How Startups Are Capturing Human Work for AI Agents

Companies know that their most valuable data is not always stored inside databases or documents. Much of it exists in the way employees complete tasks, handle exceptions and apply judgment in situations that were never formally documented. A whole ecosystem of startups is trying to capture this hidden layer of work so AI agents can learn how organizations actually operate. This landscape looks at the companies building the workflow and knowledge data infrastructure needed to train more capable enterprise agents.

Category #1: Workflow Observation & Demonstration Capture

What is this category about?

• The data layer for understanding how digital work is actually performed. These companies record the sequence of actions people take across software so an organization can understand a workflow from direct evidence rather than relying only on documentation or employee interviews.

• Capturing the difference between the official process and the real one. An SOP describes how a task is supposed to be completed. Actual work usually includes exceptions, workarounds, application switching and small judgment calls that were never written down. Workflow observation products make these variations visible.

• Turning human activity into machine readable context. A recording can become a step-by-step guide, process map, automation specification or training trace. The strategic value therefore comes less from the recording itself than from the structured representation of the workflow that the product creates.

• Moving from documenting people to teaching agents. Task mining was originally used to find opportunities for robotic process automation and process improvement. The same data can now help AI agents understand which steps to follow, which exceptions they may encounter and when a decision should be escalated to a human.

• This shift is expanding the role of the category. Gartner already defines task mining as the recording and analysis of screen interactions, clicks, keystrokes and data entries. Its market guidance increasingly connects the technology with the discovery and design of AI agent use cases.

What do products in this category do?

The products follow several distinct approaches depending on how the workflow is captured and what the customer wants to do with the resulting data.

• The first approach is intentional demonstration. A user performs a task once while the product records the steps and turns them into a reusable representation of the workflow. Scribe converts the demonstration into visual documentation, while Trainer goes further by using it to create an agent that can attempt the same task.

• The second approach is continuous personal capture. Instead of asking the user to record one specific process, the product observes activity throughout the working day. Screenpipe creates a private memory of what the user saw and discussed, which can later be searched or accessed by an AI agent when it needs context.

• The third approach is enterprise task observation. These platforms collect activity across groups of employees so the company can reconstruct how a complete process operates across several applications. Mimica uses this data to generate process maps and identify automation opportunities. Skan AI and Soroco create a broader model of how work moves through the organization.

• The fourth approach connects observation directly with agent deployment. The captured workflow becomes the basis for deciding what an agent should do and how its behavior should be evaluated. Mimica can use process maps to accelerate automation development. Skan AI turns observed work into agent opportunities and governance rules. Soroco uses its work graph to help design workflows that combine employees and agents.

• The products are therefore converging from two directions. Documentation tools are becoming workflow intelligence platforms because their recordings contain useful operational data. Task mining platforms are becoming agent infrastructure because their process models can define how an agent should behave.

• The observation layer generally sits across the software the customer already uses. Employees continue working in their existing applications, while the capture product builds a separate model of how the complete workflow operates.

• Deployment remains an important part of the product. Recording screens can expose confidential information and personal employee data. Screenpipe emphasizes local storage, while enterprise platforms rely on masking and administrative controls to limit what is collected and who can access it.

What types of customers are using these products?

• Large shared services and enterprise transformation teams are the most established buyers. These organizations manage processes that cross several departments and software systems, making it difficult to understand where time is being lost. Merck uses Mimica to analyze work inside its Global Business Services organization, while McKesson applies the platform to accounts payable and other shared services workflows.

• Financial institutions use workflow observation to understand operations that contain large amounts of manual work. ClearBank adopted Mimica while transaction volumes were increasing because it wanted to scale without expanding headcount at the same pace. Northern Trust uses Scribe to standardize internal processes and reduce the time employees spend on non client work. Skan AI also works with banks and insurance providers, although many of its largest customer stories remain anonymous because of the sensitivity of the workflows involved.

• Customer support and training teams use these products to capture knowledge that would otherwise remain with individual employees. Diesel Laptops uses Scribe to document the expertise of technical support employees and make it available to less experienced team members. Crexi uses the product to answer customer questions and maintain internal training material as its software changes. New York Life has expanded the same approach across more than 80 teams.

• Companies going through major software changes are another important customer group. Bayer has used Soroco during an SAP S/4HANA migration, while Worldnet International used Scout to understand application usage and improve technology adoption. Soroco has also worked with BT Group and LNER on broader operational improvement programs where work is distributed across many teams and systems.

• Industrial and operational businesses use workflow capture when important processes still depend on undocumented human activity. Kubota uses Scribe to create more accurate procedures, while Simplot uses it to reduce knowledge silos across its global organization. Soroco has worked with Morgan Sindall and The Wonderful Company to identify bottlenecks and reveal process improvement opportunities inside complex operations.

