How Autonomous AI Agents Are Ending Traditional Work | Fractal CEO

It's 10x the industrial revolution at 10x the speed. Human intelligence is no longer relevant.

It's going to be in the human realm, assisted with AI. AI will replace that human time. It's going to happen.

So the movie is going to get very, very interesting.

Wonderful to be here, Deepak.

No doubt about that. Absolutely. Yes. Thank you, Deepak. Wonderful to be here. AI is one of those transformational technologies — once in a lifetime, not even once in a generation, once every 300-400 years, right? It's that level. It's as some people have called it: it's 10x the industrial revolution at 10x the speed.

So it is going to change our lives in a very, very significant way and we are still in the very initial stages of that. This is the first chapter of this book of AI that we are in. So the movie is going to get very, very interesting — the book is going to get very, very interesting from here on. And the reason it is, is because historically what we have done is we have — every technology has changed something or replaced something.

The industrial revolution made physical muscle power obsolete because machines could do a much better job than human beings. Then came the information and communication technology revolution, which essentially made distance not a relevant thing anymore. You could travel anywhere through various mechanisms, you could do virtual calls, you could make phone calls there — things that basically brought us together, created globalization, made globalization possible.

This is about generating intelligence — from electricity all the way to intelligence. Can you take electricity and convert that to intelligence? And therefore now human intelligence is no longer relevant. Right? That is a hugely different thing because for the last 100-odd years, or maybe even longer, it's human intelligence that we've prided ourselves on, and if that becomes a commodity or that becomes easy, what are we going to do next? So it's going to change the world in a very significant way and therefore we have to pay attention and leverage this opportunity to build a better world.

So you said regression on a spreadsheet — that was not the case, in fact quite the opposite. It was all about human decision making, right? Literally the thinking was: 'in my experience, my judgment, my intuition is really the most important thing. I don't have to look at the data. Don't tell me how it is done.'

I remember the early days of Fractal, I had met one of the retailers in India — a very good retailer. The CEO was very forward-looking, visionary, and he wanted to bring analytics inside the organization. He brought me into that conversation and his entire top management was there. We had about two hours, and literally I was getting attacked every second by every one of his top management except him — he believed, but every one of his top management said: 'You're going to tell me that I can run my retail business better because of data? You think I don't understand the trends, I don't understand the customer, I don't understand the stores? You're never going to help me with this. No AI or no analytics is going to be useful for me.'

And after the two hours, I was completely defeated. I knew that I was right, but I knew there was no way I could convince them. So the world was very different 20 years ago — or 25 years ago when Fractal started — there was no belief in regression models or even the power of data to drive decisions.

That is what we tried to build inside companies. We said: you can predict — you know, you have experience, that experience is getting captured in data, and you can use that data to build mathematical models that predict what's likely to happen, and you can power your decisions using that. So that's really how we started — we were trying to bring in the science of data-driven decision making inside companies. And India was a very difficult market with no belief in spreadsheets or regression models either.

Yes.

See, BI is business intelligence. It's about using information of the business to make some decisions. Now, traditionally —

Yes. Traditionally, what used to happen was companies would have these BI tools internally — there were popular ones called Cognos and BusinessObjects, and then came companies like Domo and Tableau, and then came PowerBI. These are various kinds of companies that would use data and display the data in intelligent dashboards that human beings would look at and make decisions on.

Right? So you would say: 'Oh, what happened in my business yesterday? A competitor raised their prices, or we launched this promotion — and let's see how it is tracking, and therefore let's see whether I should expand the promotion to other cities or other stores, or should I deepen the promotion. What do I need to do to win market share?' You're making a decision like that, and you're having the information about what happened in the business yesterday to make a decision for tomorrow. That was the world of BI.

Now, this was — even when Fractal started, Fractal was well beyond BI. We were using AI, we were using analytical models to predict what's going to happen next, not just describe what happened yesterday. We were using models to — okay, not just look back, but look ahead.

