Once AI Learns to Click, Bad UI Is No Longer a Moat

OpenAI's computer-use agent can now operate almost any application like a human, including software that has never exposed an API. The team leads say they are close to surpassing human experts in speed.

The two leads are Ari Weinstein and James Sun. Ari works on consumer computer-use products, and James works on enterprise. Before joining OpenAI, both founded startups. They talked for more than 40 minutes on the a16z podcast, covering changes the Astra model brings to computer use, how enterprises control invisible agents, and why computer use can become the universal API for legacy ERP and financial systems.

Several points from this podcast are worth calling out separately.

Astra turned computer use from a demo into a product. The agent's ability to understand the screen, and the speed and accuracy of turning screenshots into actions, have taken a big step forward.

Computer use is a true universal API. In the past, to automate a piece of software, you had to wait for the vendor to open an API (programmatic interface). The interface either didn't exist or was absurdly expensive. Computer-use agents reverse this logic: no interface needed; just look at the screen and click like a human. One commenter pointed out: operating through the UI essentially turns every closed platform into an open integration.

But enterprises want determinism and auditability. Once speed improves, control becomes the new bottleneck. One commenter said that on a 2009-style ERP screen, every click needs a log. Selecting the wrong vendor in a dropdown may sit in the books for a month before being discovered. Enterprises won't give up auditability just because it's faster.

What about ad-supported websites? One commenter asked: if agents read pages on people's behalf, who sees the ads that support reporting? This commenter runs four Romanian ad-supported news sites. Local news survives on advertising; agent traffic doesn't pay the bills.

No one could answer that question. For content sites, agent traffic is both an opportunity and a threat.

The most practical use cases are extremely boring. Expense reports, reconciliation, procurement, order entry. These tasks aren't sexy, but they are high-volume, and people do them slowly and annoyingly. One commenter said, I just want to throw invoices at it and never fill out an expense report again. This is exactly where computer-use agents are most deployable today.

UI won't disappear, but it will change. A counterintuitive judgment: cheap software will make UI better and more personalized. Because when users expand from humans to agents, interfaces must serve two kinds of customers at once. Agents need structured, stable interfaces; humans need handy, personalized interfaces. These two directions are actually pushing UI forward.

The sharpest comment: legacy ERP vendors are about to discover that bad UI is their only moat.

That makes sense. In the past, replacing a hard-to-use old system cost so much that no one wanted to touch it. Now that agents can operate directly on old systems, the urgency to replace them has actually decreased. Computer-use agents both extend the life of legacy software and accelerate the arrival of new AI-native applications.

Ari and James call this computer use's scaling moment: models get better, execution gets faster, security controls get stronger—all three arrive at the same time.

The technology curve has hit an inflection point. The remaining questions are how fast, how cheap, and how controllable.

Full discussion on YouTube (44 minutes): https://youtu.be/9v42ZOdiBJY