Insurance Claims Are a Broken Mess. A Computer Use AI Agent Can Fix Them (Without a 90% Error Rate).
UnitedHealth built an AI to process Medicare Advantage claims. It had a 90% error rate. They used it anyway. Elderly patients were denied medically necessary care, appeals piled up, and the company is now fighting a federal class-action lawsuit that a judge let move forward in February 2025. This is the insurance industry's AI story in one paragraph: deploy something broken, hide behind the algorithm, and let sick people absorb the consequences. But here's the thing, that disaster has almost nothing to do with what good AI automation for insurance claims actually looks like. The industry got seduced by the wrong kind of AI, the kind that makes decisions instead of doing work, and it's paying for it in courtrooms and reputation. There's a better path. It just requires understanding what a real computer use agent actually does.
The Insurance Claims Mess Nobody Wants to Admit
Insurance claims processing is, by almost any measure, an embarrassing industry. The average claims adjuster spends the majority of their day doing things a reasonably smart intern could do: copying data between systems, logging into portals, downloading PDFs, uploading the same PDFs somewhere else, sending status emails, and waiting. A 2025 report from Washington State found that a single average claim requires 84 minutes of staff time just to process. Multiply that across millions of claims a year and you're not talking about inefficiency, you're talking about a structural waste machine. McKinsey estimates that AI automation across insurance could unlock hundreds of billions in value globally, and yet most insurers are still running on a patchwork of legacy systems, manual workflows, and RPA bots that break every time someone changes a web form. The industry knows it has a problem. It just keeps reaching for the wrong solutions.
Why RPA Was Never the Answer (And Everyone Knew It)
- ●RPA bots are brittle by design. Change one field label in a web portal and your entire claims workflow breaks. Insurance portals change constantly.
- ●The global RPA market hit $23 billion in 2024, yet failure rates remain brutal. Gartner predicts over 40% of agentic AI projects will be outright canceled by end of 2027, many of them RPA replacements that over-promised.
- ●RPA requires exhaustive rule-writing upfront. Every exception, every edge case, every new document format needs a new script. Claims are full of exceptions.
- ●90% of agentic AI implementations fail because companies treat them like traditional automation, according to a 2025 analysis by Beam AI. Same rigid logic, fancier branding.
- ●UiPath's own blog has a post titled 'Why RPA Deployments Fail.' That's not a competitor talking. That's the vendor admitting the problem exists.
- ●RPA can't see a screen the way a human does. It can't navigate a UI it hasn't been explicitly trained on. A computer use agent can.
UnitedHealth's nH Predict AI had a 90% error rate on appeals, meaning 9 out of every 10 denials it issued were wrong. The company kept using it. A federal judge ruled the class-action lawsuit can proceed. This is what happens when you use AI to make decisions instead of using AI to do work.
The Real Problem: Insurance AI Is Being Used Backwards
The UnitedHealth disaster and the broader RPA graveyard share the same root cause. The industry is using AI to replace human judgment on consequential decisions, while still making humans do all the tedious mechanical work. That's exactly backwards. You don't want AI deciding whether a 74-year-old gets a hip replacement. You want AI logging into the provider portal, pulling the claim, cross-referencing the policy document, filling out the prior authorization form, and filing the appeal, all without a human spending three hours on hold. The judgment call stays with a person. The grunt work goes to the machine. That's the version of insurance automation that doesn't end in lawsuits. And it's only possible with a true computer use agent, one that can actually see and operate a real desktop environment the way a human would, not one that's just pattern-matching against pre-programmed scripts.
What Computer Use AI Actually Looks Like in a Claims Workflow
Here's a concrete picture. A new claim comes in as a PDF. A computer use agent opens the file, reads the relevant fields, navigates to the insurer's internal system, enters the data, pulls the corresponding policy from a separate database, flags any coverage mismatches, drafts a status email to the claimant, and logs everything in the CRM. No API integration required. No custom connector. No six-month implementation project. The agent uses the same interfaces a human would, because it can actually see the screen. This is what separates genuine AI computer use from the RPA era. RPA needed everything to be predictable and structured. A computer-using AI handles the messy, real world: scanned documents, inconsistent portal layouts, multi-step browser workflows, and legacy desktop software that hasn't been updated since 2011. For insurance companies still running on systems that predate the iPhone, this is not a small distinction. It's the whole ballgame.
Why Coasty Is the Computer Use Agent Built for This
I'm not going to pretend every computer use agent is the same, because they're not. Coasty scores 82% on OSWorld, the industry's hardest benchmark for real-world computer use tasks. Anthropic's Claude Computer Use, OpenAI's Operator, they're all on that leaderboard too. Coasty is at the top. That gap matters when you're running claims workflows that involve navigating five different portals, handling edge cases in scanned documents, and executing multi-step processes without a human watching every click. Coasty controls real desktops, real browsers, and real terminals. It's not making API calls and calling it automation. It's doing the actual work. For insurance teams that need to run parallel workloads, the agent swarm capability lets you process multiple claims simultaneously, something no human team can match at scale. There's a free tier to start, BYOK support if you're worried about data, and a desktop app that doesn't require a six-month IT procurement cycle. If you've been burned by RPA projects that promised the world and delivered a pile of broken scripts, Coasty is worth an honest look at coasty.ai.
The insurance industry has two AI problems. One is using AI to wrongfully deny claims and ending up in federal court. The other is spending millions on automation projects that fail before they go live. Both problems come from misunderstanding what AI should be doing. AI shouldn't be making coverage decisions. And it shouldn't be locked into brittle, script-based RPA logic that breaks the moment anything changes. What it should be doing is the exhausting, repetitive, screen-staring work that currently eats 40 to 60 percent of a claims adjuster's day. That's the version of AI automation that saves money, doesn't get sued, and actually makes the industry less terrible to interact with. The technology to do this right exists right now. The best computer use agent on the market is sitting at coasty.ai with a free tier and an 82% OSWorld score. The only question is whether your company keeps buying broken tools or starts using one that works.