Look ahead and use that look-ahead power to make decisions. So that's what Fractal was doing. But let's say the 2017 post that I was talking about was about saying that even when you do BI using these very tools — primitive tools — now they will look like primitive tools, like Tableau, PowerBI — they will look primitive in a 2026 context. But in 2017 I was telling them: 'Look, these tools are now ancient. You need to bring in the power of natural language.'

Ask questions in natural language, get answers. BI has to be more proactive, telling you what's going to happen next — preemptive and proactive. It's telling you: 'Hey, this is something interesting. You must pay attention here. Don't look at the full spreadsheet, but look at this one little dot here, which is telling you what's going to happen next.' So we were telling the world that you need AI to make BI more useful.

Yes. One is you're using unstructured and structured data. Structured data is what is in rows and columns that can fit a spreadsheet, and unstructured data is what's in documents and videos, voice, text, etc., which may not fit into rows and columns but can also be very helpful. So bringing them together was one thing; probing deeper, like you said — understanding the why.

Something happened, but why did it happen? That was very interesting. And then: what is going to happen next, and what do I need to do next? So this is — if you think about it — the holy grail of business: being able to make the right decisions for the company, not just by looking in the past but figuring out what's going to happen in future. When you get that right and that engine is flowing well, you build a much more successful company regardless of the industry that you're operating in.

Yeah. No — let's look at the realm of decision-making, because it's a very vast realm, right? What is an enterprise but a set of decisions that are happening every second, every minute of the day. Now, if you look at the span of all decisions, I would characterize them as: some of them are highly strategic, one-off decisions, high-consequence — as Jeff Bezos often calls them, one-way doors. Right? You go through the door, you can't come back. Let's say you bought a company and paid a billion dollars for it — if you realize it's a bad decision, you can't unwind it later on, it becomes too challenging.

So these are strategic decisions where, while AI is helpful, it's still pretty much in the human realm. In these decisions, you don't expect AI to do more than give you analysis, data points, what-if scenarios, simulations — but let the humans, along with some AI help, make the right decisions. Right? That's the way the world is going to move: it's going to be in the human realm, assisted with AI.

The second kind of decisions are what I would call highly tactical decisions. Every day you do make those decisions. They're not very high volume — each decision itself is not very high impact — but you do make those decisions, like employee performance appraisals, like you're giving feedback. Now, if I'm giving feedback to someone who reports into me, I want to analyze all of the work that she has done over the year and understand what's the feedback coming from other parts of the organization — her peers, her juniors, etc. — collate everything.

And then give her a really actionable piece of advice on what she should do next. I'm going to use a lot of AI to do that right now, because regardless — I have one-on-one interactions with that reporter of mine, but she is interacting with an organization, she's doing a bunch of things, I don't have all that context. But if I use AI, I can get much richer and give her much more precise advice and feedback, coupling with my own judgment about her.

Right, so that is now in the realm of tactical decision making, where I would use a lot of AI, but eventually AI will assist me in actually making those decisions. Right? It's more AI and much more valuable AI, but still I would say that it's humans with, along with AI help. Right? That is the second realm.

The third realm is where it gets really interesting — these are operational decisions, high volume, high velocity. Individually they may have less consequence, but overall they're very high consequence because there are many, many decisions — like every time a credit card transaction pings, you need to decide whether it's a fraudulent transaction or not; every time a customer is applying for a loan, I have to decide whether she's good for credit or not; a claim comes in, I have to decide whether to approve the claim or to reject the claim or to do something else with that claim. Right? These are decisions where human beings do not have the attention span.

Let's say I'm watching video feeds of, like, 20 screens that are telling me what unauthorized activity is happening in the office. Right? So let's say some security guard is doing that today — they have a huge room, a war room, with all kinds of screens, and it's 99% of the time nothing interesting is happening. But that 1% of the time when an intruder comes in or something is going on, they have to pay attention — human beings are not very good at that. So when it comes to highly operational decision making, it should be mostly done by AI with some human oversight. Really, that is where all of the action should move.

