Cyberdesk Is Building the AI That Finally Automates the Software Nobody Wanted to Touch

The Macro: The Dirty Secret of Enterprise Software Is That Most of It Still Runs on Windows

I want to talk about something that the AI discourse almost completely ignores. The majority of mission-critical enterprise software in the United States runs on Windows desktop applications. Not web apps. Not cloud-native microservices. Desktop applications that were built in the 2000s or earlier, running on Windows machines in hospitals, banks, logistics companies, and government offices.

Epic, the electronic health records system that handles medical data for over 300 million patients in the US, has desktop clients that staff interact with all day. SAP has desktop modules. Legacy insurance platforms, municipal permitting systems, warehouse management tools. These are the applications that actually run the economy, and they are almost universally terrible to integrate with.

The reason is structural. These applications were never designed to be automated. They do not have APIs. They do not have webhooks. They were built for humans to click buttons and fill in forms. When a hospital needs to move patient data from one system to another, someone sits at a desk and types it in manually. When a logistics company needs to update an ERP, someone copies data from an email and pastes it into fields. This is not an edge case. This is the normal state of affairs for millions of workers.

RPA tried to solve this. UiPath, Automation Anywhere, Blue Prism. They built bots that replay recorded mouse clicks and keystrokes. The problem is that RPA bots are brittle. If a dialog box moves, if a popup appears, if the application updates its UI even slightly, the bot breaks. Maintaining RPA workflows is a full-time job, and the failure rates are high enough that many companies have abandoned their RPA investments entirely.

The AI agent layer changes the equation. Instead of replaying recorded actions, an AI agent can look at the screen, understand what it sees, and figure out what to do. It can handle popups. It can adapt when a form layout changes. It can recover from unexpected states. That is a fundamentally different capability than what RPA offered.

The Micro: Two Founders Who Picked the Hardest Possible Problem

Mahmoud Al-Madi (CEO) and Alan Duong (CTO) founded Cyberdesk and went through Y Combinator Summer 2025. The company is based in San Francisco.

The pitch is direct: Cyberdesk is a computer use agent for Windows that developers can program with natural language. You tell it “Open Epic, navigate to the patient record for John Smith, extract the latest lab results, and paste them into this form” and it executes. The key differentiator from RPA is the self-learning aspect. Cyberdesk’s agent learns from executing tasks and memorizes them, so repeated tasks become deterministic rather than probabilistic. The first execution might involve the AI figuring things out. The tenth execution is cached and fast.

The architecture is smart. You install a lightweight driver on the Windows machine. No firewall rules, no VPN configuration, no proxy setup. The agent connects and operates. This is important because the IT departments at hospitals and banks are not going to open network ports or modify security policies for a startup’s product. Minimal deployment friction is not a feature. It is a requirement.

They offer TypeScript and Python SDKs, which means developers can trigger workflows programmatically. There is an observability dashboard for monitoring running workflows, viewing logs, and managing machines. And they are already HIPAA compliant and SOC 2 Type 2 certified, which tells me they are serious about the healthcare vertical. Getting those certifications early is expensive and tedious, and companies that do it are signaling that enterprise customers are not aspirational targets but actual targets.

The use cases they are highlighting are healthcare (EHR automation and patient data entry), finance (account reconciliation and form filing), and logistics (ERP integration). Healthcare is the right beachhead. The pain is acute, the willingness to pay is high, and the compliance requirements create a moat against competitors who have not done the certification work.

The intelligent caching feature deserves attention. The idea that repeated tasks become cheaper and deterministic over time directly addresses the cost concern that killed a lot of RPA deployments. If the first run costs $X in compute and the hundredth run costs a fraction of that, the economics get compelling fast for high-volume workflows.

The Verdict

Cyberdesk is going after a market that is enormous and genuinely underserved. The number of manual data entry hours being burned in healthcare alone is staggering. If the agent is reliable enough to handle the messy reality of legacy Windows applications, this is a company that could build something very durable.

The risk is reliability. Computer use agents in production are still early. The failure modes are unpredictable. A popup the agent has never seen, a workflow that branches in an unexpected way, a form that loads slowly. Each of these is a potential failure point, and in healthcare, failure means a patient record gets corrupted or a medication order gets lost. The stakes are not abstract.

I also want to know about the competitive landscape more specifically. Anthropic has shown computer use capabilities in Claude. OpenAI is working on similar technology. If the foundation model providers ship first-party computer use agents that are good enough, Cyberdesk’s differentiation becomes the enterprise wrapper: the compliance certifications, the caching layer, the deployment model. That is a defensible position, but it is a different company than a pure technology play.

At 30 days, I want to see case studies from the healthcare deployments. How many patient intake workflows per day? What is the error rate? At 60 days, the question is whether the caching mechanism actually delivers on the cost reduction promise. At 90 days, I want to know how many IT departments have approved Cyberdesk for production use, because approval is the real bottleneck in enterprise sales, not interest.

The founders picked possibly the least glamorous problem in AI. Legacy Windows automation is not going to get them on magazine covers. But it might build them a very serious business, because the people who need this are desperate for it.