Very few people left in the room, so let's—I'll address the ones who are here. Okay, how many of you in this room are deeply interested in mathematics? Just one or two. Okay. All right. Something very interesting happened just earlier this month.
If you see a to the power n plus b to the power n equal to c to the power n, does it ring any bell? Anyone knows what I'm talking about? Does anyone know what this is?
Okay, so there was an amazing mathematician called Fermat who many, many hundreds of years ago, in the 1600s, he came up with this theorem that a to the n plus b to the n equals c to the n has no positive integer solutions beyond n equals 2. When n equals 2, it's a squared plus b squared equal to c squared, and you know them as Pythagorean triplets, right? 3 squared plus 4 squared equal to 5 squared. But when n goes beyond two, you don't have any positive integer solutions.
He came up with that theorem, and then what he did was he said that, oh, I have discovered a marvelous proof, and that unfortunately I don't have enough space in my book to write. So he said, I have a very fantastic proof, but my margins of my book are too small. I can't write it here, and that's it. After that, his notebook was found, but he was not there. The proof was lost, or maybe there was no proof. We don't know, because it's taken another 250 years for the proof to be found. Eventually, in 1995, a guy called Andrew Wiles, he wrote the proof. There are many, many people who tried this right from the 1700s and so on. This guy called Andrew Wiles, he came up with this proof in 1995. It was a 195-page proof. It was not an elegant proof that you could put in one side of a page, but it took 195 pages.
He did that, but this mathematical theorem was not yet verified in a way that you could be foolproof. So formalization of this proof was still a major task, which has taken us another 30 years. Only earlier this month, on September 4th, 2026, the AI company called Anthropic came up with a full verifiable proof. And why am I wasting your time talking about this? I'll come to that. I'll explain why this is so important even for what anybody does in retail or anywhere else in AI. Why is this so important? I'll get to that. Just give me another minute on that.
So they came up with this idea that it can be proved, and they used the Anthropic latest models, Fable 5.1 and so on. They took a computer, took roughly 12 hours to find this proof, and they realized that to prove something like Fermat's Last Theorem—it's called Fermat's Last Theorem—they need to prove a lot of intermediate results, including all the way going back to saying 1 plus 1 equals 2. To prove this, you need to prove this, to prove this, you need to prove this, and the sequence is very, very long to formally completely prove something.
And it turned out that AI was getting lost in trying to prove this because it proves this, but it doesn't realize that, okay, I have to prove that too, and in that whole confusion it got lost. So an Anthropic researcher came up with this concept, which you're seeing on this page here. It's called a directed acyclic graph. So if I'm trying to prove something here, I need to prove something this, and that needs this proof, and she built out a whole tree, and that is what it started to prove. And it took roughly 13 million lines of code and 31,000 overall results. So for Fermat's Last Theorem, they had to prove 31,000 other results to get to that point. It wrote 13 million lines of code and did it, right? This is how it did it, right? This is a kind of a visualization.
Now let's get to why is this important. In math, you have to prove everything with deterministic accuracy, right? You have to say it's proven means it's ultimately proved, computer-verified proof, and it's extremely complex. Now think about the business context that you are working with. Every business is complex, right? It is complex by design, and you need a very high level of accuracy. But what AI has showed in just the last one month is that for even extraordinarily complex tasks, long-horizon tasks, you can still get 100% accuracy by using the scaffolding that they built around that. So this is the kind of thing that you're seeing, that yes, you need to prove a very reliable outcome, but with the right scaffolding you can solve a very, very complex problem.
So far, if you have been interested in the world of AI but reasonably disappointed with it because it's not really solved anything for you, welcome to the party. There's many, many people in that camp who have not really, you know, they've been excited by AI but have been disappointed with the results. But things are getting better every month. And now we have reached a place where AI can reliably perform like a co-worker, accomplish extremely complex tasks at a much better accuracy than human beings, especially if you have the right scaffolding.
One way to look at that is this metric called METR. This is an organization called METR. They publish this, and again, it's a little complicated, so I'll try to spend a little bit of time to explain this. This is a task. The duration of a task is 1 second, 4 seconds, 15 seconds, up to 64 hours, right? The task a human being takes to complete something, like if you want to say, proof, you know, whatever, Pythagoras theorem, right? It might take you some time. Or if he says, 'Oh, find this invoice, understand the customers, understand whether this is to be paid or not, and verify, cross-check with the database of something else, with the purchase order, and release the payment, etc.' There's a complexity of that process, right? So that might take a human being half an hour to do or 1 hour to do. But there are other complicated tasks that might take you 2 hours or 4 hours or a full day, etc.
