Insurance Claims Are Still Processed Like It's 1987. A Computer Use AI Agent Can Fix That.

The UnitedHealthcare Disaster Is a Warning, Not a Blueprint

Let's talk about the elephant in the room. UnitedHealthcare built an AI system to automate claims decisions. Sounds great on paper. In practice, their algorithm reportedly denied post-acute care claims at a 90% rate, and a federal class action lawsuit alleges those denials were medically unjustified and driven by a system that was never designed to actually evaluate individual patient needs. As of February 2025, that class action was allowed to proceed. Humana and Cigna are facing similar suits. The CEO of UnitedHealth was assassinated in December 2024, and the public rage that followed wasn't just about one man. It was about years of people feeling like a black-box algorithm decided their fate. This is what happens when you automate the wrong thing, in the wrong way, with zero accountability baked in. The goal was never to help patients. It was to reduce payouts. That's not automation. That's a denial machine with a PR problem. But here's the thing: the backlash against that specific abuse of AI is now bleeding into skepticism about all insurance automation, and that's a shame. Because the right kind of AI automation, the kind that handles the actual administrative grind, doesn't deny anyone anything. It just stops making humans do work that software can do better and faster.

What Manual Claims Processing Actually Costs You

• ●30% of a P&C claims adjuster's time is spent on manual data entry, not actual claims judgment (Agentech, 2025)
• ●Complex claims can take days to weeks to settle manually, even when the outcome is obvious
• ●Deloitte estimates AI-driven claims automation can reduce processing costs by up to 70%
• ●McKinsey found a UK insurer lost over £60 million ($82 million) in 2024 to inefficiencies that automation could address
• ●95% of GenAI pilots fail, according to MIT Media Lab, usually because companies automate the wrong layer
• ●Administrative expenses in US healthcare alone run into the hundreds of billions annually, a significant portion tied directly to claims handling
• ●Adjusters experiencing data overload and burnout are making slower, less accurate decisions on complex cases, per a 2025 Atom Advantage report

Why RPA Failed Insurance and Why AI Agents Are Different

The insurance industry already tried to automate this stuff. RPA, robotic process automation, was supposed to be the answer back in the late 2010s. Companies like UiPath sold the dream of software bots handling repetitive tasks. And look, RPA works fine when nothing ever changes. The moment a vendor updates their portal UI, a form adds a new field, or a workflow deviates from the script, your bot breaks. It doesn't adapt. It just stops. Then someone has to fix it, which means you've now added a bot maintenance team on top of your original claims team. Congratulations, you've automated your way into more complexity. The real difference with a modern computer use AI agent is that it sees the screen the way a human does and figures out what to do next. It's not following a brittle script. It's navigating a real desktop, a real browser, a real terminal, reading context and making decisions. When the UI changes, it adapts. When a form looks different than last week, it handles it. That's a fundamentally different architecture from RPA, and it's why the a16z team wrote in early 2025 that computer-using AI is positioned to disrupt the entire BPO industry, including insurance claims processing and revenue cycle management. The wave is here. The question is whether you're riding it or drowning in it.

What Real Computer Use Automation Looks Like for Claims

Here's a concrete picture of what a properly deployed computer use agent does in an insurance context. A first-notice-of-loss comes in. The agent opens the email, reads the attachments, cross-references the policy details in the legacy system, pulls the relevant coverage rules, populates the intake form, flags any discrepancies, and routes the claim to the right adjuster queue with a pre-filled summary. All of that used to take an adjuster 20-40 minutes of context-switching between five different applications. The agent does it in under two minutes, consistently, without getting tired at 4pm on a Friday. For simpler claims, like a clear-cut auto glass replacement or a travel delay reimbursement, the agent can handle the entire workflow end to end: verify coverage, calculate payout, trigger payment, send the customer confirmation. No human needed. For complex claims, the agent handles all the prep work so the adjuster walks in with a complete picture instead of a pile of raw documents. The adjuster's job becomes actual judgment, not data shuffling. That's the version of automation that helps people instead of harming them. It doesn't make the denial decision. It makes the human who makes that decision faster and better-informed.

Why Coasty Is the Computer Use Agent Built for This

I've looked at the options. Anthropic's computer use capability is genuinely impressive in demos and genuinely inconsistent in production workflows. OpenAI's Operator is still finding its footing. UiPath is bolting AI onto RPA infrastructure that was never designed for this, and calling it agentic. None of them are scoring 82% on OSWorld, the benchmark that actually measures how well a computer use agent handles real-world desktop tasks across browsers, terminals, and native apps. Coasty is. That's not a marketing claim. It's a verified score on the most rigorous computer use benchmark that exists right now, and it's higher than every competitor on the leaderboard. What that means practically: when you deploy Coasty on your claims workflows, it actually completes the tasks. It navigates your legacy portal. It reads the PDF, opens the right tab, fills the right field, and moves on. It runs as a desktop app, in cloud VMs, and in parallel agent swarms when you need to process a high volume of claims simultaneously, like after a hurricane or a major weather event. There's a free tier if you want to test it without a procurement process. BYOK is supported if your security team needs API key control. It's built for the people who are tired of watching bots break and pilots fail. The 82% on OSWorld isn't the ceiling either. It's just where the bar sits today, and Coasty is already above it.