"RPA is dead" — you've probably heard that a lot lately.
But UiPath's last quarterly revenue was actually $481 million, up 14% year over year. It's not dead. Something inside it is quietly dying, though.
RPA isn't dead, is it?
UiPath's numbers look fine on the surface — annual ARR of $1.853 billion, up 11% year over year. But dig one layer deeper and the story changes. Customers using AI features are spending roughly 3x more than those who aren't, and 90% of customers are already using AI features.
So what's still alive is the "RPA" category itself. What's dying is a specific approach inside it — the old-school bots that record clicks by screen coordinates and replay them exactly.
So what's actually dying?
Traditional RPA works by recording "click here, type here" and playing it back identically. The problem: it breaks the second the screen changes even slightly. Move a button or pop up a new dialog, and it's over. Some orgs running 60 bots need more developers just to keep fixing them than there are bots.
And there was a whole category RPA couldn't even touch to begin with — old software with no API at all. Legacy SAP/Oracle modules, hospital EMRs, internal groupware — the kind of thing that only worked if a human sat there clicking through it.
This is where computer-use agents take a different approach. Instead of memorizing coordinates, they "read" the screenshot. They read the label on a button, figure out what state the screen is in, and notice when an error message pops up. Which means if the layout changes and the save button moves, they can go find it again.
| Traditional RPA | Computer-Use Agent | |
|---|---|---|
| How it works | Record coordinates → replay exactly | Reads and understands the screen |
| When the UI changes | Breaks, needs rework | Re-locates labels and state |
| Legacy systems with no API | Mostly can't touch them | Operates the screen like a person |
| Maintenance burden | Scales with bot count, sometimes worse | Lower, but still early days |
So what's actually happening right now?
Anthropic opened Claude for Chrome to a small group in August 2025, then expanded it to all Pro, Team, and Enterprise plans by December. One browser extension that can look at the screen, click buttons, and fill in forms.
There's already movement on the ground in Korea, too. One savings bank built a hybrid team — 2 AI developers alongside 3 RPA staff — running both approaches side by side, and renewable energy company H-Energy is using an agent called "Helios" to automate everything from solar plant permitting to financial structuring. Gartner projects 40% of enterprise applications will ship with task-specific AI agents by 2026 — up from under 5% in 2025.
It's not free, though. In Anthropic's own red-team testing, the attack success rate for prompt injection was 23.6% with no defenses. After adding defenses, it dropped to 11.2%, and for certain browser-specific attack scenarios, it went from 35.7% to 0%. Not eliminated entirely — that's the part worth remembering.
How to actually try this at your company
- List your candidate tasks
Write down repetitive screen-based tasks that have no API. Things like capturing approval status in groupware, or pulling data from an old booking system. - Pick a tool
For file/document-heavy work, Claude tends to lead. For web forms and multi-tab research, other options compare well too. Different tools are strong in different areas. - Pilot on low-risk tasks first
Start with read/extract work. Push anything involving payments, sending, or deleting to later. - Design permissions narrowly
Restrict access by site, and require human approval for high-risk actions. Blocking financial and personal-data sites by default is a good baseline. - Double-check results early on
"Self-healing" doesn't mean fully trustworthy yet. Build in a manual review step for the first few weeks.
Want to go deeper?
The Rise of Computer Use and Agentic Coworkers The a16z piece that first mapped out this shift, going deep on the interaction frameworks and orchestration stack. a16z.com
Piloting Claude in Chrome Anthropic's official announcement, with the prompt-injection defense numbers and permission design in detail. claude.com
Anthropic launches a Claude AI agent that lives in Chrome TechCrunch's coverage of the initial launch and early reaction. techcrunch.com
Is RPA Dead? An analysis grounded in UiPath's earnings and stock reaction, on what the actual threat to the RPA market is. saasocalypse.com
RPA's endgame, agentic AI's rise A rundown of real adoption cases across Korean finance, healthcare, and energy companies. shareitem.co.kr
Computer Use Agents 2026: Claude vs OpenAI vs Gemini A comparison of which tool is strongest at which kind of task — useful when you're picking one. digitalapplied.com
