Somewhere in America right now, a healthcare administrator is clicking through a patient record system whose interface hasn't meaningfully changed since the Bush administration. Not the recent one—the first one. There's no API to speak of. No webhooks. Just point-and-click workflows processing millions in daily transactions, oblivious to the existence of cloud computing or, for that matter, the entire modern internet.
For artificial intelligence companies trying to build enterprise integrations, this presents what you might call an inconvenient reality. Their customers' most critical systems—the ones that actually move money and save lives—often exist in a kind of digital amber, frozen in an era when "automation" meant macros and VBA scripts.
Minicor, a five-person outfit from Y Combinator's Spring 2026 cohort, thinks it's found a way around this. Or perhaps more accurately, a way directly through it. Rather than waiting for legacy software to modernize (a wait that could, charitably, take decades), the San Francisco startup has built what it describes as self-healing desktop automations—software capable of reading and writing into Windows applications at scale, adapting on the fly when interfaces change or unexpected errors crop up.
According to Minicor, AI companies building agents for healthcare, logistics, or financial services shouldn't have to become desktop automation experts. They should be able to make an API call and let someone else worry about the arcane details of scraping data from a Citrix environment or pushing insurance claims into a processing system that predates the iPhone.
How the machinery works
At the technical core, Minicor operates as a managed service. An AI company needs to extract patient data from an electronic health record, say, or inject information into some ancient dealer management system. They fire off a POST request. Minicor's platform spins up a Windows VM—or runs the process in a browser, depending on where the legacy application lives—executes the workflow, and returns structured JSON complete with verification flags indicating what worked and what didn't.
The company reports 93 to 96 percent click accuracy, a meaningful jump from the 80 to 85 percent range typical of competing tools. That improvement, according to Minicor, stems from what it calls a "reflection agent"—essentially a verification layer that checks every action against what's actually rendered on screen and course-corrects before the entire workflow derails. Rather than relying entirely on computer-use models to navigate interfaces (a recipe for brittle automation), Minicor stores workflows as deterministic code and deploys the agent only for recovery and edge cases.
The company says the architecture has processed substantial volume—25,000 patients daily at one point, though it hasn't disclosed which customers or when that figure was current, and the claim hasn't been independently verified. Its use cases span the usual suspects in legacy software hell: EHR platforms like Athena, Epic, and Cerner; automotive dealer management systems including CDK Global; supply chain stalwarts like SAP; dental practice management tools. Minicor says it maintains SOC 2 Type II and HIPAA compliance credentials, table stakes for anyone touching healthcare data.
Every workflow run generates a video replay, stored for debugging purposes. When something breaks—and things do break, even in self-healing systems—Minicor fires off Slack notifications with screenshots and full execution context. A nod to reality: even the most sophisticated automation eventually encounters problems that require human judgment.
The deployment model is flexible because it has to be. Some customers run Minicor on their own infrastructure. Others use cloud VMs. Some operate within Citrix environments, that particularly thorny subspecies of legacy deployment. The platform adapts to Windows desktops or browser-based applications, depending on where the target software actually lives.
Crowded territory, different angle
Minicor is hardly pioneering uncharted waters here. UiPath, the robotic process automation giant, updated its documentation this spring to describe a "Healing Agent" for unattended UI automation. Automation Anywhere has spent the better part of two years emphasizing what it calls "Agentic Process Automation." The Spring 2026 YC batch alone includes several companies tackling variations of the same fundamental problem: software that refuses to talk to other software.
What distinguishes Minicor, at least in how it frames the opportunity, is the customer. Rather than selling to enterprises directly—the traditional RPA playbook—the company is targeting AI startups that need desktop automation as a means to an end. The subtext: you're building healthcare agents or logistics intelligence, not wrestling with the peculiarities of how Epic renders patient data. Let someone else handle that headache.
Whether that positioning proves defensible remains an open question. The RPA incumbents aren't exactly standing still, and the technical moat in desktop automation has always been less about the core technology and more about the accumulated domain knowledge of handling thousands of edge cases across different systems.
Who's building it
Co-founders Faizaan Chishtie (CEO) and Saheed Akinbile (CTO) started the company in 2024 under the name Laminar, rebranding to Minicor earlier this year. Akinbile's background includes infrastructure and RPA work with QuickBooks, Sage, and various Citrix environments—precisely the kind of unglamorous experience that proves invaluable when you're trying to automate software designed in an era when "mobile" still meant laptops.
The company has been actively hiring. Recent job listings for Forward Deployed Engineer and Product Engineer roles offer equity grants between 0.15 and 0.45 percent. Chishtie posted on LinkedIn recently with a $5,000 referral bonus, describing Minicor as "the bridge between AI and enterprise legacy systems"—a phrase that's either ambitious or a statement of the obvious, depending on your perspective.
Minicor launched on Y Combinator's platform a few months back, with Tom Blomfield serving as its primary partner. As of the latest update, the company hasn't announced any funding beyond YC's standard batch participation. Its pricing model appears to be usage-based—"only pay when the data gets in," according to the launch post—though there's no public pricing page as of yet.
The underlying bet is simple, even if the execution isn't: as AI agents graduate from demos to production deployments, someone needs to solve the unglamorous last-mile problem of actually moving data into and out of the creaky systems that still run most critical workflows. Minicor thinks that's an API problem, not a multi-month consulting engagement.
Whether enterprises and their AI vendors agree remains to be seen. But for now at least, Windows XP's children keep running, processing claims and managing inventory and scheduling appointments, blissfully unaware that in Silicon Valley, startups are building increasingly elaborate bridges to reach them.
