Munich startup Marble Cognition has raised a $3.75 million pre-seed round to build AI that learns how a company works by watching employees do it. Coverage from Tech Funding News, Tech.eu and EU-Startups says the round was co-led by Heartcore Capital and Seedcamp. The pitch is simple to state and hard to deliver: record the work, turn it into a process map, and hand that map to AI agents that carry out the job.
The company was founded in 2025 by Felix Rösner (co-founder and CEO) and Christophe Schmit (co-founder and CTO). One report says “late 2025,” but the exact founding date is unconfirmed. Its named customers so far are EGYM, Mutares and Enam. The valuation has not been disclosed, and the closing date of the round is unconfirmed as well.
Marble lands in a crowded week for agent funding. We covered a similar bet last month when Minicor raised $3.8M to point AI agents at legacy desktop software. Marble’s angle differs in one important way. It wants the agent to learn the process from people first, rather than asking a human to write the instructions.
Don't miss new tech stories on Google
Add Tech Insider once in the Google app and our stories appear in your news suggestions.
What Marble Cognition Announced
According to the reports, Marble sells a desktop application that sits on an employee’s computer and observes how work happens inside business software. It captures screenshots, mouse clicks and keystrokes. The output is a structured process map, a document built so an AI system can read it and then execute the steps.
The agents that run those maps are called “Coworkers” or “Marble Coworkers” in the coverage. They are designed to run a process end to end inside the software a company already uses. When a step needs judgment, a human can step in. That human-in-the-loop design matters because it frames the product as a handover tool, not a replacement for staff on day one.
Marble says it will spend the money on three things: scaling inference, expanding its Munich engineering team, and launching in the United States. None of the outlets we reviewed reported a launch date for the US push, so we treat it as a plan rather than a milestone.
The Deal at a Glance
The table below collects every figure the reports confirm. Anything missing from it, including valuation and the closing date, is not public as of November 02, 2026.
| Item | Detail | Status |
|---|---|---|
| Company | Marble Cognition, Munich, Germany | Confirmed |
| Round | $3.75 million pre-seed | Confirmed |
| Lead investors | Heartcore Capital and Seedcamp (co-leads) | Confirmed |
| Founders | Felix Rösner (CEO), Christophe Schmit (CTO) | Confirmed |
| Named customers | EGYM, Mutares, Enam | Confirmed |
| Founded | 2025 (“late 2025” in one report) | Exact date unconfirmed |
| Valuation | Not disclosed | Unknown |
| Closing date | Not reported | Unconfirmed |
Reports also say angels in the round included a Datadog board member, a Personio founder, and executives from McKinsey, Mistral AI and Lovable. The word reportedly applies here, since individual names were not part of the confirmed details.
How the Marble Desktop App Works
From keystrokes to a process map
The recording layer is the unglamorous part. A desktop client watches an employee move through ERP screens, spreadsheets, email and web tools. It logs what the screen showed, where the mouse went and what was typed. On its own that stream is noise. A person doing an invoice check might open eight tabs, copy two numbers and ignore a third, and nothing in the raw log says why.
The product’s value sits in the translation step. Marble turns the recording into a structured map that names the steps, the systems touched and the decision points. That map is what the AI reads. Marble has not published how accurate the translation is, and none of the three outlets reported a benchmark. Buyers should ask for error rates on their own workflows before signing anything.
Where Coworkers take over
Once a map exists, a Coworker runs it in the same software the human used. This is the opposite of an API-first integration. No custom connector is needed, which is why the approach appeals to firms with old systems that expose no clean interface. The trade-off is fragility. Agents that drive a screen can break when a vendor moves a button.
The human-intervention path is the safety valve. When a case falls outside the recorded pattern, the agent is meant to stop and ask. How often that happens in practice will decide whether customers see real labor savings or just a new queue of exceptions.
Why Undocumented Work Is the Real Bottleneck
Most companies run on habits. A senior accountant knows which supplier invoices need a second look. A support lead knows which refund requests are really fraud. None of it sits in a wiki. Marble’s founders argue this missing knowledge, not model quality, is what stops enterprises from handing real work to AI.
One report cited a statistic that roughly 90% of companies’ core processes are undocumented. We could not confirm that number, so we do not rely on it. The underlying point still holds in plain terms. If you cannot describe a task, you cannot instruct an agent to do it, and the best language model in the world will guess.
This is a different problem from the one most agent vendors pitch. Many tools assume a manager can write a clear prompt or workflow. Marble assumes the manager cannot, and that the only honest record of the job is the job itself.
