An $8-an-Hour AI Clicks for You—Who's Accountable When It Clicks Wrong?
OpenAI's Ari Weinstein says computer use is already faster than the average human, with costs driven down to $6–$8 per hour—cheaper than Indian outsourcing. But error logs, audit trails, and accountability in real work won't automatically disappear just because it's faster.
a16z posted a cost comparison chart. Computer-use agents cost $6 to $8 per hour, Indian outsourcing is about $10 per hour, and onshore U.S. back-office roles are $30 to $45.
OpenAI's head of consumer computer use, Ari Weinstein, said in an interview: “We’ve already gotten computer use to the point where it’s faster than the average human using a computer, and I think we’re very close to superhuman.”
The scenarios they discuss are very concrete: old ERPs with no API, financial reconciliation, procurement processes, expense reports. These kinds of software exist in huge numbers inside real companies; it’s just that no one has ever written an interface for them. The agent directly clicks buttons, fills out forms, and selects vendors from drop-down menus—taking the same path people walk every day.
Someone in the comments put it plainly: legacy ERP vendors are about to discover that a bad UI is their only moat.
But there’s a gap in this logic. A reply from someone named George said that on a 2009 ERP interface, he wants a log for every click. Pick the wrong vendor and it sits on the books for a month before finance notices.
Speed is the selling point. Error rate is the cost. The two are not in the same chart.
The a16z interview also mentions that enterprises need deterministic guardrails, because once agents are running, no one is watching. You can tell they know where the problem is. But by putting “close to superhuman” in the headline and “$6 to $8 per hour” in the first line, the things that actually make people hesitate get pushed to the back.
One more comment deserves to be called out separately: “AI executes, humans set direction.” It sounds like a division of labor, but it hides a harder question: how does the person setting direction know that the execution layer picked the wrong option from the drop-down?
Cost comparisons are calculated per hour, but the real ledger should be calculated per completed task. The time it takes to fix one erroneous ticket might be enough for an agent to run all day.
The video is here, 44 minutes, discussing what changes the Astra model brings, how enterprises should set guardrails, and why UI won’t disappear.
YouTube: https://youtu.be/9v42ZOdiBJY
One last thing the interview doesn’t expand on. As the cost curve goes down, what companies will really change is not “whether to hire people,” but the accounting method of “paying by the hour” itself. Settle by task, settle by outcome, adjust contracts by error rate. Once the price anchor changes, the books for many roles in the office will have to be recalculated.
Not just back office.
发布时间: 2026-10-27 21:46