On November 2, 2026, the Munich startup Marble Cognition launched with $3.75 million in pre-seed funding led by Heartcore and Seedcamp, and its pitch is that an agent should learn a job by watching people do it. Its desktop application takes screenshots and captures clicks and keystrokes, then turns that record into a structured process map that AI agents can read and execute. Co-founder Felix Rösner described the promise this way: “Our Coworkers watch how the work happens before they take it over. They do it the same way your best people do.” When an agent’s instructions come from observation, the recorded process becomes the rule the agent follows, and someone in the business has to decide whether it is the right rule.
Basil C. Puglisi wrote about that decision in October. His essay Narrative Is Not Custody: Why AI Compliance Needs Traversable Rules places Checkpoint-Based Governance, his framework for putting a named human decision before and after an automated system runs, at two points in that path. The first point, before the system runs, includes approving the encoded rules themselves. In his words, that first checkpoint “governs what the system is allowed to treat as the rule.” A process map derived from screen recordings is an encoded rule. His Factics method holds that every fact must lead to a tactic and every tactic must leave evidence, and a map that nobody has approved hands an agent a set of tactics with no owner attached.
Marble records the work before an agent takes it over
Heartcore partner Max Niederhofer treats the recorded process as the scarce input for automation. In Tech Funding News’s coverage, he said Marble learns a whole company’s processes, anonymises the data, and “derives SOPs [standard operating procedures] and deploys agentic AI against them.” Once the data is gathered, Marble’s Coworkers are designed to run the most repetitive processes end to end in a client’s existing software, with humans in the loop when important decisions need to be made.
Rösner gave SiliconANGLE one early result. At a private equity-backed US financial services firm, the Coworkers were configured to create settlement offers for human review, and the firm cut case-handling times by 90%, which let its staff process around 50% more claims without new hires. Those figures come from the vendor, and neither SiliconANGLE nor Tech Funding News cites an independent measurement, so a buyer should test both in a pilot. The design detail matters more for a buyer than the number does, because the offers went to a person before they went to a customer.
Recording people at work comes with its own obligations. Marble says recorded data is anonymised before anyone can access it, that it records business applications only, and that employees can turn recording off at any time. Tech Funding News also pointed out that German law gives works councils a say in technical systems that monitor employee behaviour or performance, which makes the agreement on what gets recorded a step that comes before any agent is built.
Complexica and Sema4.ai write the approval step into the deployment
Two other launches within days of Marble’s put the approval question in writing. On October 30, Complexica launched Complexica Automations, a business unit that builds agents for multi-step processes inside customer systems, “with their people approving the steps that matter.” Before offering them to customers, the company built ten agents on its own work, with seven in use and three in development. Its customer method runs in five steps. The design step settles which systems the agent reads and acts in and where a person approves, and the final step reviews finished work on a schedule.
Sema4.ai’s 3.0 release on October 29 moves more of each agent’s repetitive work out of the model and into code that customers review and approve. The company says its verified queries can use up to 80% fewer tokens and return results up to four times faster than queries a model writes fresh on each run, a vendor figure a client should test in its own environment. The product documentation states the approval rule directly. A verified query “becomes live because someone with the right to approve it did, not because it looked right once,” and when someone edits a live query, the live version keeps running until the new one is approved.
A recorded process carries the workarounds along with the work
A screen recording captures what people did, including the steps they take because a system is slow or a policy isn’t clear. Some of those steps are good judgment that nobody wrote down, and others are workarounds a manager would want to end. An agent that copies the way the best people work copies both, and it repeats them at a volume no single employee could match. The recording also reflects the weeks in which it was made, so a pricing rule or approval limit that changed last month can still appear as current practice.
Puglisi treats the derived process as a draft rule until a named owner accepts, modifies or rejects it under Checkpoint-Based Governance. The owner is the manager who answers for the outcome of the process. That person reads the map step by step and marks each step as a rule to keep, a workaround to fix or a decision that stays with a person, and only the accepted version goes to the agent.
