AI Will Be Decided in the Real World, Not in the Lab
There is something increasingly odd about the way the AI industry talks about Artificial General Intelligence, Superintelligence, the Singularity and the risks associated with them. The language is becoming increasingly grandiose: AI will transform the world, AI may surpass human intelligence, AI may become uncontrollable, AI may eventually pose an existential threat to humanity.
Yet when I look at these conversations from the perspective of someone who works with businesses and is actually trying to use AI to improve productivity, I find a fundamental disconnect. The AI industry, particularly the technologists within it, seems to have very little understanding of how the real world works — not just of the risks of AI, but even of its benefits. And that disconnect matters because businesses are ultimately the ones that have to adopt, pay for, manage and live with this technology.
Redefining "Intelligence"
Let us start with the word "intelligence" itself. The AI industry has increasingly adopted an operational definition of AGI along the lines of a machine being capable of performing a sufficiently large range of intellectual tasks at or above human level. But task performance is not intelligence. You cannot simply redefine a word because it makes the technology easier to measure. Intelligence, in its actual English-language meaning, is the ability to learn, understand, reason and make judgements based on reason. Human intelligence is not simply a collection of tasks that can be benchmarked. It involves insight, intuition, emotion, experience, maturity, judgement, understanding and consciousness.
We have spent decades studying neurons, synapses, neurotransmitters, the sodium-potassium pathway, dopamine, brain waves and an extraordinary amount of neural activity. Yet we still do not have anything remotely resembling a complete understanding of consciousness, or of how all these elements combine to produce human intelligence. So when the AI industry takes increasingly sophisticated task performance and labels it "general intelligence", there is a very large conceptual leap taking place. Perhaps AI will eventually produce something genuinely comparable to human intelligence. But we should at least be honest about how much we still do not understand about the thing we are supposedly trying to recreate.
There Is No Going Back
None of this is an argument against AI. Quite the opposite. AI is extraordinarily useful, and there is no going back. The Pandora's box has been opened. Ideally, perhaps, we would never have developed some technologies at all. But once a technology exists globally, refusing to develop the capability to understand and use it does not protect us. Someone else will use it, and we will still be affected.
This is particularly true for countries such as India. If the United States, China and other countries continue developing AI while India decides to sit out, India does not become safer; it becomes dependent. We therefore have to develop the skills, infrastructure, capabilities and governance required to use AI properly. But that does not mean blindly following the AI industry into every technological race it proposes.
The Geopolitics Nobody Talks About
This is where geopolitics enters the picture. One cannot look at American AI development in isolation from American strategic interests. The standard US playbook has historically been to develop technology, establish leadership, restrict or control access where strategically useful, and use that technological advantage to further its own interests. The United States is not exactly a role model when it comes to separating technology from national interest.
Therefore, when American AI leaders talk about slowing down AI research or superintelligence, governments and other countries will naturally ask what lies behind those statements. There may be genuine concerns about safety. There may be guardrails being developed. There may be efforts to train people in proper supervision. But there are also geopolitical and economic considerations.
Nothing on Earth operates in isolation. That is why we should be extremely cautious when reading such grandiose statement, and make a genuine effort to ground ourselves in business realities.
The Risk That Actually Matters: Uncommanded Execution
I do believe AI has real risks, and some of them are already visible. The risk that concerns me most is not a Terminator scenario. It is uncommanded performance and execution. I have personally seen enough to pause the engine and go back to manual operation for now. If an Agent does something that its human operator did not intend, that has a direct business impact. The industry has to provide answers to this. No excuses. If an Agent can communicate with customers, execute transactions, make decisions or trigger workflows without direct human approval, then the question of what happens when it gets something wrong is not theoretical.
This isn't a new concern for me. I wrote about it at length in AI Going Rogue — How Big Of A Risk Is It, Really?, where I tried to actually quantify enterprise AI agent risk rather than talk about it in the abstract — drawing on an earlier incident involving an AI dialer's uncommanded execution, the OpenAI agent incident from July 2026, and a smaller version of this same failure mode that I ran into myself. The three scenarios below are what that risk looks like once it moves out of a single automated call and into GTM, lead generation and healthcare workflows more broadly.