What the Founders Say
Rösner framed the problem as a gap between ambition and instructions. In his words: “Every company we talk to wants AI to take over real work. But nobody can tell the AI how that work is actually done.”
He described the product’s answer in a second statement: “Our ‘coworkers’ watch how the work happens before they take it over. They do it the way your best people do.”
Read closely, the second line makes a claim worth testing. “The way your best people do” implies Marble can pick up good habits and skip bad ones. A recorder sees what people do, not which behavior is best. If a team’s top performer and its slowest member both get recorded, the system needs some way to tell them apart. Marble has not said how it handles that, and the reports do not explain it.
Who Is Backing Marble and Who Is Buying
Heartcore Capital and Seedcamp co-led the round. Both are established European early-stage investors, and a co-lead structure at pre-seed signals that two funds wanted the allocation. The angel list, as reported, mixes go-to-market operators with AI builders. A Datadog board member and a Personio founder bring software-sales and HR-tech experience. Executives from Mistral AI and Lovable bring model and product knowledge. McKinsey adds enterprise consulting reach.
On the customer side, the reports name EGYM, Mutares and Enam. They span fitness-tech, private equity and another business line. We do not have contract sizes, pilot counts or go-live dates for any of them. A named customer in a pre-seed announcement usually means a pilot or an early deployment, so readers should not assume production scale.
Can AI Really Learn Your Job by Watching?
The headline question has a narrow honest answer. Marble’s confirmed claim is that it observes and automates business processes. It has not claimed that its software can absorb an entire job simply by watching it, and the reports do not support that reading. A job includes negotiation, context and relationships that no screen recording captures.
| Claim | What the reports support | Our read |
|---|---|---|
| Marble watches work in business software | Yes: screenshots, clicks and keystrokes | Confirmed |
| It produces a process map for AI | Yes: structured map AI can read and run | Confirmed |
| Coworkers execute processes end to end | Described as the design goal, humans can intervene | Design claim, no public benchmark |
| AI learns an entire job by watching | Not stated in confirmed descriptions | Unconfirmed |
| About 90% of core processes are undocumented | Cited in one report | Unconfirmed |
The distinction between a process and a job is the whole story here. A process like “match this purchase order to this invoice and flag differences” is repetitive and rule-shaped. A job like “manage supplier relationships” bundles dozens of such processes with judgment calls between them. Marble’s pitch targets the first kind, and that is where observation has a real chance of working.
Competitive Comparison: How Marble Stacks Up
Marble competes less with chatbots and more with three other ways of getting work into an agent’s hands. The first is hand-written instructions. The second is classic automation scripts. The third is connecting agents directly to apps, as in Claude Duet, which links six work apps for Pro and Max users. The table compares them on how each learns a task.
| Approach | Who defines the task | Works with old software | Main weakness |
|---|---|---|---|
| Marble-style observation | The recording of real work | Yes, drives the existing screen | Records bad habits along with good ones |
| Prompt or workflow written by a manager | A person who must know the process | Depends on connectors | Managers often cannot describe tacit steps |
| Classic robotic process automation scripts | An analyst or developer | Yes | Brittle, costly to maintain |
| Direct app connectors for agents | The user, in plain language | Only where an integration exists | Limited to supported apps |
The comparison is our analysis, not vendor data. It shows why Marble could find a niche. Observation beats connectors in shops full of unsupported software, and it beats hand-written workflows where nobody can write the workflow. It loses where a clean API already exists, since an API call is faster and steadier than driving a screen.
Historical Context: From Macros to Agents
Recording a user and replaying the steps is an old idea. Spreadsheet macros did it for individuals decades ago. Robotic process automation industrialized the same trick for enterprises, with analysts building bots by capturing clicks. Those bots ran fine until a screen changed, then they failed quietly. Process mining tools later tried to reconstruct workflows from system logs instead of screens.
What changed is the reader of the recording. Earlier tools replayed exact coordinates. A language-and-vision model can read the same recording and infer intent, so a moved button no longer ends the run. Marble’s bet is that this inference step finally makes learning from observation sturdy enough to sell. Whether it is sturdy enough for a hospital billing queue or a bank’s reconciliation desk is still open.
The idea also lines up with the wider agent push. Open-source and lab releases keep expanding what agents can do, such as the DeepSeek super-agent prototype we covered with zero published benchmarks. Capability is arriving faster than the knowledge of how a given company actually operates. Marble is selling the second half of that equation.
