Opening thoughts. Agents are running amok. Or are they? All I’ll say is, particularly from a consumer angle, this is what you get for trusting your life with AI agents. A YouTuber has complained in a post on Threads that Meta’s Muse agent made a deal for an item that was to be sold, with another user on the Facebook Marketplace (trusting AI agents made by Meta, AND trusting Marketplace—something needs to be said about life’s choices…anyway), agreed to a lower-than-expected price and gave the owner’s (or configur-er’s) home address to the buyer, who of course showed up.
Did the AI do anything when the person showed up, without the seller actually being aware of this? Absolutely nothing. Later, a typically sycophantic AI response included “bad news” and “I should probably stop the auto replies”. ‘I’m sorry and I’ll delete myself from existence’, would have been a better response. My question is really simple (though I suspect the answer may not be)—are only AI agents hilariously dumb, or are the very people using AI agents for tasks they should be doing themselves, dumb-er still?
A few days ago, I sat down for a conversation with Articul8’s CEO Arun Subramaniyan, and unlike most tech leaders who tend to keep an eye on the clock (I absolutely don’t hold it against them), Arun was more than happy to keep that absorbing conversation going. Mind you, it was very late night for him in California, which he lovingly calls ‘late evening’. A few important themes emerged from the chat, as Arun detailed his thoughts on various facets of the AI era. I wrote about the Heritage AI aspect, as the AI company builds Laya and Sol models with expertise in Sanskrit and Tamil(Do give that a read). But here’s a broader perspective.
“First and foremost, my personal view is that unfortunately the word I would use is fear-mongering. That’s going on especially from the US labs. And if you notice, it’s only the US labs which are saying what they’re saying.” Subramaniyan’s clear assessment of AI companies including Anthropic and OpenAI seemingly sounding the alarm on apparent AI activity that wasn’t expected in their tests.
“Of all the cases that have been documented as a model breaking out and doing something, there is no case where a model explicitly disobeyed what a human asked it to do. That didn’t happen. What happened was they put guardrails. If you put a human in there, you can say that I didn’t specifically tell you not to do that, you should have known enough to do that. That’s not something that you can expect of these models, not yet at least, that you need to put the right harnesses around it.”
“Just because you can get a high school student to give you an answer about brain surgery, it doesn’t mean they have become a brain surgeon.” This is in reference to the ‘expertise deficit’ that is fast emerging in the AI era, where an illusion of competence is shrouding actual hard work. When non-experts use these tools without a deep understanding (because well, an AI chatbot furnished research they otherwise may have needed to do), they often push through low-quality work or errors that they cannot explain or verify if need be. That’s particularly dangerous for organisations.
“The capabilities are so close that you are not going to get a differentiation. That differentiation has to come with, the practices of your enterprise.” Subramaniyan argues that general-purpose models have reached a level of “sameness” where they no longer offer a competitive advantage. True differentiation for enterprises and cultures lies in training models on their own specific data, rules, and a “spirit” that defines every workplace.
“In the last six months at least, Every enterprise’s total token cost has skyrocketed. Unfortunately, their overall productivity has stayed flat or the trajectory has been downward.” Many enterprises are failing to budget for the actual costs of AI. Subramaniyan insists that moving from small-scale pilots to enterprise-scale production requires rigorous engineering, repeatability, and an ability to audit every step of a process—boring but critical requirements that many systems currently lack, he points out.
There is an interesting piece of data that Google and Kantar shared a few days ago. It is a study on Indian car buyers relying on AI experiences when it is time to decide on their next motor vehicle purchase. The study titled “Automobile Purchase Journey in the Age of AI” suggests that 93% of Indian car buyers are now actively using AI tools for research—and if you think this number looks on the higher side, I mist add they’re adding Google Search AI Overviews (90% users) and AI Mode (89%) within that bracket.
Google and Kantar suggest that 8 in 10 car buyers have an initial brand shortlisted, even before active research starts. And around 6 in 10 buyers begin by casually exploring videos on YouTube to get a better idea of the automobile market. “The car buying journey in India is deliberate, non-linear, and deeply digital—particularly ahead of the festive season, when consideration peaks. Buyers move between discovery on YouTube and conversational validation on Search,” says Siddharth Shekhar, Director, Google India.
In my book, a rather worrying metric at a societal level is that among the typical modern car buyer, 63% watch YouTube daily, while a lesser number (54%) read newspapers daily—streaming more than reading, whatever you make of this metric. But the good thing is, at least 41% of new car purchases place safety features as one of the key deciding factors. After all, a new car is a typically collaborative decision for Indian families.
For brands, this means there should be importance given to their digital footprint. They cite the example of Hyundai in India, which took advantage of YouTube and AI-powered Search ad solutions, to clock a 30% growth in retail sales of the Hyundai Verna, with around 32% lower cost per enquiry.
Neural Dispatch is your weekly guide to the rapidly evolving landscape of artificial intelligence. Each edition delivers curated insights on breakthrough technologies, practical applications, and strategic implications shaping our digital future.
