#337 Debdas Sen: Why AI Without ROI Will Die (Again)

0:00So I'm going to start by asking you to introduce yourself. I know you've been in the AI world for a long time, worked for several Fortune 500 companies, now CEO of TCG Digital, part of the Chatterjee Group, which is a multi-billion dollar global conglomerate. So why don't you introduce yourself, give some of your background so far that's relevant, and then we'll talk about what TCG is doing in the digital transformation world. Wonderful. Wonderful to talk to you, Craig. In fact, so I started my career in the area of what used to be called data warehousing in 97, long time back. It's coming up on 30 years.

0:48And I've been through all the changes that this area of data and AI has gone through. When I was doing my first set of courses in data warehousing, artificial intelligence was in vogue. And then somehow that word got lost and we kept talking about analytics, the data, and now it's a full circle. I've spent a lot of time for all my career in this area of data and AI. I used to run the competency in data and analytics for one of the large consulting houses in Europe and Middle East before I joined the TCG group about 11 years back. And ever since then, I've been evangelizing this area of what we call AI and trying to help clients around the world.

1:45We have an AI platform that we use to do very complex real-world AI. So, yeah, absolutely wonderful to talk to you. Yeah, and TCG Digital is a digital transformation consulting company, right? I mean, you... Well, it's an AI platform company, which helps transform our clients, particularly in the area of energy, in the area of airlines, travel, in life sciences. We also do sports. And so we do transform companies, but we do it with our AI platform, which we call MQ. I see. I see. In-house tech. So you're not like a systems integrator where you're pulling together tech from outside. No. Well, see, nowadays it's become an ecosystem play.

2:47So there is always an outside element to it. And I wouldn't call it outside because now everybody needs to talk to everybody. So what we do is we have our platform, which is MQube. And then we talk to what's happening in Cloud, Anthropic, what's happening in Gemini and Google, what's happening in OpenAI, what's happening in NVIDIA. And then we have the cloud providers, AWS, Oracle, Azure, etc. So we work in that ecosystem, but the core value that we add is we add it through our platform. Yeah. And the MQ, the platform, can you describe that for us and what it does? So we can pull data and this is where it started from.

3:40It's gone through a little bit of an evolution in the last few months, I would say. But let's trace back to what it was two years back. So you would pull data from wherever data is, bring it to a central location, and create something like a data lake or a lake house. And all of this is happening within the MQ platform. And then we would have reports, charts, graphs, which is business intelligence, or we would have machine learning models, supervised, unsupervised. We also have a knowledge graph through which we are able to contextualize data. We in fact have a very large knowledge graph in the area of life sciences, where we store a lot of information and being able to create insights out of that.

4:29And then there is an AI layer, orchestration layer, with which users can create models. There is model management and all that to be able to do that. So we work in large, complex problems, constructs, For example, let's say optimization of a large refinery. So that's a very complex problem. The net value added is usually very large, and it takes several months to get that done. So we come in with our platform, we look at all the data that's relevant, and then we are able to use the machine learning models. In this particular case, we use a hybrid model. So we use kinetic models in conjunction with AI models to be able to, let's say, improve the yield or improve reliability or improve just the visibility of the refinery.

5:27So that's what it was, let's say, a year or two years back. Where we are now is it's become a lot more agentic. so instead of just being a monolithic platform and we used to say that oh give us the data and then we will be able to do magic with that data we often reach out to where the data is and use the external information for example if if Claude has some good information in the in the particular context that we're working on or open ai or gemini we would work with that but keep the private data private so we work with large fortune 100 fortune 500 companies where they are very uh they need to be very sure that their data does not go out their model says does not go out so we take the best that's there in the external world bring that in and then we corroborate with the internal data to create insights.

6:32So it's a completely an agentic platform today where we have our in-house agents which interact with the agents from the outside world and these agents could sit in any of these language models or any other platform to be able to create insights. So today we don't say we are a platform player anymore because I don't think there is anything called a platform player anymore. So we are an AI ecosystem player. We contribute to the AI ecosystem so that we can interact in an interoperable manner to be able to create insights. And that's what we do. The core capabilities of the platform still remains.

