Insurance Claims Are a $25.7 Billion Dumpster Fire, and Your AI Agent Is the Only Way Out

The 'AI' Insurers Are Using Is Not What You Think

When UnitedHealthcare, Cigna, and Aetna talk about AI in claims, they don't mean a thoughtful system that reads context and makes smart decisions. They mean algorithmic denial machines. UnitedHealthcare's claim denial rate sits at 32%, according to data that went viral on Reddit in late 2024 and has since been cited in congressional hearings. Anthem is at 23%. Aetna at 20%. A class-action lawsuit advancing through federal courts in February 2025 alleges that UnitedHealthcare's AI algorithm wrongfully denied claims at scale, and that when patients appealed, nine out of ten denials were reversed. Nine out of ten. That means the AI was wrong 90% of the time on appealed cases, and the only thing standing between patients and wrongful denial was the willingness to fight. Most people don't fight. The insurers know this. This is not claims automation. This is automated harm dressed up in a press release. The American Medical Association said in February 2025 that physicians are 'deeply concerned' that unregulated AI is actively increasing prior authorization denials. Health Affairs published a piece in January 2026 calling it an 'AI arms race' in utilization review. Stateline reported in November 2025 that patients are now deploying their own bots to battle insurer AI. Bots fighting bots. Over whether a human gets medical care. Welcome to 2025.

What's Actually Broken in Claims Processing (The Boring Stuff That Costs a Fortune)

• ●$25.7 billion: the annual cost of claims adjudication for providers in the US, per Premier Inc. (February 2025)
• ●$18 billion of that is potentially avoidable, caused by unnecessary denials, rework, and administrative back-and-forth
• ●UnitedHealthcare denies 32% of claims, with a 90% reversal rate on appeals, meaning most denials are wrong and just never get challenged
• ●Manual claims processing involves logging into multiple portals, re-entering data, downloading EOBs, cross-referencing policy documents, and filing appeals, often across 4 to 6 different systems with no integration
• ●RPA tools like UiPath can automate rule-based steps, but they break the moment a portal changes its UI, a PDF format shifts, or a new insurer workflow gets introduced
• ●The average claims specialist spends an estimated 30 to 40% of their day on tasks that are pure data movement, not judgment, not expertise, just copy-paste between systems
• ●Operational inefficiencies in insurance run at 3 to 6% of premiums for casualty and motor claims, and 6 to 9% for property claims, per Higson's 2025 analysis

Why RPA Is Not the Answer (And Never Was)

Every major insurer and healthcare provider that went all-in on RPA for claims processing in the 2018 to 2022 era is now dealing with a maintenance nightmare. RPA bots are brittle. They follow exact pixel coordinates and DOM structures. The moment an insurer updates their claims portal, which happens constantly, your bot breaks. You file a ticket. Someone fixes it two weeks later. In the meantime, claims pile up. This is why a 2025 LinkedIn analysis on agentic AI noted bluntly that 'RPA has no built-in intelligence.' It can move data if nothing changes. But insurance claims don't live in a static world. Portals change. PDFs have inconsistent formats. Adjusters ask follow-up questions through secure messaging systems. Policies get updated mid-year. A real automation solution for insurance claims needs to read, reason, navigate, and adapt. It needs to actually use a computer the way a human does, seeing the screen, understanding context, and making decisions. That's not RPA. That's a computer use agent.

What a Real Computer Use Agent Actually Does for Claims

Here's what proper AI computer use looks like in an insurance claims workflow. Instead of a brittle bot that clicks on hardcoded coordinates, a computer use agent opens the insurer portal, reads what's actually on the screen, finds the claim, pulls the EOB, cross-references it against the policy document, identifies the denial reason, drafts the appeal with the correct clinical language, and submits it, all without a human touching it. If the portal has changed since last week, the agent adapts. If there's a CAPTCHA or a multi-step authentication flow, it handles it. If the claim requires supporting documentation from three different systems, it navigates all three. This is the difference between automation that works in a demo and automation that works on a Tuesday afternoon when everything is slightly broken. The AI arms race Health Affairs described is real, and the only way providers and payers on the right side of it win is by deploying AI that can genuinely operate software at the level of a trained human, not just trigger pre-scripted macros.

Why Coasty Is the Computer Use Agent Insurance Teams Actually Need

I'm not going to pretend there aren't options here. There are. Anthropic has Claude's computer use. OpenAI has Operator. But if you're making a decision about what to run on real production workflows, the benchmark scores matter, and Coasty scores 82% on OSWorld, the standard benchmark for real-world computer use tasks on actual desktop software. Claude Sonnet 4.6 scores 61.4%. GPT-5.4 recently hit 75%. Coasty is at 82. That gap is not marketing. That's the difference between an agent that handles your claims portal 8 times out of 10 versus one that handles it 6 times out of 10. At scale, across thousands of claims per month, that gap is enormous. Coasty controls real desktops, real browsers, and real terminals. It's not making API calls to a sanitized integration layer. It's actually using the software the way your team does. It supports agent swarms for parallel execution, so you're not processing claims in a single-threaded queue. It has a desktop app, cloud VMs, and a free tier to start. BYOK is supported if your compliance team has opinions about where API keys live, and they always do. For insurance teams that are drowning in portal-hopping, appeal filing, and EOB reconciliation, this is the tool that actually works on the messy, inconsistent, constantly-changing software reality of the industry.