Yes. So the more you trust that your AI is trustworthy, the less human oversight is required. Right? So the first question really is: the AI that you've built — how accurate is it, how aligned is it to your purpose? Like, you wanted a certain thing; sometimes you have a purpose but the AI has taken that purpose and done something in the process of doing so, and you've completely distorted what you wanted to do. That lack of alignment is very common in AI systems. It's called the alignment problem.

It is very well researched but not well solved. So if the AI is not very strongly aligned or it's not accurate, you want a lot of human oversight. And as time goes by and you feel like it's much more aligned and it's much more accurate, human oversight can go away completely.

For example, take autonomous driving, right? Today, Waymos are way safer than any human-driven cars because, as human beings, we like to do hair and makeup or talk to people or listen to podcasts and do all of the above while still driving the car. So we tend to not pay attention, whereas an AI can pay full attention. And as I get more confident that this AI is never going to kill a human or never going to put me in danger and is going to navigate me in the safest possible way at the fastest possible way, then I will step away and say: let AI do the job. Why do I need to do this?

True.

Let's look at why we are where we are. Right? It's not that we have healthcare — for example, India doesn't have enough doctors, and we have a lot of unlicensed quacks operating as doctors if you go to remote villages and so on. So universal access to high-quality healthcare is an important challenge that India needs to solve.

So how would you solve that? You will solve that by making sure that there is some AI layer accessible to everyone for every need — rich or poor. By the way, it is not across economic classes, because even if I'm rich—

In between the doctor visits — I had just had a doctor visit now — all my interesting questions have come up in my head only after visiting the doctor, right? I forgot to ask them. Now I want to use the AI to query: 'Okay, this is what the conversation was, this is the report, this is the diagnostics from the lab — now tell me what are the questions, or what are the answers to some of the vexing questions in my head before I meet that doctor next.'

Or: 'I have something going on in my mind, I don't know whether it is enough for me to go to a doctor, but I need to query with an AI tool to understand what's going on so that I'm prepared better for the doctor visit, and even decide whether to meet the doctor at all or not, because maybe it's nothing.' So these are the kinds of questions for which I might need AI access rather than having a doctor friend who I can call and pester every time. Right? That's not how it will work.

So really, for the complex stuff where human judgment is required, where errors are possible, you want to make sure that there is a doctor meeting you and solving for it. But for absolutely routine things, or figuring out what is worth paying attention to, you might use AI. So this is really how the world will move: for questions like, for example, 'I need my account balance' — I'm calling a customer-care number to figure out my account balance — that call should be avoided. You should have a bot answering that question, because I really do not need a human to tell me this. The human is telling me because the technology was not ready. Now I really want the answer to my question as quickly as possible.

Waiting for an agent to answer the question is inefficient, is not a good use of my time. So we have to figure out what are those things that are absolutely self-served, should be avoided in the first place, and what are those where a real human is needed to help me out, because human time is precious, human time is valuable and expensive, and I want that human to also — those jobs have to also be meaningful. So it has to be only in the context of what is valuable, where the human interaction is really valuable, and take out and supplement with AI or replace with AI where human interaction is not relevant.

Yeah, see, we have to pay for tokens, and sometimes the cost of a token has to be compared with the value that the token is producing. Earlier today, Jensen Huang did his Nvidia quarterly results and he talked about how tokens are becoming profitable. So his point is that every token is actually generating revenue or profits to the end use — right, it is actually really valuable. But the real question is: the price of the token, or the cost of the token, versus the value that the token generates. For some things, if you're using a lot of tokens because either you are very inefficient in the way you ask the question and you asked it to boil the ocean, and therefore it consumed a lot of tokens — and therefore the value of the task accomplished versus the cost of the task, what you paid for, just did not make any sense — that is possible. And there are other situations where, let's say, the task is much better done by a human being, because human beings are far more effective at doing that task than an AI can do. So do not use AI where the tokens simply do not generate enough value; use it where it generates enough value. That's one.

