise automation has stalled. According to research from McKinsey & Co., up to 57% of all working hours in the United States could theoretically be automated. However, most companies fall far short of that figure because more than 90% of business processes remain uncodified—trapped inside informal email exchanges, disconnected spreadsheets, and employee memory.
To solve this, Marble has developed a dedicated desktop application designed to observe and map how employees carry out daily work across various tools. The software monitors worker activity by taking screenshots and capturing keystrokes and clicks. It then processes this raw behavioral data into a structured process map that AI models can interpret and execute much like traditional code.
To address corporate privacy and compliance concerns, Marble incorporates several safeguards. Recorded workflow data is anonymized and can be reviewed before being processed by any AI models. Customers also have the option to execute model processing entirely within their own private cloud or on-premises infrastructure. Furthermore, the application is trained to recognize non-work computer usage—such as an employee checking Facebook during a lunch break—and automatically excludes those activities from its data collection.
"Every company we talk to wants AI to take over real work," explained co-founder Felix Rösner. "But nobody can tell the AI how that work is actually done. So we solve that first. Our Coworkers watch how the work happens before they take it over. They do it the same way your best people do."
Deploying AI 'Coworkers' for real-world enterprise tasks
Once workflow data is converted into structured process maps, Marble deploys specialized AI agents called "Coworkers." These agents run repetitive processes end-to-end directly within an enterprise's existing software stack, keeping human employees in the loop to approve critical decisions.
Early adopters of the technology include organizations such as EGYM GmbH, Mutares SE & Co. KGaA, and Enam GmbH. The startup has already demonstrated significant practical impact: a private equity-backed U.S. financial services provider used Marble's Coworkers to generate settlement offers for human review. By automating the underlying process, the firm reduced case-handling times by 90%, enabling its staff to process approximately 50% more claims without increasing headcount.
Reshma Sohoni, General Partner at Seedcamp and an early backer of robotic process automation leader UiPath Inc., highlighted the importance of workflow visibility. "The hard part has always been mapping out how the work gets done, so you can actually automate it," Sohoni noted. "Marble fills that gap, giving companies vital context for AI transformations that could define their future."
What it means for developers
For software engineers and AI builders, Marble's approach underscores a fundamental shift in AI agent architecture: moving beyond static prompts and synthetic benchmarks toward real-time operational context capture. As AI agents move into production, feeding them accurate, structured context is becoming as critical as the underlying model choice. Developers building complex automation pipelines, agentic workflows, or background monitoring tools can test and compare different LLM backends cheaply through a single unified API at https://apixoai.online.
Furthermore, Marble's model demonstrates how developers can bridge the gap between computer vision, telemetry, and structured process generation. By converting desktop actions into executable code-like maps, the startup provides a blueprint for building agents that interact with enterprise software without relying on fragile custom integrations for every legacy system.
Source: Marble Cognition raises $3.75M for its AI agent workers that watch and learn how people work — SiliconANGLE AI. Written by the Apixo team from that report.
One key for Claude, GPT, GLM, DeepSeek and more. Pay per token with crypto.
Get your API key
