Insurance Claims Are a $25 Billion Dumpster Fire. A Computer Use AI Agent Can Fix It.

The UnitedHealthcare Disaster Is a Warning, Not a Blueprint

By now you've heard the story. UnitedHealthcare deployed an AI algorithm called nH Predict to review post-acute care claims for Medicare Advantage patients. A class action lawsuit filed in federal court in Minnesota alleged the algorithm wrongfully denied care to seriously ill elderly patients at a 90% rate, overriding doctor recommendations at scale. The U.S. District Court allowed the case to proceed in February 2025. STAT News won a Pulitzer Prize finalist nod for their 'Denied by AI' investigation. Cigna got caught in a separate lawsuit for denying over 300,000 claims in bulk. This is what happens when you use AI as a cost-cutting weapon instead of a productivity tool. The backlash was enormous, the legal exposure is real, and the reputational damage is permanent. But here's the thing: the lesson isn't 'don't use AI in insurance.' The lesson is 'don't use AI to make decisions it can't justify.' Using a computer use AI agent to handle the tedious, manual, data-entry-heavy parts of claims processing is a completely different proposition. One replaces human judgment. The other replaces human copy-pasting. Those are not the same thing.

The Real Numbers Behind Manual Claims Processing Are Embarrassing

• ●Manual claims processing costs between $2.05 and $10.00 per claim. Automated processing costs $0.85 to $1.58. That's a 5x cost gap that insurers are choosing to ignore.
• ●Health insurance claims have a 19.3% average error rate in manual processing. Nearly 1 in 5 claims has a mistake in it.
• ●Claims adjudication costs providers $25.7 billion annually, with $18 billion classified as potentially unnecessary expense (Premier Inc., Feb 2025).
• ●Gartner found that over 40% of agentic AI projects will be canceled by end of 2027, mostly because companies built on RPA and chatbot logic instead of real autonomous agents.
• ●The global AI insurance market was pegged at $340 billion for 2025. The money is flowing. The results are not always following.
• ●A claims adjuster still spends the majority of their day on manual data extraction, form reading, and system navigation. That's not a job. That's a punishment.

Why RPA Failed Insurance and Why Everyone Pretended It Didn't

Robotic Process Automation was supposed to be the answer. UiPath, Automation Anywhere, Blue Prism, all promised insurers they could automate claims workflows without touching their legacy systems. And for a while, the demos looked great. Then reality hit. RPA bots are fragile. They break every time a UI changes. They can't handle unstructured documents, handwritten notes, or any input that deviates from the exact format they were trained on. Insurance claims are messy by nature. PDFs with inconsistent layouts, scanned documents, multi-system workflows, portal logins that expire, CAPTCHAs that block bots cold. RPA couldn't handle any of it gracefully. Gartner's 2025 data showing 40% of agentic AI projects getting canceled is partly a hangover from this era. Companies bought RPA, got burned, and are now skeptical of everything. The problem wasn't automation. The problem was that RPA was never actually intelligent. It was a macro with a marketing budget. A real computer use agent, one that can actually see a screen, navigate a browser, fill out forms, switch between applications, and handle unexpected states, is a fundamentally different technology. The comparison isn't even fair.

What AI Computer Use Actually Looks Like in a Claims Workflow

Here's what a computer use AI agent can realistically do in an insurance claims context right now, without replacing adjusters or making coverage decisions. It can log into your claims management portal, pull the new submissions, open each one, cross-reference policy data in a separate system, flag discrepancies, and populate the adjuster's review queue, all without a human touching a keyboard. It can navigate insurer portals to check claim status, download EOBs, and update internal trackers. It can handle first notice of loss intake, pulling data from emails, PDFs, and web forms and routing it to the right system. It can run eligibility checks across multiple databases simultaneously using agent swarms, something that would take a human team hours. None of this requires the AI to make a coverage decision. It's handling the mechanical, repetitive, error-prone work that currently eats 60-70% of a claims professional's day. The adjuster still makes the call. The AI does the grunt work. That's the correct division of labor, and it's the one the industry should have built toward from the start.

Why Coasty Is the Right Tool for This

I'm going to be direct here. Most 'AI automation' tools for insurance are either glorified chatbots, brittle RPA wrappers, or API integrations that only work with systems that have modern APIs (which most insurance legacy stacks do not). Coasty is a computer use agent. That means it controls a real desktop, a real browser, and real terminals, the same way a human would. It doesn't need an API. It doesn't need a custom integration. It navigates your actual software interface. Coasty scored 82% on OSWorld, the standard benchmark for AI computer use in real-world environments. That's the highest score of any computer use agent available. Anthropic's Claude Sonnet 4.5 made headlines for its OSWorld improvements. OpenAI Operator exists. But nobody else is at 82%. That gap matters when you're running claims workflows that involve 12 different browser tabs, a legacy desktop app from 2009, and a PDF that someone scanned sideways. Coasty also supports agent swarms, meaning you can run parallel instances handling multiple claims simultaneously. You get a desktop app, cloud VMs, a free tier to start, and BYOK if you want to keep your own model keys. For an industry drowning in manual work and burned by bad automation promises, that's not a pitch. That's a lifeline. Check it out at coasty.ai.