The question is, at what duration of a task can AI match and exceed human performance in a reliable way? And that number has been going up as new models are getting released. We've gone to almost 16 hours right now, which means that a task that a human being takes 16 hours of meticulous work, AI can complete it in half an hour to 1 hour, or maybe even less, and do it better than human beings. That's beginning to happen in a narrow range of tasks, but that range of tasks is expanding, which means that even as early as 2027 or a year from now, you will have AI completely beating human accuracy in almost every single task that you can think of. That is a rate of progress; it's mind-boggling.
And so therefore, you have to think of AI not necessarily as an experiment, but it's become a core part of your infrastructure. It's a bit like electricity, right? 100 years ago, you didn't have electricity. Nobody had electricity. And as electricity came in, there were bits and you used it in some places, right? The companies that reimagined their entire factories with electricity made a lot of progress compared to companies that thought that, okay, electricity, new thing, I don't need electricity. Now you know that nothing works without electricity today, and all the inventions of the next 50-odd years—of microwave ovens and toasters and refrigerators and radios and everything else—was all because of electricity. So AI is the new electricity, and it becomes a part of your core infrastructure.
Intelligence is something that you can take for granted in every process, product, and service, and that's beginning to happen today. Therefore, anybody and everyone has to pay attention to what's coming up next, and it has led to a place where I'm going to skip some of these. We have reached a point like this, right? So imagine a few years ago, AI was an interesting thing which could do a few things that human beings could do. Okay. And we went to a place which you can call as artificial narrow intelligence, where in some places it was extremely good. For example, AI reading a chest X-ray and finding out tuberculosis in it or lung cancer in it. It's a very narrow use case of AI. It used to, it will beat human accuracy. That happened a few years ago, let's say about 7, 8 years ago, 10 years ago roughly.
Right from there, we have reached this place which you can only call as artificial jagged intelligence. It's extremely good, it's extremely powerful, and beats human accuracy in a number of different things, but in some places it's still childishly wrong, right? This is where we are today. And as we, like, we just crossed that stage and reaching a place where AI beats human accuracy almost across everything, maybe a couple of things may be left out everywhere, and then we eventually reach AGI, which might happen as early as two years from now or a year from now, which is essentially artificial general intelligence, where AI beats human performance everything, everywhere, all at once, handsomely. That's the future state, right? Almost there. We're almost there today.
So what does that lead to organizations doing? What can you do with it? Mainly, just imagine—I have boiled the ocean on this topic, right? So I'm going to give you a distilled version of what. There's three things that come up with it. Number one is the process dimension. Think of every single process: procure-to-pay process, invoice-to-cash process, supply chain process, merchandising process. Think of your customer management process, sales process. Think of every single process in the organization of any retailer.
Every single process can be reimagined with AI. Your customer experiences can be reimagined with AI. Your supply chain can be reimagined, your merchandising can be reimagined with AI. Each of these things can be reimagined. Not just automated—automate something that's already kind of dumb—but completely built with AI at the center of it, and imagining that AI is only going to get better. The worst AI you've ever seen will be today, right? It is only going to get better. It's not going to get any worse. It'll only get better by the day. And this progress is happening on a week-by-week basis. So, how can you reimagine every workflow with AI is the biggest opportunity. It's really where trillions of dollars lie as value to be created for every major retailer.
The second is about building your organizational brain. Every organization needs a brain which not only can know everything that's happening in the organization but can learn from every transaction and maybe even suggest what to do next to every single individual every minute of the day. That's possible. And that's the vision that many organizations are starting to build, is to build the organizational brain. That's a technology dimension. And third is you have to transform your people, because eventually all transformations require people to be transformed, people to work in a new, different way. So these are the three big vectors of opportunity across every industry, certainly for retail as well.
Now, you can see there are many, many examples today of companies wholescale reimagining their workflows with AI. You don't have to look too far out to see hundreds and thousands of companies taking this view and spending four to six percent of revenue on technology and a third of that on AI. So you will start to see that a third of your tech spend will go towards AI, and it'll help you reimagine every business process with AI. Number one. Okay.
