Healthcare Is Drowning in $1 Trillion of Admin Work. A Computer Use AI Agent Can Fix It.
American hospitals spent over $1 trillion on labor in 2025. A terrifying chunk of that money went to highly trained physicians and nurses who spent their shifts copy-pasting data between software systems, clicking through prior authorization portals, and doing work that a decent computer use AI agent could handle in seconds. Let that sink in. You went to medical school for a decade and your Tuesday afternoon is a copy-paste marathon. This isn't a technology problem. It's a decision problem. The technology to fix it has existed for a while now, and most health systems are still dragging their feet.
The Numbers Are Actually Obscene
A 2025 study published in the ACP Journals found that primary care physicians spend over two hours per half-day clinic session on EHR tasks alone. Not two hours per week. Two hours per half-day. The AMA confirmed the same year that 'pajama time,' meaning physicians doing administrative work after hours at home, hasn't meaningfully budged despite years of burnout initiatives and wellness programs. Meanwhile, Health Affairs reported in January 2026 that health insurers are now running AI systems to accelerate prior authorization reviews. Sounds great, right? Except the AMA also found in early 2025 that 49% of physicians say AI-assisted prior authorization is leading to MORE denials, not fewer. So insurers automated the denial process before providers automated the appeal process. That's the state of healthcare AI right now. One side brought a computer use agent to a gunfight. The other side brought a fax machine.
Why Legacy RPA Is the Villain Nobody's Talking About
- ●Traditional RPA tools like UiPath were built to follow rigid, pre-scripted rules. Change one button in your EHR interface and the bot breaks completely.
- ●Healthcare UI updates constantly. Epic, Cerner, Oracle Health push interface changes regularly. Every update is a potential bot-killing event.
- ●RPA can't read context. It can't look at a screen, understand what's happening, and adapt. It executes scripts. A computer use agent actually sees the screen like a human does.
- ●Implementation costs for enterprise RPA in healthcare routinely run into the hundreds of thousands of dollars before you automate a single workflow.
- ●Administrative costs account for 15 to 30 percent of all US healthcare spending according to Health Affairs. RPA was supposed to dent that. It hasn't.
- ●The US spends more on healthcare administration than any comparable country. Comparable countries aren't using better doctors. They're using less broken processes.
"49% of physicians say AI is leading to MORE prior authorization denials. Insurers automated the rejection process. Providers are still using fax machines to appeal." (AMA, 2025)
OpenAI Operator and Anthropic Computer Use Tried. Here's the Honest Report Card.
To be fair, the big labs saw this problem coming. Anthropic launched computer use capabilities for Claude. OpenAI launched Operator in January 2025. Both are genuine attempts to build computer-using AI that can navigate real interfaces. Both have real limitations that matter a lot in healthcare contexts. An independent review published in mid-2025 found that Operator 'performed poorly' on multi-step real-world tasks and that even the improved ChatGPT Agent released in July 2025 was 'still not very useful' for complex workflows. Anthropic's computer use features are impressive in demos and genuinely useful for simple tasks, but healthcare workflows are not simple. They involve legacy systems, multi-window navigation, conditional logic, and zero tolerance for errors. When a computer use agent fills in the wrong insurance code, someone doesn't get their medication. The bar is higher than 'pretty good in a demo.' This is exactly why OSWorld exists as a benchmark. It tests AI agents on real, open-ended computer tasks in actual desktop environments, not sanitized toy problems. It's the closest thing we have to a real-world stress test for computer use AI.
What Actual AI Computer Use Looks Like in a Health System
Here's what a real computer use agent can do in a clinical operations context right now, today, without a six-month IT implementation project. It can log into your EHR, pull outstanding prior authorization requests, navigate to the payer portal, fill in the required fields, attach supporting documentation, and submit the request. All of it. Without a human touching the keyboard. It can reconcile patient records across systems that don't talk to each other. It can monitor insurance portals for denial notices and flag them for immediate human review. It can handle the mechanical parts of revenue cycle management that currently eat entire departments. The key word is 'mechanical.' A real computer use agent doesn't need an API. It doesn't need the software vendor to build an integration. It uses the screen the same way a human does, which means it works with any system your hospital already has. That's not a minor detail. That's the whole ballgame. Healthcare IT environments are a patchwork of systems from five different decades. An AI agent that requires clean APIs is useless in most real hospitals.
Why Coasty Is the Computer Use Agent Healthcare Actually Needs
I've watched a lot of computer use AI tools get hyped and then quietly underperform when real work shows up. Coasty is different in a way that matters specifically for healthcare. It scores 82% on OSWorld, which is the industry-standard benchmark for computer use agents on real desktop tasks. That's not a marketing number. That's a reproducible, third-party tested score that puts it ahead of every competitor currently on the market, including the offerings from Anthropic and OpenAI. In a domain where a wrong click has real consequences, benchmark performance isn't vanity. It's a proxy for reliability. Coasty controls real desktops, real browsers, and real terminals. It doesn't need API access to your EHR vendor. It doesn't require your IT team to build custom integrations. You point it at a workflow, it watches, it learns, it executes. For health systems dealing with prior auth backlogs, revenue cycle drag, or the soul-crushing EHR documentation burden that's burning out their clinical staff, that's a direct solution. The agent swarm capability means you can run parallel execution across multiple workflows simultaneously, which matters when you're processing hundreds of claims a day. There's a free tier to start, and BYOK support if your compliance team has opinions about where your data goes. They will have opinions. Healthcare compliance teams always have opinions.
Here's my take, and I'll stand behind it: the healthcare system is not suffering from a lack of smart people or a lack of money. It's suffering from a catastrophic misallocation of both. Physicians doing data entry. Nurses chasing prior authorizations. Billing staff manually reconciling claims. Every hour spent on that work is an hour not spent on patients, and it's also a slow-motion financial hemorrhage that no wellness program or EHR upgrade is going to fix. The tools to fix it are real, available, and not that expensive to try. A proper computer use agent, one that actually scores well on real benchmarks and works on real systems, is not a future technology. It's a 2026 technology. The question is whether your organization is going to treat it that way or wait until 2030 and act like you were ahead of the curve all along. Stop waiting. Go look at what Coasty can do at coasty.ai. If your prior auth queue is a nightmare and your clinical staff are doing pajama-time data entry, you don't have a staffing problem. You have an automation problem, and it has a solution.