7:18It's deep in machine learning. It's deep in business intelligence. very complex data management structures, but things like model tuning or even inference or reasoning capability that we use some of the external models to do that. So AI is super evasive today. I mean, we can't say that we do it all. I don't think anybody can say that we do it all. So that's what we do right now. And it's picking up a lot because what earlier enterprises used to do is they should see that here's our data give us insights from it now the problem statement is here's a business problem doesn't matter where the data is get us the best insight so we try to make the best of what's there in the enterprise and what's there in the external world and to be able to generate value sorry i interrupted No, I was going to ask if we could sort of zoom out to 40 ,000 feet and talk about digital transformation, that whole market, because I hear a lot of, I talk to a lot of companies, I hear a lot of companies that are saying similar things.

8:41and from an outsider's point of view, it's very hard to differentiate what companies are doing in one domain as opposed to another domain or what one company is doing as opposed to another company who has similar marketing language to be. Right. So when we do our AI, when we do our digital transformation, we are very domain led. And this is where the Chatterjee group comes in. We have a sister company, Lummis Technology, which is the largest chemical engineering process licensor in the world. They have 3 ,000 refineries and petrochemical plants and other things in energy where their licenses run.

9:39So MQ works in that space, in that ecosystem, to be able to increase the yield of a refinery by 2 % or improve reliability of a petrochemical plant by X percent. So this is real world large problem sets that we focus on. Or we have another sister company, the name of the company is LabVantage, which works in the area, it's the largest laboratory information management system company in the world. And we work in the area of pharmaceuticals, in the area of life sciences, where we're accelerating the R &D of novel drugs. And that's, again, a very large envelope driver statement. And we bring in complex domain encapsulated in the knowledge graph that we have, have AI overlay with that, and then take advantage of what's there in the external world, let's say what perplexity has or what anthropic has, and be able to see what a scientist needs to be able to accelerate the output that the scientist is looking for.

10:56So from a differentiation perspective, we are talking about very, very deep domain, whether it's kinetic models and chemistry or whether it's biologics from a life sciences standpoint. point. I mean, we've got thousands of PhD scientists whose day job is to just do that core chemistry or biology. And that's what we bring into play. And we do a hybrid solution where the core science is a part of what we provide to the client. So it's not pure AI. So that's our differentiation. Yeah, and then when you look at what other companies that either have platforms or are building tech but providing consulting to customers on how to implement the tech, Do you see it breaking down by domain expertise, or are many of the companies more horizontal?

12:11I mean, there's this world of companies helping large enterprises transform digitally. and that means today that, you know, building agentic systems, you know, in the business processes. And I mean, how do you see that? Is there just so much business available in the market? Everybody needs to go through transformation that these more horizontal plays will find plenty of takers or do you think it's evolving into domain expertise, more vertical plays? When business value is added, it's almost always there is a domain specificity to it. And that is very critical to the solutions that we provide.

13:12We typically work with core functions of large companies. uh if you're talking about enable function enabling functions like finance hr supply chain etc that's where the horizontals can play very well but what we are working with you know the type of work we do so we work with accelerating r d or we work with making manufacturing more optimized there's a huge domain play in in such cases and and that's our niche well you were talking earlier about insights about that that is the output of of your system is it insights or is it implementing uh agentic uh workflows uh yeah i mean is it business yeah at the end of the day it's it's impact and business value the means to an end could be could be one or two things and I'll come back, come to that in a moment.

14:18But the end goal is to create business value for our clients because for us, return on investment for our clients is super important. And if somebody is paying$5 million, they should get back$50 million. That's our rule of thumb. So the means to that end is a combination of what's there in the legacy infrastructure And what can we bring in with the agentic infrastructure? Trying to say that for these large organizations who we work with, trying to say that, oh, we will completely ignore legacy and build an agentic system, that ain't going to happen in six months. So the ability to connect to legacy, and there are tools available to do that right now, like the MCP protocol and all that, is a part of the solution.

