Insurance Claims Are a Broken Mess. A Computer Use AI Agent Can Fix It (But Most Won't).
Auto repair insurance claims now take an average of 23.1 days to process. That's more than double the pre-pandemic time. Not because the claims got harder. Because the industry is still duct-taping manual workflows together with spreadsheets, PDF attachments, and legacy portals that haven't been updated since Obama's first term. Meanwhile, claims adjudication is costing the U.S. healthcare system alone $25.7 billion every single year, with $1.8 billion of that flagged as potentially unnecessary expense by Premier Inc. in early 2025. Read that again. One point eight billion dollars. Wasted. Annually. On paperwork. And the punchline? Most insurers think they've already solved this problem. They haven't. They've made it worse.
The AI Denial Scandal Nobody Wants to Talk About
Here's where the story gets genuinely ugly. When insurers finally did adopt AI for claims, a significant chunk of them used it to deny claims faster, not process them better. UnitedHealth was sued for using an AI algorithm called nH Predict that critics say carries a 90% error rate. Ninety percent. Cigna faced a class-action lawsuit for allegedly using an algorithm to auto-deny over 300,000 claims. Humana got hit with a similar lawsuit. The Guardian reported in January 2025 that new tools are now being built specifically to counter health insurance denials that were decided by algorithm in seconds. The American Medical Association came out in February 2025 saying physicians are deeply concerned that AI is increasing prior authorization denials at scale. So the industry's first major swing at AI automation wasn't 'let's make claims faster for customers.' It was 'let's use a black-box model to say no, and do it so fast that nobody can appeal in time.' That's not automation. That's a liability factory. And it's exactly the wrong lesson to take from AI.
Why RPA Was Always a Band-Aid, Not a Fix
- ●Traditional RPA tools like UiPath build brittle bots that break the second a portal changes its UI. One button rename and your entire automation pipeline is down.
- ●RPA has zero ability to reason. It can't read a messy PDF, interpret a handwritten adjuster note, or handle a claim that doesn't fit the exact template it was trained on.
- ●Gartner predicted in June 2025 that over 40% of agentic AI projects will be canceled by end of 2027, largely because companies are bolting AI onto old RPA frameworks instead of replacing them.
- ●Legacy system integration is RPA's perpetual excuse. 'It'll take 6 months to connect to your claims management system.' A real computer use agent just uses the system visually, like a human does, on day one.
- ●The insurance industry spent billions on RPA in the 2010s. Claims processing times still doubled post-pandemic. That's not a coincidence. That's a verdict.
Claims adjudication costs U.S. providers $25.7 billion per year. $1.8 billion of that is flagged as potentially unnecessary. And the AI most insurers deployed first was used to deny claims with a reported 90% error rate. The industry didn't automate its way out of the problem. It automated its way into a class-action lawsuit.
What 'Computer Use AI' Actually Means (And Why It's Different)
There's a lot of noise right now about AI agents. Anthropic has Claude with computer use capabilities. OpenAI has Operator. Everyone's demoing something impressive in a controlled environment. But here's the real question: what happens when the claims portal is a 15-year-old Java applet? What happens when the adjuster workflow involves switching between four different desktop apps, a browser, a terminal, and a shared drive? What happens when the document is a scanned fax from 2019 with a coffee stain on it? Most computer use implementations from the big labs are optimized for clean, web-based tasks in demo conditions. They struggle badly with the actual messy, multi-system, multi-format chaos of real insurance operations. A genuine computer use agent doesn't just call APIs. It controls a real desktop. It sees what a human sees. It clicks, types, reads documents, navigates portals, and executes multi-step workflows across any application, old or new, without needing a custom integration for every single system. That's the difference between a demo and a deployment.
What a Real Claims Automation Workflow Actually Looks Like
Let's get concrete. A proper AI computer use agent handling insurance claims should be able to open an incoming claim email, extract the relevant data from an attached PDF or image, log into the claims management system, cross-reference policy details, flag inconsistencies or potential fraud indicators, fill in the adjuster form, escalate edge cases to a human queue, and close the ticket. All of it. Without a human touching it. Without a custom API integration for every legacy system. Without breaking when the portal updates its CSS. And critically, it should do this in parallel. Not one claim at a time. Agent swarms running simultaneously, processing hundreds of claims while your adjusters focus on the genuinely complex cases that actually need human judgment. The CAQH published data in mid-2025 showing that transitioning to electronic claim attachments alone offers a 55% cost savings compared to manual and web portal processing. That's just one step in the workflow. Imagine automating the whole thing.
Why Coasty Is the Tool Actually Built for This
I'm going to be straight with you. I've watched a lot of 'AI automation' tools get pitched to insurance teams, and most of them are glorified form-fillers with a GPT wrapper. Coasty is different in ways that matter for this specific use case. It scores 82% on OSWorld, the industry-standard benchmark for AI agents operating in real computer environments. That's the highest score of any computer use agent, and it's not close. But the benchmark score isn't why I'd recommend it for claims processing. It's the architecture. Coasty controls real desktops and browsers, not just APIs. It works with whatever system your team already uses, whether that's a modern cloud platform or a legacy desktop app from 2008. It runs agent swarms, meaning you can process claims in parallel at a scale no human team can match. And it has a free tier, so you can actually test it against your real workflows before committing. For insurance teams specifically, the ability to just point the agent at your existing systems without a six-month integration project is the thing that makes it viable where other tools stall out. You can get to coasty.ai and see what it actually does on real desktop tasks. Not a curated demo. Real work.
The insurance industry has a choice right now. It can keep doing what it's been doing: deploying brittle RPA bots that break constantly, using AI to deny claims faster instead of process them better, and watching processing times stretch to three weeks while customers lose their minds. Or it can actually use a computer use agent that was built to handle the full, messy reality of claims workflows, not just the clean parts. The $25.7 billion in annual adjudication costs isn't a fixed cost of doing business. It's a problem with a solution. The AI denial lawsuits aren't an argument against automation. They're an argument against lazy, irresponsible automation that optimizes for the wrong outcome. Done right, AI computer use in insurance claims means faster settlements, fewer errors, happier customers, and adjusters who spend their time on actual judgment calls instead of copying data between tabs. That's not a fantasy. That's what the technology can do today. The only question is whether your org has the nerve to actually deploy it. If you want to see what a real computer use agent looks like in action, start at coasty.ai.