• Some customers are now using the observation data specifically to prepare their AI strategy. TXNM Energy used Scribe Optimize to map workflows and identify where AI could create value. Mimica’s work with Merck connects task mining with automation and generative AI opportunities, while Soroco increasingly shows customers using observed human workflows to improve agents after deployment.

• The public customer evidence remains concentrated among the more established enterprise vendors. Screenpipe and Trainer are still earlier in their commercial development and have not yet published comparable named customer deployments. Their current products are more accessible to individual users, developers and small teams, which may allow them to enter companies through bottom-up adoption before building a larger enterprise sales motion.

What are the major trends shaping this category?

• Workflow observation is becoming part of the agent development stack. Task mining was originally used to improve processes or identify opportunities for robotic automation. The same data is now being used to teach AI agents how work is actually performed.

• Workflow data is also becoming executable. A process map no longer serves only as an analytical document. Vendors increasingly want to turn it into instructions, evaluation criteria and guardrails that an agent can follow.

• Continuous observation is replacing one-time process mapping. Workflows change as software and company policies evolve. The platform therefore needs to keep observing the process and update its model after agents enter production.

• Exceptions are becoming more valuable than the standard workflow. The normal path is usually easy to document. The real advantage comes from capturing how experienced employees respond when information is missing or a case does not follow the expected pattern.

• The market is moving from employee observation toward human-and-agent observability. Companies will need to understand how work moves between people and AI systems, especially when responsibility changes during the workflow.

• Privacy will remain the main limit on adoption. Screen recording can easily be perceived as employee surveillance. Vendors that keep data inside the customer’s environment and clearly separate process analysis from individual performance monitoring will have an advantage.

• The proprietary workflow dataset may become the most important competitive asset. Models and agent frameworks are widely available, but a detailed history of how one company performs its work is much harder to reproduce.

How does the funding environment look?

• The category has attracted approximately $219 million in disclosed funding across Scribe, Mimica, Skan AI and screenpipe.

• Scribe represents most of the recorded capital. The company has raised $130 million and announced a $75 million Series C in November 2025. The round valued the business at $1.3 billion.

• The company has also demonstrated that a workflow capture product can grow beyond a documentation tool. Scribe announced in May 2026 that it had passed $100 million in annual recurring revenue. This makes it the clearest commercial proof point in the category.

• Mimica has raised $32.2 million across its Series A and Series B. Its $26.2 million Series B was announced in September 2025. The financing supports a strategy that is moving the company from task mining toward agent training and automation development.

• The funding market is divided between two models. Established enterprise platforms require capital to support long sales cycles and demanding deployments. Newer products can begin with one user and validate the workflow capture experience before building an enterprise sales organization.

• Consolidation has only started to appear. Skan’s acquisition of Metaculars is the clearest example inside the landscape. Most companies are still expanding internally from observation toward agent design rather than acquiring a separate agent platform.

Category #2: Tacit Knowledge & Decision Capture

What is this category about?

• The intelligence layer for knowledge that has never been formally documented. These companies help organizations capture the judgment and practical experience that employees use to make decisions every day.

• Understanding why people act rather than only observing what they do. Workflow observation products record the steps taken inside software. Tacit knowledge products try to uncover the reasoning behind those steps through conversations or by studying how previous cases were resolved.

• Making professional judgment accessible to the wider organization. An experienced employee may recognize which customer request requires escalation or which exception can safely be approved. This knowledge is often difficult to explain and rarely appears in a formal procedure.

• Turning employee experience into context for AI agents. A model can read the company’s documentation, but documents rarely explain how rules are interpreted in practice. These products aim to give agents access to the patterns and decision logic that experienced employees have accumulated over time.

• The category is therefore closely connected to organizational memory. The goal is not simply to create more documentation. It is to build a system that continues learning from how the organization solves problems.

What do products in this category do?

• The first product approach uses AI interviews to conduct organizational research at scale. Ontora sends an AI interviewer to employees and synthesizes their answers into a model of how the company operates. The platform can use the resulting information to identify bottlenecks and prepare an automation roadmap.

• The second approach learns from the organization’s history of resolved work. Interloom analyzes previous cases and extracts the patterns that explain how employees reached a successful outcome. It combines these examples into a context graph that can guide another employee or an AI agent when a similar case appears.

• This approach is particularly relevant when the employee cannot fully explain their own expertise. The knowledge may only become visible after the system compares many previous decisions and identifies which context led to a particular response. Interloom therefore relies less on asking an expert to describe the process and more on learning from the evidence left behind by the work itself.

• The third approach turns expert conversations into a structured knowledge base. KNOA conducts guided interviews and combines the answers into evolving documents. It can compare responses from several people and surface areas where their understanding does not align.

• The products are beginning to converge after the knowledge has been captured. An interview may initially produce a report, while a context graph may initially guide a human employee. Over time, both become a source of operational context that an AI agent can use.

