Healthcare Is Drowning in Paperwork and a Computer Use AI Agent Is the Only Lifeline Nobody's Talking About
A doctor in the United States spends, on average, nearly two hours doing administrative computer work for every single hour of patient care. Read that again. Two hours of clicking, copying, and form-filling for every one hour of actual medicine. Oliver Wyman estimates that healthcare's administrative bloat is costing the system $450 billion over the next decade, and the industry's big response so far has been to throw more clipboard software at the problem. Meanwhile, insurance giants like UnitedHealth got caught using a faulty AI algorithm to auto-deny elderly Medicare patients' claims at a 90% denial rate, and the lawsuits are still grinding through federal court in 2026. So we have one AI being used to create more work for healthcare workers, and a completely different AI that could eliminate that work entirely. The difference between those two outcomes isn't some philosophical debate. It's a question of who controls the computer use.
The Numbers Are Genuinely Obscene
Let's stop being polite about this. The American Medical Association found that 78% of physicians say prior authorization often or sometimes causes patients to abandon recommended treatment entirely. Not delay it. Abandon it. Nurses spend 3 hours per week just on prior auth paperwork. Physicians spend another hour on top of that. Almost 20% of healthcare workers spend more than 20 hours per month correcting billing errors alone, according to Healthcare IT News. The American Hospital Association's 2025 Costs of Caring report puts the median investment required per employed physician at over $317,000 per year, and a huge chunk of that cost is administrative overhead that produces zero clinical value. This isn't a staffing problem. It's a computer problem. Specifically, it's a problem of humans being forced to do things computers should be doing, because the computers in healthcare are still mostly dumb portals that require human fingers to operate them.
The UnitedHealth Scandal Should Have Woken Everyone Up
- ●UnitedHealth's nH Predict AI algorithm was denying Medicare Advantage claims at a 90% rate, according to the class-action lawsuit that's now moving forward in federal court as of early 2026
- ●The Guardian reported in January 2025 that insurers' algorithms turn down claims in seconds, while appeals by human patients take weeks or months
- ●The AMA released a statement in February 2025 warning that AI is being used to automatically deny more needed care, with physicians saying it's getting worse, not better
- ●CMS launched its WISeR model in January 2026 specifically because prior authorization abuse had become a systemic crisis
- ●The cruel irony: insurers have fully embraced computer use AI to process and deny claims faster, while providers are still faxing documents and manually re-entering the same data into three different portals
- ●A Georgetown litigation tracker shows the UnitedHealth AI denial case is still active in March 2026, meaning the legal fallout from weaponized automation is still unfolding
"78% of physicians say prior authorization often or sometimes causes patients to abandon recommended treatment." The insurance company's AI processed that denial in seconds. Your staff will spend 3 hours trying to appeal it. This is the computer use gap in healthcare, and it's killing people.
Why Every Existing 'Solution' Is a Joke
The healthcare IT industry has been selling the same answer for 15 years: integrate your systems. Build APIs. Buy our middleware. How's that working out? EHR usability is still rated as one of the top drivers of physician burnout, per a June 2025 PMC study. Doctors are doing EHR work at home after hours, as the AMA confirmed in August 2025. The problem isn't that systems aren't connected enough. The problem is that even when they are connected, someone still has to navigate five different browser tabs, log into three separate portals with different credentials, copy data from one screen, paste it into another, attach the right PDF, click submit, and then check back tomorrow to see if it went through. That's not an integration problem. That's a computer use problem. And the answer isn't another API. It's an AI agent that can actually sit at a computer, see what a human sees, and do what a human does, but in seconds instead of hours, and without burning out.
What Real Computer Use AI Actually Looks Like in a Healthcare Workflow
The New England Journal of Medicine published research in August 2025 benchmarking medical AI agents against a virtual EHR environment, submitting 300 clinically derived tasks that could benefit from AI agent automation. The results were promising but pointed to a clear bottleneck: most AI models can reason about healthcare tasks, but they can't actually execute them across real desktop software and web portals. That's the gap a genuine computer use agent fills. We're talking about an AI that can open your EHR, pull a patient record, cross-reference it with the insurer's prior auth portal, fill out the required forms using the actual data on screen, submit the request, and log the outcome, all without a human touching the keyboard. Not a chatbot. Not an RPA script that breaks every time the portal updates its UI. A computer-using AI that sees the screen, reasons about what it's looking at, and takes action. PwC's 2026 AI predictions specifically called out healthcare revenue cycle management as one of the ripest areas for AI agents to automate complex, high-value workflows. They're right. They're just late to naming what the tool actually needs to be.
Why Coasty Is the Tool Healthcare Operations Teams Should Be Testing Right Now
I'm not going to pretend there aren't a dozen AI tools claiming to automate healthcare workflows. But most of them are either narrow RPA bots dressed up with a ChatGPT wrapper, or they're cloud-only API integrations that still require your IT team to spend six months on implementation. Coasty is a computer use agent, meaning it controls a real desktop, a real browser, and real terminals the same way a human operator would. It scored 82% on OSWorld, the gold-standard benchmark for AI computer use, which is higher than every competitor including Anthropic's Computer Use and OpenAI's Operator. That gap matters in healthcare because your workflows aren't clean. Your prior auth portal looks different from your EHR. Your billing software is from 2014. Your insurer's submission form changes every quarter. A brittle RPA script fails every time something shifts. A true computer use agent adapts because it's reading the screen in real time, not following a hardcoded click path. Coasty runs on a desktop app, spins up cloud VMs, and supports agent swarms for parallel execution, meaning you can run 20 prior auth submissions simultaneously instead of one at a time. There's a free tier to start, and BYOK support if your compliance team needs it. The question isn't whether you can afford to try it. The question is how many more hours your staff is going to spend on things a computer use agent could handle by lunch.
Here's my honest take on healthcare automation in 2026: the industry is at a genuinely absurd inflection point. The same technology being used to deny patient care in milliseconds could be used to fight those denials, file appeals, manage prior authorizations, reconcile billing records, and give clinicians their time back. The $450 billion in administrative waste isn't going to fix itself with better middleware or another EHR training session. It's going to get fixed when healthcare operations teams stop waiting for their software vendors to build integrations and start deploying AI agents that can use any software that already exists. The physicians are burned out. The nurses are drowning in paperwork. The patients are abandoning care because the system is too slow. None of that is inevitable. It's a computer use problem with a computer use solution. If you want to see what that actually looks like in practice, go to coasty.ai and run it yourself. The demo is free. The time you're wasting right now is not.