Healthcare Is Drowning in Paperwork and a Computer Use AI Agent Is the Only Lifeline That Actually Works
A physician in the United States now spends roughly 28 hours every single month doing clinical documentation on nights and weekends. Not treating patients. Not doing research. Typing into a computer at midnight in their kitchen. The American Medical Association calls this 'pajama time.' I call it a catastrophic failure of every automation tool the healthcare industry has thrown money at for the last decade. We are in 2026. Hospitals spent $30 billion in 2025 alone just on the technology and administrative overhead required to deal with payer policies. Twenty-five percent of all healthcare spending in the US is considered wasteful, and the single biggest bucket of that waste is administrative. We have AI that can write poetry, generate code, and beat world champions at chess. And yet a staff member at your doctor's office is still manually filling out 39 prior authorization requests per physician, per week, spending 13 hours on them. Thirteen hours. Per week. Per doctor. This is not a resource problem. This is an imagination problem, and a tool problem.
The Numbers Are So Bad They Sound Made Up (They're Not)
Let's actually sit with these statistics for a second because they deserve more than a passing glance. The AMA's 2024 prior authorization survey, released in early 2025, found that physicians and their staff complete an average of 39 prior authorization requests per physician per week. Forty percent of physicians have staff members who work exclusively on prior authorizations. That's a full-time employee whose entire job is asking insurance companies for permission to help sick people. And that staff member is doing it manually, logging into payer portals, copying data from EHRs, filling out web forms, downloading PDFs, uploading them somewhere else, and then following up by phone when the portal crashes. Meanwhile, California's healthcare system alone wastes an estimated $21 billion per year on administrative overhead. Nationally, researchers peg wasteful administrative spending at hundreds of billions annually. The system isn't broken. It's working exactly as designed, which is the most depressing possible explanation. The people getting crushed by it are the clinicians, the billing staff, and ultimately the patients waiting longer for care because their doctor's practice is buried in forms.
Why the 'Automation' You Already Bought Is Failing You
- ●Traditional RPA tools like UiPath build bots that follow rigid, scripted paths. The moment a payer portal updates its UI or adds a new field, the bot breaks. Healthcare IT teams then spend weeks rebuilding it.
- ●Over 80% of organizations planned to hire more automation professionals in 2025 just to maintain their existing RPA deployments. You bought automation and ended up hiring more people to babysit it.
- ●API-based AI tools only work where APIs exist. Epic, Cerner, and most payer portals don't offer clean APIs for everything. Most real healthcare workflows live in legacy UIs that no API can touch.
- ●AI scribes help with documentation, but they're one narrow slice. They don't touch prior auths, billing workflows, referral coordination, or the 50 other manual tasks that eat staff time every day.
- ●OpenAI's Computer-Using Agent launched to massive fanfare in early 2025 and scored 38.1% on OSWorld, the industry standard benchmark for computer use agents. That means it failed on roughly 62% of real-world computer tasks. That's not a tool you deploy in a clinical setting.
- ●Anthropic's computer use offering, while more capable, still trails significantly behind the current state of the art. Healthcare can't afford 'pretty good.' A failed prior auth means a delayed surgery.
- ●The 1.2 FTE problem is real: the National Academy of Medicine documented in early 2026 that clinician burnout is driven by administrative load equivalent to working 1.2 full-time jobs. No scribe fixes that.
Physicians and their staff spend 13 hours per week, per doctor, just on prior authorizations. That's 676 hours a year. At a loaded staff cost of $35/hour, that's $23,660 per physician, per year, spent asking insurance companies for permission. Multiply that across a 10-doctor practice and you're looking at $236,000 annually. Gone. On forms.
What 'Computer Use' Actually Means and Why Healthcare Needs It Specifically
Here's the thing that most healthcare administrators don't fully understand yet. There's a difference between an AI that generates text and an AI that actually operates a computer. A computer use agent doesn't call an API. It sees your screen, moves a mouse, clicks buttons, fills out forms, switches between applications, reads what comes back, and adapts when something unexpected happens. It works on any software, any portal, any legacy system, because it interacts with the UI exactly the way a human would. That distinction matters enormously in healthcare. Your payer portal doesn't have an API. Your 15-year-old practice management system doesn't have an API. Your state Medicaid portal definitely doesn't have an API. But a real computer use AI agent doesn't care. It can log into UnitedHealthcare's prior auth portal, read the patient record in your EHR, cross-reference the clinical criteria, fill out the form, attach the supporting documentation, submit it, and log the confirmation number back into your system. Without a human touching it. That's not a pitch. That's just what the technology can do in 2026 when it's built correctly.
The Benchmarks Don't Lie: Most AI Agents Aren't Ready for This
The OSWorld benchmark is the industry's best stress test for computer use agents. It throws real-world desktop tasks at AI systems and measures how often they actually complete them correctly. The gap between the leaders and the pack is shocking. OpenAI's operator scored 38.1%. Anthropic's computer use offering has improved but still sits well below the threshold you'd want before automating anything patient-adjacent. Most of the tools being marketed to healthcare IT departments right now would fail the majority of the tasks you'd actually need them to do. This is why the 'we tried AI automation and it didn't work' story is so common in healthcare. It's not that AI automation doesn't work. It's that the specific tools deployed were not good enough for the complexity of real healthcare workflows. A prior authorization workflow involves conditional logic, document parsing, multi-system navigation, and error handling. A 38% success rate agent doesn't cut it. You need something that actually finishes the job.
Why Coasty Is the Computer Use Agent Healthcare Has Been Waiting For
I've looked at the benchmark numbers and the competitive field pretty closely. Coasty sits at 82% on OSWorld. That's not a marginal improvement over the competition. That's a different category of capability. When you're automating prior authorizations or insurance eligibility checks or referral workflows, the difference between 38% and 82% is the difference between a tool that creates more work and one that actually eliminates it. Coasty controls real desktops, real browsers, and real terminals. It doesn't need an API. It doesn't need your EHR vendor to build an integration. It works with what you already have. The agent swarm capability means you can run parallel workflows simultaneously, so a practice handling 39 prior auths per physician per week isn't doing them one at a time. You can also bring your own API key if your organization has compliance requirements around that, and there's a free tier to start without a procurement nightmare. For healthcare organizations that have been burned by RPA deployments that required a full-time engineer to maintain, the architecture here is meaningfully different. The agent adapts to UI changes instead of breaking on them. That alone is worth the conversation.
Here's my honest take. Healthcare administration is not going to fix itself through policy. The prior authorization reform bills move slowly, payer portals stay terrible, and EHR vendors have zero incentive to make interoperability easy. The only lever that actually moves fast is deploying AI that can work within the broken system as it exists right now, not as we wish it were. The doctors doing EHR work at midnight aren't waiting for a regulatory fix. The billing staff manually copying data between systems aren't waiting for Epic to build a better API. They need a computer use agent that can do the work today, on the actual software they're already using. That technology exists in 2026. The gap is that most healthcare organizations are either still running 2021-era RPA bots that break constantly, or they're waiting for some perfect integrated solution that will never arrive. Stop waiting. The administrative waste is real, the burnout is real, and the tools that can actually address it are finally good enough to trust. If you want to see what a computer use agent that scores 82% on OSWorld can do for your workflows, go to coasty.ai. The free tier is there. Start with prior auths. Your midnight-typing physicians will thank you.