A year ago, the best AI system trying to operate a computer on its own succeeded on 42 out of 100 basic office tasks. Today, the leading system succeeds on 85. Human testers doing the same tasks score around 72. That crossover is the real story: AI that can open a browser, log into a portal, fill out a form, and move data between systems has gone from a shaky demo to something companies now run at scale.
The test behind these numbers, OSWorld-Verified, asks an AI system to complete everyday desktop chores on real operating systems: updating records, filling forms, moving files. A new report from venture firm Andreessen Horowitz, based on talks with founders and buyers running this technology, argues this jump in accuracy pushed computer-operating AI from something IT teams experimented with into something running unattended in production.
The examples are concrete. One data company that gathers pricing information from retailer websites uses an AI agent as a repair crew for its scrapers: when a retailer redesigns its site overnight and breaks the automated data collection, the agent notices, fixes the process itself, and keeps data flowing before a human engineer sees an alert. That company cut its scraper-maintenance team in half. A separate IT outsourcing firm runs 27 agent workflows today, handling between 1,500 and 2,100 support tickets a day, aiming to move a fifth to a quarter of its staff off that low-margin ticket work entirely.
The economics explain the timing. Running one of these agents costs roughly six to eight dollars an hour, close to what an offshore outsourcing worker costs and far below the 30 to 45 dollars an hour a fully loaded US back-office employee costs once benefits and overhead are added. The agent works around the clock and adds capacity without a hiring process.
This should worry the outsourcing industry many companies depend on to keep costs down. The global business process outsourcing industry, the one that answers customer service calls, processes claims, and enters data for a fee, was valued at over 328 billion dollars in 2025. India's slice of that industry employs millions and has been described as facing an existential shift as AI does work once done by call center staff. The Philippines, where the sector employs close to 1.8 million people and generates 38 billion dollars a year, faces the same pressure.
None of this means AI agents are ready for everything. They still fail on roughly 15 out of 100 tasks, and a process only counts as automated if every step works, not most of them. Failures cluster where nobody can easily check whether the output was correct. An agent that misreads payment terms on a contract, or files an insurance claim that later needs a phone call to fix, will not know it made a mistake. It looks done. It is not.
The lesson for any business leaning on outsourced back-office work, whether claims processing, order entry, ticket handling, or portal-based data collection, is that the technology to replace a chunk of that work already exists and already costs less than paying people to do it by hand. The winners will not be the companies with the fanciest AI model. They will be the ones that build the surrounding plumbing: knowing exactly how the work should be done, catching mistakes before they become a problem, and giving a human a clear way to step in when something goes wrong. That is a solvable problem, and it is being solved right now, one back-office workflow at a time.