Your Insurance Claims Team Is Drowning in Copy-Paste Work, and an AI Computer Use Agent Already Fixed This
Insurance carriers spent over $1.3 trillion processing claims last year, and somewhere between 25% and 40% of that operational cost traces back to manual data work that a reasonably smart piece of software could handle. Not someday. Right now. Today. We're talking about claims adjusters who spend literal hours every single day logging into three different legacy portals, pulling policy data, copying it into a spreadsheet, pasting it into a claims management system, and then emailing a PDF to someone who will do the same thing in reverse. It's 2025. This is still happening at scale. And the insurance industry's answer, for most of the last decade, has been to buy a RPA tool, watch it break every time someone changes a UI, and then quietly hire more people to handle the exceptions. That cycle is over. AI computer use agents have arrived, and the gap between carriers using them and carriers ignoring them is about to become a competitive canyon.
The Dirty Secret Behind '30-Day Claim Settlements'
Ask any claims operations lead where the time actually goes, and they'll tell you the same thing off the record. The investigation itself, the actual judgment call about liability and payout, takes hours. Maybe a day or two for complex cases. The other 28 days? That's queue time, data entry backlogs, portal timeouts, re-keying errors that trigger reviews, and a genuinely absurd amount of time spent navigating software that was never designed to talk to other software. McKinsey has estimated that up to 40% of insurance work activities could be automated with current technology. Accenture puts the potential cost reduction from claims automation at 30% or more. The Coalition Against Insurance Fraud estimates that claims processing inefficiencies cost the U.S. industry alone over $80 billion annually when you factor in leakage, fraud that slips through manual reviews, and straight-up operational waste. Eighty billion. And the industry response has largely been to deploy brittle RPA bots that handle exactly one workflow, in exactly one system, until someone updates the UI and the whole thing falls apart at 2am on a Tuesday.
Why RPA Failed Insurance (And Everyone Pretends It Didn't)
UiPath, Automation Anywhere, Blue Prism. These tools had a real moment. The pitch was simple: record what a human does on screen, replay it forever. Insurance bought in hard. And for narrow, perfectly stable workflows, it kind of worked. But insurance claims are not narrow or perfectly stable. A vendor portal updates its login flow. A state regulatory form adds a new field. A claimant submits a PDF in a format the bot wasn't trained on. And suddenly your 'automated' workflow is a frozen bot and a human firefighter. Industry surveys consistently show that 30% to 50% of RPA implementations fail to deliver their projected ROI, with maintenance costs often eating 30% to 40% of the original build cost every single year. The dirty truth is that most large carriers have a graveyard of half-working bots that their IT teams quietly babysit. They didn't automate their claims process. They automated a screenshot of their claims process, and then hired people to fix it when reality changed. That's not automation. That's expensive theater.
Up to 40% of insurance work activities could be automated with technology that already exists, according to McKinsey. The average carrier is capturing maybe 10% of that. The other 30% is sitting on the table right now, waiting for someone to stop buying RPA maintenance contracts and start using actual AI.
What 'Computer Use' Actually Means for Claims Workflows
Here's where things get genuinely interesting. The new generation of AI computer use agents doesn't work by recording clicks and replaying them. A computer use agent actually sees the screen, reads what's on it, reasons about what needs to happen, and takes action. It handles unexpected UI changes. It reads a PDF that it's never seen before. It navigates a legacy portal with no API because it doesn't need an API. It's doing what a human does, just faster and without bathroom breaks. For insurance claims, this is enormous. Think about first notice of loss intake: a computer-using AI can receive an email with attachments, open the attachments, extract the relevant claim details, log into the carrier's claims management system, create a new claim record, cross-reference the policy database to verify coverage, flag discrepancies, and route the claim to the right adjuster queue. All of it. Without a single API integration. Without a six-month IT project. Without a bot that breaks when the portal updates. That's not a demo. That's what production computer use agents are doing right now. The workflows that used to require either a massive IT build or a human being clicking through screens for 45 minutes can be handed off to a computer use agent that finishes in under three minutes.
Anthropic and OpenAI Tried This. Here's What Actually Happened.
To be fair, the big labs saw this coming. Anthropic launched Computer Use with Claude in late 2024. OpenAI shipped Operator. Both are impressive demos. Both have real limitations in production environments that nobody in the press wants to talk about honestly. Anthropic's Computer Use is powerful but it's a capability, not a product. You're building on top of it. You're managing infrastructure, handling failures, building retry logic, and figuring out how to run things in parallel when you have 500 claims to process simultaneously. OpenAI's Operator is consumer-facing and genuinely good at booking restaurants. Enterprise claims workflows are a different category of problem entirely. The gap between 'impressive API capability' and 'production-ready agent that runs unsupervised on real business workflows' is enormous, and most enterprises have learned that the hard way after burning months on integration projects. The OSWorld benchmark, which is the most rigorous independent test of how well AI agents actually operate computers in real-world conditions, tells the story clearly. Coasty scores 82% on OSWorld. That's not a marketing number. That's a benchmark score on tasks that include navigating real desktop apps, handling unexpected states, and completing multi-step workflows without human intervention. For context, that's higher than every other agent on the benchmark. When you're processing claims at scale, that accuracy gap is the difference between a workflow that runs and a workflow that requires a human to audit every single output.
Why Coasty Is the Obvious Answer for Claims Teams That Are Done Waiting
I've seen a lot of automation tools come through insurance operations. Most of them solve one specific problem and create three new ones. Coasty is built differently, and the reason I keep recommending it to anyone who asks is that it operates at the level where insurance complexity actually lives. It controls real desktops, real browsers, and real terminals. It doesn't need your legacy system to have an API. It doesn't need your vendor portal to cooperate. It navigates them the same way a human would, except it does it in seconds and it doesn't make transcription errors. For claims specifically, the use cases that move the needle fastest are first notice of loss processing, coverage verification across policy systems, subrogation research, medical bill review intake, and status update communications. These are all tasks where a computer use agent replaces hours of human screen time per claim, per day. Coasty also runs agent swarms, meaning you can process hundreds of claims in parallel across cloud VMs without building any infrastructure yourself. The free tier lets you test it on your actual workflows before you commit to anything. BYOK support means your data isn't going somewhere you don't control. The 82% OSWorld score means that when your portal does something unexpected, the agent figures it out instead of freezing and waiting for someone to notice at 2am. That's what production-grade computer use looks like.
Here's my actual opinion, and I'll stand behind it: any insurance carrier still running manual claims intake and data entry workflows in 2025 is making a strategic choice to be slower and more expensive than their competitors. That's it. The technology exists. The benchmark scores are public. The ROI math is not complicated. A claims processor handling 40 claims a day spends, conservatively, 90 minutes of that day on pure data movement between systems. Multiply that across a team of 50 adjusters. That's 75 hours a day of work that an AI computer use agent handles for a fraction of the cost, without errors, without sick days, and without needing a six-month IT integration project to get started. The carriers who figure this out in the next 12 months will have a structural cost advantage that's very hard to close. The ones who wait will spend 2027 trying to catch up. Go test it yourself at coasty.ai. The free tier is real. The benchmark scores are real. The only thing that isn't real is the excuse that this isn't ready yet.