Marble Cognition raises $3.75 million in pre-seed to create AI collaborators that learn by observing – Unite.AI

The hardest part of automating office work often happens before an AI agent takes its first action: understanding what employees actually do. A workflow might traverse an inbox, a spreadsheet, and several business applications, with its exceptions understood by the team but never captured in a formal specification. Marble Cognition has raised $3.75 million in pre-seed funding to turn that operational knowledge into AI peers.

Heartcore and Seedcamp co-led the round. According to the announcement, Angel participants include a board member of Datadog, a founder of Personio, and executives from McKinsey, Mistral AI, and Lovable. The Munich company will use the funding to scale up inference, expand its engineering team and launch in the United States.

Marble’s approach begins with observing the work. Capture the flow of processes, turn those observations into structured maps, and use Marble Coworkers to run the resulting workflows in the tools your employees are already using. People remain involved where judgment or approval is required.

Transform unwritten workflows into an automation project

An agent who can navigate the software still needs to know which records to look at, what counts as a completed task, and when an exception requires escalation. Such instructions are difficult to provide when a company’s actual operating procedures exist in the form of habits, informal steps, and knowledge scattered throughout its systems.

Marble is betting that this context will become more valuable as software development and agent creation become easier. Its proposed benefit is an accurate understanding of the work to be automated: the sequence of actions, the information used, and the decisions that connect one step to the next.

The company says its desktop application records screen shots, clicks and keystrokes within enterprise applications. Use those observations to create structured process maps that AI systems can interpret, then create collaborators who run the process end-to-end. This links workflow discovery with implementation, rather than leaving the customer to translate process analysis into a separate automation project.

“Every company we talk to wants AI to take over real work. But no one can tell AI how that work actually gets done,” Felix Rösner, Marble’s co-founder and CEO, said in the announcement. “So let’s solve it first.”

AI collaborators operating across existing business tools

The existing Marble product website makes the execution model more concrete. Describes colleagues with native use of computers and browsers, allowing teams to delegate work in the environments they already use. His refund triage example combines a customer request, a refund policy, and company documents: The colleague classifies the request, reviews the policy, drafts a response, and sends exceptions to a person.

This example illustrates why context matters. Producing a plausible answer is just one step. A helpful colleague must connect the request with relevant rules and customer information, then distinguish routine work from a case that should not proceed automatically.

In operational terms, Marble is trying to bridge the gap between understanding a workflow and executing it. The process map provides a specification of the work; the colleague inserts it into the company applications; and human review preserves a decision point where the task requires it. The success of the deployment depends on the specification’s ability to capture exceptions and the common path.

The first customers and the implementation of financial services

Marble says it works with tech unicorns, private equity-backed companies and AI roll-ups, naming EGYM, Mutares and Enam among its clients. These organizations offer a natural testing ground for a product intended for repetitive, administrative-heavy tasks.

In one implementation at an unnamed private equity-owned U.S. financial services firm, Marble says a colleague reduced handle time per case by 90% and helped the team handle 50% more cases in the first month without additional hires. The colleague prepares transaction offers for a specialist to approve, after learning the workflow by observing the team.

These figures are company-reported results from a specific implementation, rather than a general performance benchmark. They describe two different outcomes: time spent on a single case and total team output, and specialist approval remains part of the process.

The example also highlights the commercial orientation of the company. Working with a repeatable structure and a clear approval boundary can provide an initial place to deploy an agent, while more ambiguous decisions remain with the team. Adapting this approach requires the system to recognize when a seemingly familiar case leaves the learned workflow.

Privacy controls are critical to observing work

A product that learns from employee activity needs clear boundaries around collection and access. Marble says its desktop application excludes private applications, private tabs, and after-hours activity. Employees can suspend enrollment, and analysis is performed in the aggregate rather than used to evaluate individuals.

The company also says that captured material is anonymized before a person or AI model sees it, even in deployments within the customer’s cloud. These are the guarantees declared by the company; they are fundamental to the operation of the system, rather than ancillary features of the product.

For an enterprise deployment, practical questions include which applications are in scope, how sensitive information is removed, and who can access information from the resulting process. Learning enough to reproduce a workflow while limiting unnecessary collection is a key engineering and governance challenge for this model.

Preliminary Capital for Inference, Engineering, and Expansion in the United States

Rösner and co-founder Christophe Schmit, CTO of Marble, met at the Center for Digital Technology and Management in Munich and founded the company at the end of 2025. Rösner previously worked at the sales agent company AI 11x. Schmit, a former MIT researcher, previously worked at Enam, according to the release.

The funding will support the infrastructure and engineering needed to transform observed workflows into deployable collaborators, alongside Marble’s planned launch in the United States. Seedcamp’s stake also connects the company with an investor who backed UiPath in the early stages of its development, putting Marble among a broader line of companies looking to reduce repetitive office work through software.

Marble’s thesis is that the next constraint on business automation is increasingly the quality of process knowledge an agent receives. His AI colleagues will be judged on how reliably they turn that knowledge into finished work and how well they preserve human judgment when the workflow calls for it.