Deals · Enterprise AI

Marble Cognition raises €3.35M pre-seed to build AI coworkers that learn your job by watching you do it

A Munich Enterprise AI startup that learns how employees work by observing business-app activity, maps the implicit logic into structured process descriptions, and deploys AI coworkers to automate those workflows end to end.

ProYarn Desk · November 2, 2026Read this in French

in 𝕏 ✆

Marble Cognition (Munich) has raised €3.35 million ($3.75 million) in a pre-seed round co-led by Heartcore Capital and Seedcamp. The round also includes angels with board-level or executive backgrounds at Datadog, Personio, McKinsey, Mistral AI, and Lovable. The company — founded in 2025 by Felix Rösner (CEO) and Christophe Schmit (CTO), who met at Munich's CDTM programme — is building what it calls Marble Coworkers: AI agents that learn business processes by observing employees at work, then automate those processes on their behalf.

The observation model

The dominant approach to enterprise AI automation today starts from an API or a structured process description: a workflow is documented, a model is instructed to execute it, and the implementation work involves mapping those instructions to the target applications. The failure mode is familiar — processes as documented differ from processes as practised, and the gap grows with process complexity.

Marble's starting premise is that the relevant source of truth is not documentation but observation. Its desktop application runs in the background during normal working hours, recording screen activity, keystrokes, and clicks exclusively within business software (email, ERPs, CRMs, internal tools). From this stream of observed behaviour, the system infers the structure of the workflow: the decision points, the sequence of actions, the edge-case handling. That inferred structure becomes the training material for the Marble Coworker, which then executes the same workflow in the same tools.

The company states that all captured data is anonymised before any person or AI model processes it, that recording excludes private applications and private browser tabs, and that employees can pause observation at any time. Initial customers include EGYM (fitness tech), Mutares (private equity), and Enam.

What the €3.35M funds

The pre-seed capital funds three parallel priorities: scaling the AI inference infrastructure that processes observed screen activity into structured process maps; growing the Munich engineering team; and preparing for a US market launch. The company's early customer base — three enterprise or mid-market organisations with distinct workflow libraries — gives it a range of process types to train against before the US expansion.

Rösner previously worked at AI sales-agent company 11x. Schmit's background is in enterprise software architecture. Heartcore and Seedcamp — backers of Revolut, UiPath, Neo4j, and others — are co-leading a round that is early even by pre-seed standards.

The threshold question is the privacy model. Marble's observation approach is technically more direct than retrospective process mining, but it is also more legally exposed, particularly in Germany where the Works Constitution Act gives workers' councils co-determination rights over monitoring-adjacent tooling, and GDPR imposes additional consent requirements. Whether the company's stated anonymisation and consent model is sufficient for large enterprise deployment — as opposed to early adopter clients — is a governance question the funding does not answer.

Why it matters to investors

Enterprise workflow automation has a long record of promising more than it delivers. Business process automation, robotic process automation, and process mining tools have all offered to codify how organisations work, and all have struggled at the gap between observation and execution — the cost of translating what people actually do into something an automated system can reliably replicate. Marble Cognition's claim is that observing in real time, rather than reconstructing from logs or interviewing practitioners, closes that gap. The distinction matters: if the implicit logic of a workflow — the judgements made, the steps skipped, the corrections applied — is only visible while the work is happening, then retrospective analysis systematically misses it. The privacy question is the round's most material open issue. Recording screen activity and keystrokes in a workplace is legally and structurally complex under German law (Works Constitution Act, GDPR, employee data-protection regulation), where workers' councils hold meaningful co-determination rights over surveillance-adjacent tooling. The company has stated that data is anonymised before any person or model sees it, that recording covers only business applications, and that employees can pause it. Whether those measures satisfy the statutory and co-determination requirements that apply at German enterprise customers — rather than smaller early adopters like Enam — is a question regulators and workers' councils will evaluate independently of the investor deck. Heartcore and Seedcamp are credible judges of early enterprise software; their decision to co-lead at pre-seed suggests conviction that the legal risk is manageable. The next 18 months will test whether the approach holds for complex workflows, at scale, under the governance structures of large organisations.

Sources

Pre-Seed Enterprise AI Germany Heartcore Capital Seedcamp

Threaded to this story

Enterprise AI · July 9, 2026

Every European round, in your inbox by 8am.

The day's seed and Series A rounds across France and Europe — threaded, sourced, and read in two minutes. Free.