AI agents? Yes, let's automate all sorts of things that don't actually need it

Opinion The “agentic era,” as Nvidia’s Jim Fan and others have referred to the current evolutionary state of generative artificial intelligence (AI), is going to be a huge disappointment.

The simple reason is that most people do not need AI agents to book their trips, order groceries online, or book restaurant reservations.

Those who do need AI agents – friendless titans of industry, exposure-starved celebrities, and the like – already have a better option, a human personal assistant, several of them, even entire companies committed to carrying out commands.

The tech-besotted imagine that everyone wants their own Jarvis, the AI agent that serves billionaire Tony Stark in the Marvel cinematic universe, even though Jarvis would be wasted making occasional DoorDash orders for those with relatively uncomplicated lives.

But beyond the delusions of self-importance that lead people to believe they deserve a software servant, there are a number of obstacles that will hinder the deployment of AI agents and limit their practicality.

OpenAI’s Operator, like Anthropic’s computer use API before it, combines generative AI interaction – the now familiar prompt and response ritual – with multimodal models (capable of visual, audio, and text interaction), chain-of-thought reasoning, browser automation, and access to third-party APIs.

That’s quite a technical confection, but what it enables – web automation – isn’t new. Everything that Operator enables was possible before using other means, both programmatic and manual.

The demonstration video shows a few such tasks as prompts to Operator, each of which is tied to a specific online service provider – “I need a cheap rental car for LA,” linked to Priceline; “Find me a spot to take my 22’ campervan for a solo trip this weekend that has electric hookups,” linked to Hipcamp; and “I’m looking for a boutique hotel in Paris with a view of the Eiffel Tower,” linked to Booking.com.

Has OpenAI not heard of Google Search? Of course they have, and that’s what this is really about – opening up a new distribution channel for online services in which OpenAI is the gatekeeper, toll taker, and data aggregator instead of Google.

OpenAI is the gatekeeper, toll taker, and data aggregator instead of Google

It’s understandable why OpenAI might wish to become the central intermediary between customers and businesses, but it’s less clear why businesses might want to become endpoints in OpenAI’s ecosystem and surrender direct contact with customers.

Perhaps companies are keen to participate to exclude rivals – if Uber is getting rides booked by Operator, Lyft is not.

But if Lyft does eventually become a partner in this ecosystem, will Operator make decisions about which brand will be awarded a given service request? Or will Operator ask the customer to make the decision – an intervention that suggests a human might have better handled the booking process without an intermediary service? Along similar lines, why couldn’t companies like Booking.com integrate their own open source AI models into their IT infrastructure to act as brand-specific agents that focus on site-specific tasks?

What about other potential costs beyond brand dilution? Is this a pay-to-play arrangement and, if not, could it become one? Are there infrastructure costs associated with fielding the fumblings of hundreds or thousands or millions of Operator instances hitting servers and interacting? OpenAI’s plan to spend something like $500 billion on AI infrastructure through its Stargate Project says something about the super-lab’s ambitions. Whether there’s demand for the AI services expected to run in those data centers remains to be seen.

So many questions

The bandwidth burden imposed by AI crawlers is already a problem. According to DoubleVerify, “a record 16 percent of [general invalid traffic] from known-bot impressions in 2024 were generated by those that are associated with AI scrapers, such as GPTBot, ClaudeBot and AppleBot.” Will AI agents make this worse?

Operator is presently available as a research preview to US Pro tier subscribers, who pay $200 per month. And presumably, the agent service is going to burn through a lot of tokens, compute cycles, water and energy to complete complicated tasks. Are those resources well spent?

Relying on AI when it’s not necessary is ethically indefensible

Given that the human brain can perform mathematical operations using a million times less power than a computer, relying on AI when it’s not necessary is ethically indefensible in light of the environmental impact of data center usage.

And that’s to say nothing of the moral bankruptcy involved in using copyrighted content without compensation to train AI models, in order to sell content generation services.

Then there’s the time required for Operator to complete all the steps necessary to accomplish the requested task – sometimes more than a few minutes, if the agent doesn’t stop entirely. If the task is simple and easily done, it’s probably a chore that a person could have done in a browser just as well in more or less the same amount of time. And if it’s complicated, requiring human intervention multiple times, maybe the human should have simply handled the whole process from the start.

Let’s not forget the policy questions. Many websites disallow automated interaction, often with good reason (to prevent spam or malicious bots). Is that suddenly going to change because Operator comes calling? Keep in mind its maker’s warning, “We know bad actors may try to misuse this technology.”

According to AI-focused publication Every, Operator uses a browser in an OpenAI datacenter rather than the user’s desktop browser (where varied IP addresses and user-agents would hinder blocking). Setting aside the privacy implications of that design decision, one of the consequences is that sites that block bots like Reddit, Figma, and YouTube are already turning away Operator.

To demonstrate the capabilities of Operator, Andrej Karpathy, who helped co-found OpenAI before departing, had the AI agent post a comment to a Hacker News thread discussing the Operator announcement that summarized the sentiment expressed in the thread.

The AI-generated comment is banal and non-committal, in keeping with the sort of beige prose that people have come to expect from generative AI. And it was suitably flamed by a forum participant: “I’m sure you think this is cute or inevitable, but it’s also how you destroy community on the internet and finish its transition from an implicitly trusting public square into an implicitly adversarial and polluted wasteland.”

Not only are AI agents not needed for the tasks proposed, no person wants to deal with them

Though Karpathy subsequently posted some details of the Operator run, describing various failures and delays along the way, he also underscored his critic’s point – that AI-driven interaction drives people away – by responding to the criticism with an Operator-generated reply.

So not only are AI agents not really needed for the tasks proposed, no person wants to deal with them. AI simply isn’t welcome in human spaces. At least Operator will be well-received for certain automated API interactions, assuming prior negotiation has occurred.

But as Steven Sinofsky, former president of Windows at Microsoft, has observed, automation is more difficult than most imagine.

“Most people who have built an automation know, or at least come to know, just how fragile it is. It is fragile because steps are not followed,” he wrote in an essay last October. “Tools and connections fail in unexpected ways. Or most critically because the inputs are not nearly as precise or complete as they need to be. And they come to know that addressing any of those is wildly complex.”

Sinofsky professes to be optimistic about AI’s role in automation over time, but expects it won’t be easy and that humans will have to remain in the loop.

So Operator looks less like Jarvis and more like a toddler, whose hand has to be held, at least for a while. If one goal of Operator is replacing human assistants, enjoy your child-like employees.

But at least AI agents are not entirely useless; consider the electric scooters that a few years ago littered city streets in a show of venture capital hubris and contempt for public space.

Electric scooters are useful for certain scenarios – trips that require more than some moderate period of time (20, 30, or 40 minutes, depending on various factors) to walk, for those in an age group not particularly concerned about injury, during periods of pleasant weather, when no significant load has to be carried, over distances that aren’t too great.

Operator is like that. Someone is sure to appreciate it. ®