• Traceability is an important part of the product. The user needs to understand whether an answer came from an employee interview or an inferred pattern across previous cases. Inplicit links its findings back to anonymous quotes, while Ontora exposes grounded querying across interview transcripts through its GraphRAG layer.

• The platform normally complements the company’s existing knowledge systems. Formal documents can remain inside tools such as Confluence, while previous cases may stay inside the customer’s operational software. The tacit knowledge product adds the missing layer that explains how employees interpret this information in practice.

What types of customers are using these products?

• Large companies preparing an AI transformation are an important target market. Vertiv uses Ontora to accelerate operational research that previously required months of interviews and analysis. The company’s strategy team reported that Ontora delivered the result within a single afternoon.

• Private equity firms represent another customer type. One of Ontora’s early design partners is a San Francisco investment firm deploying the product across portfolio companies. In this use case, the platform provides a faster way to understand how each business operates and where efficiency improvements may be possible.

• Industrial companies use these products to make operational experience available across the organization. Lenze uses Interloom to capture knowledge that were previously scattered across employee conversations and historical cases. The company can then reuse this knowledge when similar problems need to be resolved elsewhere.

• Banks are using the technology to understand workflows where the written procedure does not match the real operation. Interloom analyzed millions of support emails at Commerzbank and compared them with the bank’s existing documentation. The deployment found a substantial difference between the formal knowledge base and the way employees actually handled customer cases.

• Customer service teams form another important market. Volkswagen uses Interloom to process support tickets and learn from how earlier requests were resolved. This allows the platform to build operational memory from the company’s existing work rather than asking employees to document each decision manually.

• Insurance provides a strong example of why tacit knowledge matters. Interloom has worked with Zurich Insurance on an underwriting use case. The decision cannot be reduced to a generic rule because it depends on how the insurer evaluates a particular broker and interprets its own risk appetite.

• The public customer evidence remains limited outside Ontora and Interloom. Inplicit, KNOA and Tacivo are still establishing their commercial motion and have not published comparable named deployments. Their current websites are oriented toward pilots and direct founder led sales rather than a mature library of customer stories.

What are the major trends shaping this category?

• Tacit knowledge is becoming a major bottleneck for enterprise AI agents. Models can access company documents, but the most important operational decisions often depend on experience that was never written down. Agents will remain unreliable when they cannot access this context.

• Organizational memory is becoming continuous. Traditional knowledge transfer projects happen when an employee leaves or when a company prepares a transformation. New products aim to capture knowledge throughout normal operations so the context remains current.

• The market is developing two complementary methods. Interview platforms ask employees to explain what they know, while context-graph products infer knowledge from previous work. The strongest platforms may eventually combine both sources so they can compare what employees say with what happens in practice.

• Knowledge capture is moving closer to execution. The first output may be a report or internal wiki, but vendors increasingly want the same knowledge to guide an AI agent. This makes the category part of the agent infrastructure stack rather than only a knowledge management market.

• Traceability will remain critical. A company must be able to understand where an operational recommendation came from before it allows an agent to act on it. Products that connect conclusions to interviews or historical cases will be easier to trust.

• Privacy may determine which products can be deployed at scale. Employee interviews can reveal sensitive concerns or internal criticism. Inplicit makes anonymity and European data processing central to the product, while other vendors emphasize controlled enterprise deployment.

• The proprietary context layer may become a real competitive advantage. General-purpose models are available to every vendor. A continuously updated record of how one company applies judgment is much harder for a competitor to reproduce.

How does the funding environment look?

• The category has attracted at least $20 million in disclosed funding, much less than the other category on the landscape. Almost all of this capital has gone to Interloom, while the remaining companies are still operating with small teams or have not disclosed external financing.

• Interloom raised an initial $3 million seed round in February 2024. Air Street Capital led the financing when the company was still positioned around adaptive enterprise automation.

• The company raised another $16.5 million in March 2026. DN Capital led the round with participation from Bek Ventures and existing investor Air Street Capital. Interloom and its investors describe this financing as a seed round, despite its relatively large size and the previous seed announcement.

• Ontora has raised at least the standard $500,000 investment associated with its participation in Y Combinator. Its YC launch update says that the company had raised $700,000 before formally opening its round, which suggests that a small additional investment may not yet be reflected in the Axomap timeline.

• The category has attracted much less funding than workflow observation because it is considerably younger. Task mining and process intelligence have existed as established enterprise software markets for several years. Investors can therefore evaluate these companies against known budgets, existing competitors and measurable automation outcomes.

• The initial deployment is also harder to standardize. Capturing knowledge from one organization may require adapting the interviews and outputs to its internal language. A workflow observation product can begin by recording the same types of screen interactions across many customers, while tacit knowledge is more dependent on the context of each company.

• The return on investment is less immediate to measure. A workflow observation platform can point to reduced process time or a clear automation opportunity. The value of preserving judgment may only become visible later when an employee leaves or when an agent handles a complex exception more accurately.