loading...
Skan AI watches employee desktop sessions to build the process maps that AI agents need — capturing the informal work that event logs have always missed.
ENTRY ANGLES
Agent behavior monitoring and evaluation against observed human baselines · Lightweight browser-extension process discovery as a Skan entry point for mid-market
VERTICALS
CAPABILITIES
Computer vision, On-device ML, Process mining analytics, Enterprise endpoint deployment
The challenge with deploying AI agents in enterprise is not finding agents capable enough to attempt the work. It is that the work itself has never been formally described. Invoice approval at a large company involves an employee opening a procurement system, checking a vendor record in a different system, checking a budget allocation in a third system, applying a judgment about payment urgency that exists nowhere in writing, and routing to a specific approver based on context that lives in that approver's email. Give an agent instructions to 'approve invoices' without that process map and the agent fails at the first undocumented decision point — or makes the decision incorrectly because it doesn't know what the decision actually depends on.
Skan AI, a Menlo Park company founded roughly seven years ago with approximately $120 million in total capital, has built the observation infrastructure to produce those maps. Its software runs on employee desktops, captures screenshots, processes them locally on the machine — so no content leaves the endpoint — and sends anonymized metadata back to a central analytics platform: which applications were used, in what sequence, how long was spent at each step, and where the human paused or diverted from the expected path. The output is a process record built from how work actually gets done, not how it was documented when someone last updated the process flowchart.
The $63 million Series C, co-led by Cathay Innovation and Dell Technologies Capital, closed August 12, 2026. Citi Ventures, Bloomberg Beta, State Farm Ventures, and Wipro Ventures participated — a cross-industry investor composition that reflects how broadly the problem distributes across enterprise verticals.
Process intelligence is not a new category. Celonis, valued at $11 billion in 2021 on over $1 billion in capital, pioneered log-based process mining — analyzing event logs from ERP and CRM systems to construct process models that reveal inefficiencies. SAP Signavio and IBM's process mining tools occupy the same approach. What log-based mining can see is the happy path: the sequence of events the system records when work proceeds as designed. What it cannot see is exception handling — the analyst who opens five applications before categorizing an unusual transaction, the claims adjuster who calls a colleague before approving an edge case, the procurement manager who routes through an informal channel before the formal ticket updates. That informal layer accounts for a disproportionate share of enterprise labor cost and is precisely what breaks AI agent deployments, because agents built on log-based models don't know those paths exist.
Skan's observation at the UI layer rather than the event log layer captures what actually happens in the application session. The on-device processing architecture — screenshots analyzed locally, behavioral metadata sent rather than content — is not incidental. It is what makes deployment possible in regulated industries where desktop recording creates compliance and data sovereignty exposure. The metadata Skan collects is structural: application, sequence, duration, decision point location. Not content: not what the email said or what the document contained.
The timing of the Series C is precise. Enterprise AI agent investment has accelerated sharply through 2026, and with it the discovery that agents fail reliably outside well-structured tasks. The capital flowing into agentic AI creates immediate demand for the context layer Skan provides: not a better agent, but the map that tells an agent what the actual steps are, where the decision points occur, and what the exception paths look like. Seven years of accumulated process intelligence becomes retroactively valuable at the moment enterprises most need what it enables.
The observation infrastructure Skan has built produces the input to agent deployment. It does not produce the feedback loop that tells you whether a deployed agent is using that context correctly. An enterprise that maps its invoice approval workflow and deploys an agent to automate it needs a monitoring system to determine whether the agent is following the process map or deviating in ways the map didn't anticipate — handling edge cases incorrectly, skipping validation steps, taking paths that look like the happy path but produce wrong outcomes. That evaluation layer — comparing agent session behavior to the observed human baseline for the same process — doesn't exist as a standalone product category. It sits directly above what Skan has built: Skan produces the ground truth; the evaluation layer measures agent adherence to it. The company that builds the agent evaluation product sells into the same operations and engineering budget that bought the process map, and creates the instrumentation loop that enterprise AI governance teams currently have to build manually.
The second opening is distribution leverage. Skan's enterprise sales motion requires deploying native software on employee endpoints, which involves IT security review and often a compliance process. A lighter-weight alternative — a browser extension that captures application-switching signals rather than screenshots, producing lower-fidelity process maps at the cross-application workflow level — has much lower deployment friction and reaches the mid-market segment that can't commit to a full endpoint agent rollout. That lighter product is a funnel into the fuller platform once a team has seen what process discovery can surface; it's also a standalone viable product for the many organizations that need 'good enough' process documentation for agent scoping without the full data richness Skan provides.