15:18And agentic and reasoning solutions are very much a part of the solution, but we cannot ignore core data management because at the end of the day, you still need to do a fair amount of plumbing because it's really surprising. I mean, a lot of companies still run mainframes and RDBMSs. And so we will still need to connect to those data systems and be able to integrate that data before actually creating the agents. And that's the enterprise data. I was talking about the fact that we play both with enterprise data and the external data. The external data that's sitting in the language models, that's all done by the language model players.

16:10multimodal model or language model. So we take that information in and be able to corroborate, validate with the internal data. And that internal data, there's a lot of data management that needs to get done with that. So it's not one or the other. We think value is to be able to make the new data interoperate with the old data to be able to create contextual insights for the enterprise. Yeah. Can you give us an example? Say a large petrochemical plant comes to you. What's the process? You sit down and do an audit. I mean, yeah. So we always start with the business problem. The front end of our services is very much consulting.

17:17We sit with the client and try to understand what is it that they are trying to solve. that that piece is still traditional old school consulting maybe now it's two weeks earlier used to take three months but that two weeks is still there because i still believe that defining the problem well is actually half the solution once we do that and i'll give you an example so we are working with a top five oil and gas major and they are trying to understand how catalyst can be developed faster and catalyst is a is something that they use to use in defining and petrochemicals to be able to uh catalyze the chemical reaction that happens how catalyst can be developed faster Now, this is R &D information.

18:19And this R &D information that this particular organization has, it's very, very private to them. They don't want to give this information to anybody else because this is core IP.

18:36However, the world of catalyst development has significant amount of external information. Earlier, to be able to formulate a new catalyst, and this area is called formulation, to be able to formulate a new catalyst, they would go back to the experiments that they have done before, what materials was used, and try to come up with a new catalyst. now you would go out to the external world and see what all information is available of all research that's there in the public domain and there is a lot in the public domain i mean every time we go to a chat gpt or a jamana or a perplexity we always get surprised just how much is there yeah and then we bring that in and we contextualize it with the internal information and effectively run millions of models in the virtual space, saying that millions of different formulations in the virtual space with agents reasoning amongst each other and so on, and then come to 5 or 10 or 15 saying that go and test this out in the lab.

19:45So this is a real-world problem. now to be able to come to those five candidates to be tested in the lab earlier it used to take 12 months yeah today it takes one month so r &d is completely collapsed you know giving giving another example and this is a very large uh uh refinery in uh this this particular one is in India and companies called Hindustan Petroleum and they have come out with one of the most sophisticated refineries in the world. It's a it's a Lummis technology process called LC Max. It's a Chevron Lummis. There's a joint venture between Chevron and Lummis called Chevron Lummis Global.

20:34It's their process and they wanted to make this AI ready on day one because they wanted to run this enabled by AI on day one from a reliability standpoint, from an optimization standpoint, from a visibility standpoint. And the reason we were able to do that is because we were able to take all the information that's inside and be able to bring in what happens in these kinds of refineries from outside to make them run reliably and optimally and be able to do that. And And these are typically multi-billion dollar revenue ecosystems. So that's what we are focused on, trying to take really large problems and try to see if we can optimize half a percent, 1 % of it, which translates into several tens of millions of dollars of value.

21:32Yeah. And just on the catalyst example, what would get you that 1 %? I'm just curious, what is the process? We are trying to bring down the development time, cut down the development time to one third of what it usually takes. So innovation cycle is cut down to one third of it. That's value. I mean, if it takes five years, they would get it in a year and a half, which means that for the rest three and a half years, they would have a far higher yield going forward. Yeah, but specifically, I mean, obviously, there would be different points along the way that you would optimize. Is it providing, you know, the chemical engineers with...

22:30So let's take an example. So I'll probably simplify this. So they're trying to come out with a new material effectively. And that new material has ingredients and then they do something called a formulation. You effectively mix different ingredients in different ways to come out with a new material. That's a problem statement. and that material needs to help some reaction happen faster. That's the objective. Now, there are millions of materials and they can be combined in another million ways to come out with this new material that can be used to catalyze a reaction. So a million materials in a million ways is something that it's humanly impossible to figure out.