Now, secondly, remember that the cost of these tokens is going to go down. These models are going to be much more efficient in using electricity — they're being optimized so that they use fewer tokens — and these models are getting more intelligent because of all the combined effects. Token costs for a like-for-like model are going down 10x every year: the same model that you used last year today is only one-tenth the price. It's only because you want the latest and greatest model, which is even better, that you're still paying that same price. So in this way you will see that the apples-to-apples comparison year-on-year will make tokens absolutely so cheap that you can use it for everything.

In fact, we call it too cheap to meter. We will also enter the era where AI is too cheap to meter, and that's when you will use tokens indiscriminately. That's not happening today because Nvidia's chips are very scarce — they're making 75% gross margin. So right now AI is very expensive relative to where it will be eventually, and so we'll have to plan for the world where AI is too cheap to meter.

There are two aspects of it. Let's separate the two. Number one aspect of it is that, to our last point, we were discussing the cost of a token versus the value it generates, and also relative to the human cost of the same work. If the cost of AI to do a piece of work is significantly cheaper and more accurate than the human cost of doing that, you will see that AI will replace that human time. It's going to happen.

It's going to go down. So it's going to become more and more attractive, and the cost of human time is going to go up as we become more and more developed as a civilization. Cost of human time goes up; cost of AI goes down. So that economics will dictate that more and more work will flow to AI. And AI is not just replacing technology cost; it is also replacing labor cost, and that's what you're seeing — the Meta example, etc., are some examples of that.

Now, there's a different question as to whether it's a good thing or a bad thing, but this is a direction of travel. And I sometimes look at the sort of scene in the Matrix where Agent Smith comes and says: 'Never send a human to do a machine's job.' Right? And you were saying something very similar. This is going to happen: a machine is far more effective, and why do I need a human being to do this piece of work? Like, for example, driving — why do you need a driver?

If AI can do a much better job of driving and actually save a million lives every year, that it will be irresponsible to have drivers drive the car just because we want to provide employment. It won't make sense at a country level or a global civilization level that we are letting 1 million people die so that a few million people can get employed. That won't make sense.

So there will be various kinds of reasons why we will move into an AI-first world where AI will replace human time. Now, the next question is: what does the world look like in that scenario, and how do you protect jobs? How do you protect work? And to me, that question is fundamentally not connected to this, because this disruption is going to happen — the wave of technology change is inevitable, unless a lot of countries come together and block it and stop it. It might happen, but I'm not betting on that. It's probably not going to happen.

Now, in that world where AI is getting adopted, what we have to see is: how does human work continue, right? We are not here to protect jobs as much as protect human beings — we want to protect human beings. So one question of that is: what happens when some work gets automated? Does something else pop up as a new option? Historically, for the last hundreds of years, we have seen that for every technological revolution — it has replaced some jobs, but it has created way more jobs than it has replaced.

So what are those new kinds of roles that come in? For that, you have to look at what is the realm of human imagination. We are a species that is capable of unlimited imagination. We come up with new ideas all the time, and something gets automated, we come up with the next problem. For example, we are now thinking of orbital data centers, we're thinking of colonizing Mars or making the world a multiplanetary species.

So these kinds of ambitions — providing healthcare for all, right? It's still a distant dream; we don't have healthcare for all. So many, many problems to be solved where human time can go in solving, because something has become automated. It's a good thing if something becomes automated or cheap because machines are doing it.

Exactly.

So FDE is a very hot concept today, but it's pretty simple to understand.

This term was coined by Palantir back in the day, a few years ago, and the logic was the following: I have all these amazing engineers who are building the Palantir product, but I have to solve customer problems or client problems. Can I put some of these backend engineers and forward deploy them so they're actually at the customer location, solving that, making that product work for that company in real time? That was the idea of forward deployed, as opposed to backward. Most engineers are always backward deployed, building their technology. Can I forward deploy them so that they're working closely with the customers and solving their problem?