Setting up a learn-by-watching deployment for a client
The first step comes before any recording. The client agrees with employees, and with a works council where one exists, on which process is in scope, what the software records, how long the data is kept and who can see it. Recording stays inside that scope.
The second step is the owner’s review of the derived map. The owner signs a dated, versioned copy with every step tagged as keep, fix or human decision. Steps that commit money, make promises to customers or create legal obligations, such as the settlement offers in Marble’s example, stay as drafts for human review through the pilot.
The third step controls change. The agent runs only the signed version of the process. When a new recording or the agent itself suggests a better path, the proposal goes back to the owner, and the signed version keeps running until the owner approves the new one. Sema4.ai’s documentation describes the same rule for edits to a live verified query.
A client can track the setup with two numbers. The first is the share of agent-run processes that have a signed, current map, and the target is all of them. The second is the share of recorded steps the owner changed or rejected at the first review. A high figure there tells the client the recording needed heavy correction, so the owner records the reason for each change and confirms the fix is in place before the agent repeats the step.
Observation produces the draft, and the owner sets the version the agent runs
Marble, Complexica and Sema4.ai all describe agents that take over work a company already does, and each puts a person at a defined point in the path. Recording can give a team a useful first draft of a process it never wrote down. The version an agent repeats at volume is a business decision, and a dated signature on that version gives an auditor or a works council the record of who chose it and when.
Sources
- Chesnokova, S. (2026, November 2). Can AI learn your job simply by watching you work? Munich’s Marble has raised $3.75M to find out. Tech Funding News. https://techfundingnews.com/marble-cognition-3-75m-ai-agents-workflow-automation/
- Complexica. (2026, October 30). Complexica launches Complexica Automations, its agentic AI business unit [Press release]. https://www.complexica.com/news/complexica-launches-complexica-automations-agentic-ai-business-unit
- Puglisi, B. C. (2026, October 11). Narrative is not custody: Why AI compliance needs traversable rules. basilpuglisi.com. https://basilpuglisi.com/ai-compliance-traversable-rules/
- Sema4.ai. (2026, October 29). Sema4.ai unveils 3.0 platform enabling team-based work and dramatically reduces token spend [Press release]. https://sema4.ai/newsroom/sema4ai-enterprise-ai-agent-platform-team-based-work-token-reduction/
- Sema4.ai. (n.d.). Verified queries [Product documentation]. https://sema4.ai/docs/v3/build/data/verified-queries
- Wheatley, M. (2026, November 2). Marble Cognition raises $3.75M for its AI agent workers that watch and learn how people work. SiliconANGLE. https://siliconangle.com/2026/11/02/marble-cognition-raises-3-75m-for-its-ai-agent-workers-that-watch-and-learn-how-people-work/
Questions readers ask
What does Marble Cognition do?
Marble Cognition is a Munich startup that launched on November 2, 2026, with $3.75 million in pre-seed funding. Its desktop application records how employees work in business software and turns that record into a process map, and its AI agents, called Coworkers, then run those processes in the company’s existing tools with people involved in important decisions.
Can an AI agent learn a business process by watching employees?
Marble says its software can derive a process from screenshots, clicks and keystrokes, and one lead investor describes the output as standard operating procedures that agents then follow. The recorded process can still include workarounds and outdated steps, so Basil C. Puglisi recommends that a named process owner review and sign the map before an agent runs it.
Who should approve the process an AI agent follows?
The manager who answers for the outcome of the process should approve it. Under Checkpoint-Based Governance, that owner accepts, modifies or rejects each step of the derived map, and only the signed version goes to the agent.
What did Complexica and Sema4.ai announce?
Complexica launched Complexica Automations on October 30, 2026, a unit that builds agents for multi-step processes with people approving the steps that matter. Sema4.ai released version 3.0 of its agent platform on October 29, 2026, moving repeated agent work into verified queries and logic that customers review and approve.
How can a company measure whether a learn-by-watching agent is under control?
Track the share of agent-run processes with a signed, current process map, with a target of all of them. Also track the share of recorded steps the owner changed or rejected at the first review, and record why each step needed a change.
#AIg
Leave a Reply