23:25But in the agentic virtual world, we can run as many experiments as we want. So that's what happens. So learning happens in the virtual world. Agents talk to each other, they reason amongst each other. And then they come out with saying that, this is one combination that could work. So we never go around to one combination. We give them five combinations, ten combinations. Then they test it out in the lab in real scenario to see. Right. But you're doing that work of coming up with the five candidates. You're not giving the engineers at the petrochemical plant the tools and the training to do that on their own.

24:16You're acting as... So we are giving this tool to the research scientists and the labs of that particular petrochemicals plant, let's say. So it's not the engineers who are running the plant, but it's the engineers who are doing R &D for the plant. Right, but you're giving them the tools that they can use. Or are you running, you know, finding the five candidates and then giving them the five candidates? We work with the engineers and we work in a collaborative manner to come up with the five candidates. I see. And is this on a problem like that? And then you're done or is there a continuing relationship?

25:08relationship? There is a continuing relationship because they would still need the platform, number one. Number two is our involvement comes down over time. So the first catalyst, they might need, you know, maybe 60 % of the stuff is ours, effort is ours. By the fifth catalyst, maybe 5 % of the effort is ours. So our objective is to enable them, because this is a core function. As I said, we work with core functions of organizations. So we don't, we don't work on the enabling function. So they would never outsource something like this. So our objective is to enable them. We leave the platform in.

25:52I mean, that's how we make money. Yeah. And you're combining insights from their data with insights from your knowledge graph, from your own bank or knowledge source. Knowledge graph. We have that, yes. How important is the knowledge that you're bringing in? It is quite important because we have a lot of chemistry contextualized in our knowledge graph. So we do that. So we take an external data, let's say all experiments done for various things, whatever is available. And then we curate it into the knowledge graph. And we do the same thing on the biologic side. So it's three pieces of information.

26:47It's our knowledge graph. It's the internal data of the organization and the information that's sitting in the public LLMs. Yeah. All three come together. Yeah. And what differentiates you maybe from another player in that domain is the curated knowledge graph, I would guess. If competition is another player in the domain, what differentiates us is the platform the depth of ai in the platform if the competition is a horizontal ai player then what differentiates us is the domain i see ability to play with domain data and technology is what differentiates us yeah and on these problems you know taking the catalyst as as an example not that i understand i mean i know what a catalyst is but uh is the tech improving or or the depth of the knowledge in the domain accruing so that you're you're finding better and better solutions or are these really can only as far as you can go.

28:08It always learns. It always learns with every single project we learn. We learn from the information that we get in our projects. Of course, we leave the confidential information with the client. We don't touch that. But the metadata is available. And then we always learn from the external language models. So it is always learning. And every project that we do is a little faster than the previous one. Yeah. Where do you see this going? So you're optimizing industrial processes in these specific domains. I would imagine a couple of things are happening. One, companies that undertake this process get a competitive edge over other companies that don't.

29:10But also, there's got to be an accumulated economic effect as this kind of optimization works its way through an industry and through the economy. do you see that i mean how do you see the the future uh given this kind of thing it is it's actually so relevant with what's happening in the world right now so if we can get a two percent improvement in the yield of let's say motor oil or diesel yeah of the refineries which doesn't need to pass through that little strait of water which everybody's fighting over it's huge and And that's happening for us right now.

30:00And more than optimization, reliability is also very important. So all these plants are being pushed around the world to produce more and more and more, which means that they have a higher chance of failure. So we predict failures. We predict failures and we are able to do this using our, in fact, it's just right behind me called the Agent to Command Center, where there may be a petrochemicals plan running in Nigeria or somewhere deep in Africa, but the subject matter expert may be sitting in Houston. And to be able to look at the reliability of the plan to see that, oh, it shouldn't break down and be able to give advice is critical and it is it has a far bigger business impact than than the one percent two percent improvement in yield because once these plans go down they take like weeks to come back up so in in these trying times loss of production is crazy And that's what we're helping clients with.

31:13So the economic value is incredible. I mean, it is very focused on the economic value for us. And while the R &D economic value is not very easy to calculate, our manufacturing side of the use cases, it's quite easy to calculate that. Yeah, and as I was saying, that accumulates that economic value. I mean, you'll start seeing an impact in national GDPs as large. I was just talking about this one plant in India. It's the largest refinery in the country, multi-billions of dollars. And so particularly for a percentage of GDP is a calculation, depending on where it is, but it's definitely relevant for large enterprises.