An example of that is a piece of work that we're doing with a major retailer who is in North America, and they're one of the most prestigious retailers, and they're like a legendary brand. And part of that is also they sell gas, right, or petrol, near their retail stores. Now they want to be cost competitive. It's very important for them to be the cost price leader, cost leader for their customers, but they also want to be competitively priced with the surrounding retailers and so on, right? So they would look at all the information of the geography, zip code, and then come up with a pricing decision on their gas. It may change by a few cents every few minutes, but they have to make that pricing decision.
And they decided that instead of, you know, taking 5 to 8 hours to adjust the fuel, the price of the fuel, right, gas prices, they're now able to adjust it every 15 minutes because now they can put all those data together, algorithms together. They've figured out a way to be customer responsive to the customer and meet their promise that they made to customers and adjust the price every 15 minutes. This has happened because they have completely embedded AI into what they're doing. So this is one example that it can genuinely help you reimagine your process but also remain customer focused and meet your promise that you have as a leading cost leader in setting up the fuel prices and so on. So this is an example of something that we've done with one of the large US retailers.
Another example is, and this has a little bit of AI and some interesting human behavioral element as well. We were working with one of the largest retailers in the US, the largest retailer. And here we were working with a beverage company that was supplying, you know, stocking beverages inside the retail stores, every retail store. So every salesperson goes and spends, let's say, 15 minutes in every store, and then he has to make or she has to make 10 trips across 10 retailers, and they have to stock the aisles with their products. But what was happening was that they were not relying on the forecast of the retailer or of their own beverage company because they didn't think the data was trustable and so on. So the salesperson was making his own decisions or her own decisions as to how much to stock of which SKU, hundreds of SKUs, right? So because they didn't trust it.
So what we started doing is to help them by putting a—this is a Walmart thing. We started to put 'Walmart verified' in, just the way they now produce very good quality forecast, 98% accurate, but they were not using it. So we started to give them trust around the data that they were using, and by putting that this data has been verified by Walmart, and that instantly—I mean, this is not just one thing, there are 10 other things—apart from building a very good forecasting algorithm which was all built by AI, we also could give them increased conviction in trusting the data and therefore adopting those recommendations from AI and therefore stocking the units. In that process, they dramatically decreased their stockouts and increased the overall on-time and full situation. So that is an example as well.
I've got many other such examples, I know, but I don't have much time. But on the AI foundation side, we've worked with a bunch of these companies in building their digital brain. I don't have a retail example, but in insurance we are currently working on building their overall brain. So I'll skip that.
And in terms of helping companies also reimagine their work and workforce, one of the examples I like to give is that, see, today if you look at jobs around the world, they're defined by some of these things. Every job comes with a set of results I need to achieve and some activities I need to perform every day, people that I interact with every day, and tools and skills that I need to have to do a good job. Today my metrics are all process metrics, SLAs, volumes, and so on, and not outcome metrics. As AI becomes more and more interesting in an organization, you'll have to move from process metrics to outcome metrics.
In terms of the kind of activities I do today, I do a lot of manual tasks, things that AI can do easily. I might have to redesign processes so that I augment, automate, and redesign the processes for the age of AI. In terms of people, I may work with fewer people and work with a lot more agents and tools. In which case, I need to make sure that I adapt to this new way of work. And in terms of tools and skills, I might have to learn new AI tools, get activated on them so that I'm a manager of digital agents. Every single person in the organization, including all of you in the room, will be managers of digital agents that are working for you. So, how do you know how best to work with them and how to be an effective manager of digital agents? That's something all of us have to learn.
So this is the kind of people transformation lens also you have to apply. So people, process, and technology—all three levels of transformation—as you build a company that is relevant for the AI world. And I'll finish with just one more minute.
As you're building this specifically, it's very important that you build this AI and design it for trust, because the biggest problem in AI today is that you have an objective function. You want it to do something, but it ends up doing something slightly different. This lack of alignment creates a maximum trust gap between what the AI is doing versus what you want it to do. So this is called the alignment problem, is one of the most important problems in the AI world, and how do you build these systems that are trustable and align to your objectives. That's the one thing you have to think of.
Second, you have to make sure that you match and exceed human accuracy. No one is going to trust AI unless it matches and exceeds human accuracy in a consistent, reliable, predictable way. So how do you do that in terms of every process? And three is that you have to do this in a way that is making money, right? You have to clearly start with the outcomes, work backwards, and build in, you know, be efficient in how you use tokens and other things in such a way that it is significantly high ROI for you. So if you do that, you can really transform your organization. With that, I'll stop and maybe take some questions if you have any. Thank you.