Market Impact and the Funding Picture
A $3.75 million pre-seed is small next to the headline rounds in AI. It still carries signal. Two respected European funds co-led, and the Munich team plans a US launch quickly. That points to a belief that process capture is a distinct software category, separate from model labs and from general agent platforms.
Adjacent bets back up the theme. Infisical raised $40M at a $600M valuation for AI agent security, which shows investors are paying for the controls around agents as well as the agents themselves. Marble sits upstream of that. It supplies the instructions, and others will supply the guardrails.
For buyers, the near-term effect is a new line item. Teams that wanted agents but stalled on documentation now have a vendor offering to produce the documentation as a by-product of normal work. If that works, it shortens the path from pilot to production. If it does not, it adds a recording tool and a pile of maps nobody trusts.
Privacy, Security and Employee Trust
Any tool that logs screenshots and keystrokes collects sensitive material. Passwords, customer records and private messages can all cross a screen. The reports we reviewed do not describe Marble’s redaction, storage or retention practices, so those remain open questions for any buyer.
Staff reaction is the other hurdle. Employees asked to train a system that may later do their tasks have reason to be wary. In Germany, where Marble is based, works councils generally have a say over technical systems that can monitor workers, so a rollout there involves labor representatives as well as IT. That is general context on German workplace practice, not a statement about any Marble deployment.
Agent permissions are a further risk. An agent that acts inside live business software holds real power. In our earlier coverage, 9 of 10 AI agents raised their own crypto spending caps in a Cornell and Yale study. Teams adopting Coworker-style agents should assume limits need enforcement outside the agent, and our guide to identity providers for AI agents covers six options for that job.
What It Means for Jobs
The workforce angle is the one readers will care about most. A system that learns by watching people also, by design, prepares to take their tasks. Marble frames Coworkers as handling routine execution while humans cover judgment, and the reports describe human intervention as part of the design. They do not give any headcount effect for the named customers.
History in this area urges caution on both sides. Gartner projected that 30% of staff laid off because of AI would be rehired by 2029, partly because the replacement work cost more than expected. If that pattern holds, firms that cut roles around agents trained by observation may find the unrecorded parts of the job were the expensive ones.
Five Predictions
These are our forecasts, based on the reported facts and the pattern of earlier automation waves. None come from Marble or its investors.
- Pilots will outnumber production deployments. Named customers at pre-seed usually mean trials. Expect the first public case studies to describe narrow, repeatable back-office processes.
- Competitors will copy the observation layer. Agent platforms that rely on written prompts will add screen-recording capture, since the idea is easy to imitate and hard to defend alone.
- Accuracy numbers will become the sales battleground. Buyers will demand error rates on their own workflows, and vendors who publish them early will win trust.
- Privacy controls will decide enterprise deals. Redaction and on-premise options will matter more than model quality for regulated buyers in Europe.
- A US launch will bring a seed round. If the American rollout goes as planned, a larger follow-on round is the logical next step, though no timing has been reported.
What Teams Should Do Now
If you run operations and want to test this class of tool, start small. Pick one process with clear inputs, clear outputs and a low cost of error. Record it with consent from the staff involved. Ask the vendor for the map, read it, and check it against how your best people say they work.
Then ask for four things in writing: where recordings are stored, how sensitive fields are masked, what the agent can and cannot touch, and how a human takes over mid-run. Those answers separate a serious product from a demo. For wider context on how fast AI products reach these decisions, browse our AI and machine learning coverage.
Frequently Asked Questions
How much did Marble Cognition raise? Reports from Tech Funding News, Tech.eu and EU-Startups put the pre-seed round at $3.75 million. The valuation was not disclosed.
Who led the round? Heartcore Capital and Seedcamp co-led it. Angels reportedly included a Datadog board member, a Personio founder, and executives from McKinsey, Mistral AI and Lovable.
What does the Marble app record? The desktop application captures screenshots, mouse clicks and keystrokes while employees work in business software, then builds a structured process map for AI to read and execute.
What are Marble Coworkers? They are the AI agents that run the mapped processes end to end inside a company’s existing software. Humans can intervene when a step requires judgment.
Can Marble learn a whole job just by watching? That claim is unconfirmed. The confirmed descriptions cover observing and automating business processes, not replacing an entire role.
Who are Marble’s customers? The reports name EGYM, Mutares and Enam. Contract sizes and deployment scale were not disclosed.
What will Marble do with the money? The company plans to scale inference, grow its Munich engineering team and launch in the United States. No US launch date has been reported.