32:12It's big enough to be relevant at large enterprises level. And now large enterprises output adds up to a country's GDP. So I wouldn't go there, but certainly it is large enough to be relevant to large Fortune 100, Fortune 500 companies. Yeah. Yeah. I mean, it's just it's an exciting time because all of these optimizations across industries are, you know, it's a little bit like, you know, the Internet. when the internet was introduced and suddenly information could flow much more easily and there was a big productivity leap. We seem to be in that era again today. I think there is a 10 % productivity possible today with the agent systems.

33:17and we are working with clients across various roles to be able to create productivity gains, whether it's in manufacturing, whether it's in R &D, or whether it's in supply chain. These are the three areas that we primarily focus on. We don't really look at finance, HR, et cetera. So I think the productivity gains are incredible. some of these are and and i'm saying to the extent of 10 i mean particularly let's say in a supply chain if we are able to optimize across the supply chain let's say for pharma right from where you are bringing in the uh the drug development process through the uh the phase one phase that the clinical trials process and can we collapse that by six months that means that it gives the uh the the pharma company six extra months to be in the market and for some of these drugs it it can be it the the values are staggering so reducing the innovation cycle and uh and then of course optimizing operations i think the overall long-term impact of the pnl can be as high as 10 yeah uh you're you're uh you know there's been a lot written about the slow uptake of agents in industry because of trust and reliability uh issues how have you guys uh managed that or have you encountered reliability issues?

35:12So that's our differentiation because the agents that we use are within the enterprise boundary. Some of our clients, particularly in the Middle East, MQ is running in their data center. we might interoperate with agents that may be sitting with a language model or a cloud provider but the final decision making happens validated with the internal enterprise information so that's our answer to the very question that you asked so by definition the process that we follow does not have this trust issue because it's validated with internal information and it's tested enough with internal information.

36:05Does language models hallucinate? Yes, it does. Can it be brought to a place where the percentage error is acceptable? Yes, it can be. And that's exactly where we focus. Validate the information that's coming from the external world. make sure that hallucination is to acceptable number because when it comes to management decision making they don't care about the third place of decimal that that's actually our that's specifically our answer to the question that you asked we are we're actually doing very well with exactly this because the tech is ours the uh the information the data is validated with the client It is validated with the domain.

36:56But we take advantage of what's there in the outside world. But we validate that with what's there in the enterprise space. What are the problems or what are the mistakes that large energy clients, for example, make in trying to implement AI? I mean, do you see, do people come to you after having tried to do it on their own or try to do it with a... All the time. Yeah. And what are the mistakes you see? So in the energy space, what we have seen is they either go first principles, which is, I was talking about optimization. So you could use something called a linear programming or a multivariate programming to optimize a process.

37:46Or they go all AI. You know, throw it to a neural network, whatever comes out, let's do that. I think the answer is somewhere in between. That's where we sit in the cusp. And they, so that's from a risk perspective where they fail. But the advantage, the opportunity loss is not taking advantage of what's there in the public domain. Because that today is also quite beneficial. So hybrid modeling, taking advantage of the public domain data is the answer that we've seen. And the mistakes that clients do is go one way or the other too far. Because marrying the two is not necessarily easy. Because in the energy space, there's something called mass balance, which is like you have a chemical equation, the two sides have to balance, going back to high school chemistry.

38:43and the chemical engineers swear by that. And that is important. That's what's sitting in our knowledge graphs, making sure that the answer that we give is something that can get validated by the chemical engineer. So that hybrid modeling, hybrid between chemical kinetics and AI is what makes the difference. uh how long has mq been around or how long has the agentic layer of mq been around well we mq has been around for about seven years i mean we really launched it about just around the pandemic so maybe maybe five six years the agentic uh so we do a release every three months and we did our first agentic release in october last year uh no sorry in in july last in the middle of last year so we have been we've been around for about maybe six nine months uh so we we would i i would never say would complete but one version will we We'll reach a milestone of everything that we want to do by our release in June.