UiPath Is Dying and AI Agents Are Why: The Computer Use Revolution Nobody Wants to Admit
Manual data entry costs U.S. companies $28,500 per employee per year. That's not a typo. Twenty-eight thousand dollars, per person, per year, just from copying and pasting data between systems. And the tool most enterprises bought to fix this problem, UiPath, just had its stock collapse 34% in a single trading day, is now named in a securities fraud lawsuit, and is posting revenue growth of a measly 5% year-over-year while the rest of the AI world is exploding. Something is very wrong with the RPA story. The companies that got sold on it in 2018 are starting to realize it. And the rise of real computer use AI agents is why the whole thing is unraveling.
The RPA Promise Was Always a Little Dishonest
Here's what the UiPath sales pitch sounded like in 2019: 'We'll build you a bot. It'll do the repetitive work. You'll save money and your employees will love you.' And for about six months, it worked great. Then someone changed the font on a login page. Or an app updated its UI. Or a vendor switched from one web portal to another. And the bot, the expensive, painstakingly configured bot, just stopped working. This isn't a fringe complaint. Industry data shows that 30 to 50 percent of RPA projects fail outright, and the ones that don't fail require constant maintenance to keep running. Blueprint Systems documented this years ago and nothing has changed. Every time your underlying software updates, someone has to go back into the bot and fix the coordinates it was clicking on. RPA is essentially a very expensive, very fragile macro that pretends to be automation. The hidden maintenance costs alone can run over $100,000 per implementation when you factor in developer time, licensing, and the inevitable firefighting. You didn't automate the work. You just made it someone else's problem.
UiPath's Numbers Tell a Brutal Story
- ●Stock dropped 34% in a single day in May 2024, from $18.30 to $12.07 per share
- ●Revenue growth slowed to just 5% year-over-year in fiscal 2025, down from double digits
- ●Full-year guidance was even worse: 7% projected growth, in an AI boom year
- ●The company is now named in a securities fraud class action lawsuit filed in 2024
- ●The lawsuit alleges executives discussed demand challenges internally while publicly projecting confidence
- ●Founder Daniel Dines had to come back as CEO in a 'boomerang' move, a classic sign of a company in trouble
- ●Gartner predicts over 40% of agentic AI projects will be canceled by 2027, many of them legacy RPA pivots that don't actually work
62% of employees spend the majority of their work time on repetitive tasks. UiPath's answer to that problem is bots that break 30-50% of the time and cost over $100,000 to implement and maintain. That's not automation. That's a more expensive version of the original problem.
What a Real Computer Use Agent Actually Does Differently
Here's the fundamental difference between RPA and a modern computer use agent, and it matters a lot. RPA works by memorizing coordinates. Click here. Type this. Wait for this element to appear. It's rigid, it's brittle, and it has zero understanding of what it's actually doing. A computer use AI agent, on the other hand, sees the screen the same way a human does. It reads context. It adapts when things change. If the button moved, it finds the button. If the workflow has an unexpected popup, it handles it. This isn't a marginal improvement. It's a completely different architecture. Real computer use means the agent can operate any desktop app, any browser, any terminal, without needing a custom integration or a developer to hardcode every possible UI state. The agent figures it out, the same way a competent new hire would figure it out on their first day. That's what makes the comparison to UiPath so lopsided. One of these tools requires months of setup and a dedicated maintenance team. The other one you point at a task and it runs.
The Benchmark That Should End This Debate
OSWorld is the standard benchmark for AI computer use. It tests agents on 369 real desktop tasks across file management, web browsing, and multi-app workflows. It's the closest thing we have to a real-world stress test for computer-using AI. Most agents score in the 30 to 50 percent range. Anthropic's Claude Sonnet 4.5 made headlines when it posted a significant improvement on OSWorld, and OpenAI Operator gets talked about constantly in breathless press releases. But here's what the press releases don't lead with: these are general-purpose models being asked to do computer use as a side feature. They weren't built for this from the ground up. Coasty, which was built specifically as a computer use agent, sits at 82% on OSWorld. That's not a rounding error above the competition. That's a fundamentally different level of reliability. When you're running 500 tasks in parallel across a business, the difference between 50% and 82% success rate isn't an inconvenience. It's the difference between automation that works and automation theater. UiPath can't benchmark on OSWorld because its bots aren't intelligent agents. They're scripts. There's no score for 'breaks when the UI updates.'
Why Coasty Exists
I'm not going to pretend I stumbled onto Coasty by accident. I was looking for something that could actually do what RPA promised but never delivered: reliable, adaptable, zero-maintenance computer use at scale. Coasty controls real desktops, real browsers, and real terminals. Not API wrappers. Not integrations that require a developer to build and maintain. Actual computer use, the same way a human operates software. The 82% OSWorld score matters because it's an honest number from an honest benchmark, not a cherry-picked demo. The agent swarms feature means you can run tasks in parallel across cloud VMs without hiring a team to manage the infrastructure. There's a free tier so you can test it on real work before committing to anything. BYOK support means you're not locked into someone else's model choices. The reason Coasty exists is because the people who built it watched enterprises spend millions on UiPath and Automation Anywhere and get back brittle bots that needed babysitting. The whole point was to build a computer use agent that actually adapts, actually scales, and actually works when the software it's operating changes. That's a low bar in theory. It turns out it's a very high bar in practice, which is why most competitors are nowhere near 82% on OSWorld.
UiPath isn't going to zero tomorrow. Enterprise contracts are sticky and IT departments are slow. But the trajectory is obvious. When your revenue growth is 5% in a year where every AI company is posting 50 to 100 percent growth, you're not competing. You're surviving. The companies that are going to win the next decade of automation are the ones that stop treating 'we have bots' as a strategy and start deploying actual computer use AI that can adapt, reason, and scale without a maintenance team. The $28,500 per employee wasted on manual tasks isn't going to fix itself with a fragile script that breaks when someone changes a button color. If you're still evaluating UiPath in 2025, you're not doing due diligence. You're just comfortable with the familiar. Go try a real computer use agent. Start at coasty.ai. The free tier is right there. Run it on something you've been meaning to automate for six months and see what happens. I'll wait.