The customer service scenario. Consider a simple case. I hand over GTM and Customer Service to an Agent. It handles a call from an anchor customer who contributes 20% of my business and regularly gives me new business. The Agent says something or executes something that causes the customer to disconnect and take the business elsewhere. Who is responsible? There is no easy answer. Even with real-time human supervision, perhaps a quick apology, discount or remuneration might sort the problem out. Then again, it might not. The customer may simply decide to take the business elsewhere. And in a real-world system running at scale, real-time supervision of every transaction is well-nigh impossible. The cost would be exorbitant and execution would slow to a crawl.
The lead-generation problem. The same issue appears in lead generation. Suppose I receive 25 leads a day. An AI Agent talks to all 25 and gives me a summary, including which ones it believes are likely to convert. What happens if the summary is not good enough to identify the customer who actually would have converted? The leads that the AI rejected may be the ones I would have converted. Is that a problem with the technology? With the training? With the workflow? With the way we have defined a "good" lead? These are the real-world AI problems that businesses are going to encounter.
The hospital workflow. Consider a hospital where an AI-powered system automates reception, registration and patient classification. Two people with the same name arrive. The system treats them as the same entity and one person's information gets classified under the other. There may be a million John Smiths in the world. What happens then? That is an AI risk with immediate real-world consequences. It doesn't require superintelligence, consciousness or a machine takeover. It requires one error in one workflow.
Why the Industry and Business Speak Past Each Other
This is why I believe the AI industry has a serious context problem. Technologists are often micro-focused on one domain, and understandably so — their job is to build the technology. Business people, by the nature of their jobs, are forced to look at a much wider ecosystem: geopolitics, economics, business, sales, marketing, HR, customers, competition, people relationships, regulation and finance. When someone tells us that AI will "transform the world", we may understand the words, but we don't necessarily understand or empathise with what is being said because there is no context and no anchor with which our brains can engage. What does transformation actually mean? What changes in my business? What changes on Monday morning?
The same applies to doomsday scenarios. We understand what the person is trying to say. We understand the risk they are highlighting. But if we cannot conceptualise the scenario, we cannot engage with it. If someone tells me that an AI incident happened at Hugging Face or OpenAI, I can understand why it matters to the people involved. But I may still be sitting 15,000 kilometres away asking: how does this affect my business? What happens if I lose access to one AI tool? I was running my business before these tools existed, and there are already five other AI tools available. Why should I panic?
This is where the trust deficit begins. The industry talks about transformation while most businesses experience productivity improvement. My life may become easier because of AI, but I still have to go to the office, follow up with critical customers, manage people and perform my daily tasks. The nature of my tasks may change, but that is not necessarily transformation in the grand sense in which the industry uses the word.
Bringing AI Back to the Real World
We should absolutely continue AI research and development. India must develop AI capabilities because if we don't, someone else will, and we will still be affected. But responsible growth is essential. We need proper, legal, moral and good usage of AI under established processes and good law. And, critically, we need to keep business interests front and centre. Only then will investment happen.
If we expect massive investments in foundational technology without having the economy, market and ecosystem to support those investments, we are fooling ourselves. The question should not simply be, "How do we build AI because everyone else is building AI?" The question should be, "Where are the profits and cash flows, how can AI improve those businesses, and how can we improve the work and lives of the people involved?"
Naa baap badaa naa bhaiyyaa, bhaiyyaa sabse badaa rupaiyyaa. Business interests ultimately drive sustainable investment. We need to look at where AI can create genuine economic value, build capability around that value, develop the right safeguards, and gradually increase the level of autonomy as the technology proves itself.
The Bottom Line
AI is here. There is no point pretending otherwise, and there is no point being paralysed by hypothetical futures. We need to use it. But we also need to understand its limitations, its failure modes and its real-world consequences. The time to train software vendors, development teams, software architects, programmers, and front-end, sales and marketing teams on AI risk in depth is now — not when customers begin asking these questions bluntly.
The future of AI will not ultimately be decided by the grandest predictions made inside technology companies. It will be decided by whether ordinary businesses can trust the technology enough to put it into the real world.
Comments
Post a Comment