In some ways. Now, what is some of the differences now, especially in an AI-first world? Number one is that these people have to be able to understand the client problem and whiteboard with them and be able to fully appreciate the domain depth, the context depth of the organization. So that's number one. Okay. Sometimes engineers are not necessarily that, right, because they're very busy working on whatever the technological problem they're solving. They have to solve a business problem. So they need some kind of management, consulting, data science kind of a background that they need to have. Number one. Number two is that they should be able to communicate very, very well, and that is important. Number three, they're not exactly management consultants because they need engineering depth because they're understanding a technical problem, they're solving that too. And number four, they need to have data understanding, analytics understanding, and algorithms understanding because they have to solve an algorithmic AI problem inside the company. So if you bring all of this together, you have a forward deployed engineer. That sounds like a unicorn today, right? Because having all these skills in one person is not easy. But if you have a mix of these skills and you are able to make the technology work for that client, you are a forward deployed engineer, and the world needs many of them because technology is getting much, much better, but realizing value from the technology is still a challenge. You've built amazing models, let's say as Anthropic or as OpenAI, you've built these phenomenal models that convert electricity into intelligence, but that intelligence has to translate into revenue, and that's really where the forward deployed engineer comes in. And she or he works with the customer or client and understands this specific problem and makes these models come alive in that environment.

It's a great question. At least as of right now, no. But in future, it is absolutely possible that there's an AI agent working with you to understand your requirements or your needs fully, brings in enough domain understanding, and makes the technology work for them. So it's an agent deploying a tool. It's possible conceptually, not today as practically.

I think first is understand and get deeper into domains, right? Understanding any domain is a very good asset today. So if you want to be a healthcare expert or you want to be a financial service expert, having some understanding of the context, going deeper is very important. Second, you need to understand the mathematics behind machine learning or AI models, which means that linear algebra, probability, and statistics, deeper understanding of these math models is very helpful. Strong communication, obviously a must, and then deeper engineering understanding so that you can solve these problems because these are very scaled problems. So if you bring these skills or at least a mix of these skills, you have a very good shot at working for AI companies.

So let me explain the context behind what I said. And generally speaking, if I said something that I revise my opinion, I think I pride myself on being able to revise my opinion.

But in this case, I stand by what I said, and let me explain what I mean by that. We are thinking of AGI. The world is changing very quickly. AI models are progressing very fast. At a country level, I need to think: what if this technology really becomes so good that a nation state controlling the technology could effectively be superior to me and control me? That is the danger around sovereignty because of AI, right? So that is a real conversation of AI progress and so on. But on a side note, we were also discussing how AI models do not understand Indian data very well, don't understand Indian context very well, language very well, dialects very well. And therefore, let's build AI that works for India by building those. Now this is also an important problem, but in the scheme of things, it's a much smaller problem than building towards AGI. So it can become a distraction when you're thinking of language and when AI is becoming so much more powerful. The second aspect of this is that if you look at the models, Western models, they may have been poor at understanding Indian languages and dialects when they were initially released. But you have seen in the last two, three years, as the general capability of these models increases dramatically, they become very good at language automatically and dialects and everything as well. So if you're solving a smaller problem here, the larger problem is already solving the smaller problem. So please don't get distracted and try to solve the Indian language problem when the generic models are going to solve them anyway. So solve that problem and compete at the AGI level rather than compete at this level. That was the idea.

Look, AI learns from data, and if you bring in more data, more context, more situations. For example, India is extremely rich data, right? India has 1.4 billion population who are all digitally savvy, using, consuming more bandwidth per person than any other country in the world, right? And we have this UPI stack and the Aadhaar stack creating the perfect conditions for a data-rich economy which is model-poor right now. So absolutely bring in all of the Indian data in order to build India's sovereign AI. Absolutely do it. But that doesn't mean that we just focus on building a language model that only does the language task, because I would like India to build models that compete globally and match and exceed the Chinese and the American models out there.