OpenAI's Computer-Operating AI Is Faster Than Average Users, Approaching Expert Speed
OpenAI's product leads say their computer-operating agents are faster than average users and approaching expert speed; rather than building new software to replace legacy systems, they want agents to take over enterprises' existing software directly.
The video won't play here. Listen to the audio instead:
The argument · tap a timestamp to hear it
Agents that operate computers are already faster than average users
Ari says OpenAI's computer use is already faster than an average person using a computer and approaching superhuman speed—faster than any experienced user in the room. He sees this as a critical inflection: once agents outperform you at any task, the default should be to delegate to them, with humans intervening only when they specifically want to act. This isn't distant vision—it's the product inflection they expect to hit soon.
— Ari WeinsteinEnterprises need a tunable dial for human control, not just trust in the model
James says enterprise trust in agents isn't built on model sophistication but on the ability to make human involvement adjustable: which applications and websites can operate fully autonomously, and which actions—like downloads, uploads, or visiting new sites—require human approval. OpenAI built Guardian for this: a system that audits each agent action for safety and alignment with the user's original intent, letting enterprises draw their own line between autonomy and control.
— James SunAgents and ads are inherently in conflict on the open web
Agents roll out smoothly inside enterprise software but hit a business-model collision on the open internet: many sites depend on ad revenue, and when an agent replaces human browsing, ad impressions don't generate real revenue for publishers and advertisers. OpenAI's solution is a standard like WebMCP that lets sites support both human and agent use: agents handle repetitive work, humans keep the actions they want—ordering, choosing—instead of the agent replacing the whole site.
— Ari WeinsteinZero developer integration is what makes it fundamentally different
Ari stresses that computer use is wholly unlike any product OpenAI has shipped before: all prior capabilities required developers to opt in—writing connectors, running MCP servers. Computer use doesn't; it operates like a person, clicking and typing and browsing, so it works with any existing software, including systems twenty years old that were never designed for AI. That zero integration cost is why it can penetrate legacy ERP and CRM systems so quickly.
— Ari WeinsteinEnterprises will automate grunt work first: the system-record jobs
James divides enterprise work into two automation tiers. The first is tedious, high-value drudgery: procurement (manually entering thousands of line items into purchase software, then contacting suppliers individually), financial reconciliation (no API between ledgers, pure manual matching), sales (scraping call logs to populate CRM). These lack APIs and live in point-and-click workflows—exactly where computer use strikes first. The second tier—circuit design, hardware engineering—is more creative work enterprises haven't yet recognized agents can do.
— James SunLegacy software won't disappear; UI won't either—it will personalize
On whether new companies will use agents to displace legacy software, their views complement. Ari thinks software written over twenty years won't vanish overnight; computer use will instead pressure all software to be agent-friendly. James cares more about UI's intrinsic value: he built his wife a gomoku trainer because understanding how she learns only humans can do. He expects future software to be customizable for a hundred users or one; that's why UI won't disappear but will become personal.
A startup's moat is speed and being first to solve the pain
James says even with lab bandwidth stretched thin, startups have huge room, especially in verticals where AI penetration is slow. He cites a company using computer use to photograph kitchen inventory, forecast tomorrow's customer flow, auto-place orders in procurement software, and adjust the schedule—wiring together a manual chain end-to-end. Most small and mid-market firms don't even know what AI can do. The real moat isn't model capability; it's who first stitches these scattered, complex systems into one usable product.
— James SunNext: agents that come to you proactively, not waiting to be summoned
Ari says most computer use today is ‘you give a task, it runs in the background.’ Next comes two directions: true superhuman speed with real-time collaboration, and more importantly, proactivity—like Dots, where the agent finds problems and comes to you, rather than waiting for you to remember it exists. He thinks this will make the product easier for normal people to adopt because you don't learn how to use an agent—the agent learns to fit your rhythm.
— Ari WeinsteinIn their own words · checked verbatim
So, we've gotten computer use now to a place where it's faster than the average human at using a computer. And I think we're getting really close to a place where it is superhuman, meaning it is faster than an expert computer user like probably one of us at this table.
Ari Weinstein9:19
We have this really awesome system called Guardian, right? Or I think auto-approve in the app that is a really smart model looking at every single action that the computer use agent is taking to decide if it's safe, if it's aligned to their original intent.
James Sun10:23
You think about advertisers, publishers whose business model is entirely on advertisement. If the agent's using the website for them, then where's that revenue coming from?
Ari Weinstein12:29
You know, it works with every application. It's unlike any other thing that we've built where, you know, the developer has to do some work to adopt it. Someone has to build a connector or MCP server. You know, computer use works with everything.
Ari Weinstein15:47
Why not? Like, why should we settle for shitty software?
James Sun31:23
And speed, really, I think at the end of the day, speed and pain is kind of the moat that exists today.
James Sun35:28
I'm really excited about a world where the agent can sort of adapt to you instead of you adapting to it.
Ari Weinstein40:45
Glossary
- Computer Use (CUA)
- AI operating software like a person—clicking, typing, browsing—to complete tasks.
- WebMCP
- A protocol enabling humans and AI agents to use the same website together.
- Guardian
- OpenAI's internal model that reviews each agent action in real time for safety and alignment with user intent.
- Dots
- OpenAI product that gives users a cloud computer capable of proactively discovering and handling tasks.
- Atlas
- OpenAI's former AI browser product, now discontinued.
How to listen
Founders and investors focused on AI agent product design, enterprise automation, RPA, or legacy-system modernization.
The opening 2 minutes of pleasantries and guest introductions has low information density and can be skipped.