EP. 48 CIO’s Survival Guide to the Agentic AI Era with Sumit Taneja

00:00 - Welcome & Guest Introduction

Matt Pacheco
Welcome to Cloud Currents, a podcast that helps navigate the ever-changing seas of cloud computing, cybersecurity and emerging technologies. I'm your host Matt Pacheco and I lead the content marketing team at TierPoint, where I help businesses understand cloud and security trends to make better decisions about their IT strategy. Today we're very excited to have Sumit Tanija, Senior Vice President and Global Head of AI Consulting and Implementation at EXL Service. Sumit started his career as a pharmacist india, transitioned into the BPO world and worked his way up over 25 years to become one of the leading voices in enterprise AI and cloud transformation. He's a cloud certified architect across aws and Azure, EXL's internal authority on cloud solutions design and the person the CIOs call when they're trying to cut through the hype and actually get AI to work.

Today we're going to explore Sumit's remarkable career journey, the evolution of his company into cloud native AI powerhouse, the pitfalls and promises of agentic AI in the enterprise and why Sumit thinks companies may not be ready for AI yet and what they need to do to get there. So we have a lot of good advice, a lot of good learnings here today. So Sumit, very happy to have you here today with us.

Sumit Taneja
Same here. Very happy to be part of this podcast, Matt.

Matt Pacheco
Cool. So, so let's jump right in. Tell us a little bit about your history, where you started and what, how you got to where you are today.

Sumit Taneja
So I know my history may be a little different than most other people because as you rightly said, I am a pharmacist, still a licensed pharmacist, but I right now consult companies on how to make them AI native, how to embed AI in their enterprise. And it's a very different world, right? I learned how to dispense and manufacture medicines and here I am using AI to transform enterprises. But I think one thing which probably drove some of that was my curiosity to do things, probably because I was lazy. So I wanted to do things which can be much faster and somehow that has worked in my favor.

So after the first few years of working in pharma companies, when I started in exl, at that point doing simple scripting in day to day sort of operations was not something which everybody was doing and that paved the way that honestly I didn't want it to create daily reports or have mis which used to take hours of lot of people to data crunching. I was always looking how to automate that and how to make sure that I'm only spending five minutes doing that kind of work. That meant I had to devise ways and means while I was not a coder myself. But I could learn, I could figure out how best in class people are doing across the globe and just use some of that learnings and deploy it.

And slowly what happened was we started doing this, me and my team in EXL across EXL and then were building a few applications. I was given a charge of design as well as getting that application built out. And that started my journey into a more digital era of building applications. And obviously over the last 25 years EXL as an organization has also pivoted from being a more operation centric organization to analytics and data led to now where we are using data and AI in large parts of our business. So EXL, for people who don't know, we are a $2 billion organization and 55% of our revenue is data and AI. And that has given us a, a big stride in whatever has happened in the last. Sorry, you'll cut that.

In the last few years, this whole data and AI practice where EXL has pivoted has really given us the double digit growth. And we are continuously growing because we have the expertise of operations, we know how to use the models and we know how to operationalize AI into the client ecosystem. And it's a journey. Honestly, I would never would have imagined when I started my career or I was studying in college that I would be doing this after whatever, 25, 27 years later. But I'm almost sure. So when I speak to my daughter, I tell her that whatever you are studying today, don't get caught up into that mindset that this is what I will become in five years. Because the world will be very different. You need to be open, keep learning, keep be curious.

But don't get caught up into one kind of careers. That's very dangerous today. You don't know what careers are going to be.

Matt Pacheco
Change is the only constant these days. Really interesting. So starting off as a pharmacist, how does that I guess influence this? If it does, how does that influence what you're doing today at exl? Are there any learnings that came from that? Because it's really interesting career path.

Sumit Taneja
Yeah. But I'll tell you. In my final year when I was in my pharmacy course in India, one of the final projects I submitted was to digitize the statistical models. That was a time when people used to still use scientific calculators, but still do all the statistics on calculators or paper. The amount of usage of regression models was not available freely. So I worked with one of the Computer Science PhD candidate to create this and submit it. Now I had no idea that this is going to be part of my career in future. But if you sometimes look back, maybe that was the start of being close to something. So pharmacy actually uses a lot of statistics because when you create new molecules your knowledge of statistics is extremely important.

Statistics is also a ground for all the AI model right where we are today. So some of the learning which I had there, I did statistics in process re engineering on using, creating control charts, simulations when I was doing process engineering. So all of that sort of comes together. Right. So what I always believe is whatever you learn never gets wasted. It may help you in different ways. So for example, my statistics knowledge in college helped me in my job when I was in Excel at least the early days. My ability to at least think that this should be digitized, the statistical models helped me to think on how to look at world in a very different lens and not keep doing what people have been doing in the past. Right.

And on a very lighter note because I understand the medicine world very, very deeply. So when I lead teams and if somebody gives me a wrong prescription, I know he has fudged that prescription because I know what the doctor will prescribe at what time. Right. So that's on a lighter note. But so whenever there is somebody who's not coming in office for a few days, I will get into detail how truth. That is very cool.

Matt Pacheco
Yeah. Our experience prepares you for your next thing. What drew you to exl? Like what, what made you go specifically there? What did you, what do you like about it?

Sumit Taneja
Honestly, it was more coincidental, not really planned. So I had done maybe about three and a half years in pharma, a bit in medical transcription. But most of the jobs in that area was in sales. And this is the research india at that point was not that great, which is, which would have been cool. The BPO sector was coming up and this was almost like a stopgap. I would have moved on to something else in one or two years. But because I, after moving to exl, I rapidly almost kept changing my role and growing in roles every six months to a year that I never looked back. So I started almost taking calls as a collections agent for one of the US company. I moved into quality which, where I used all the statistical knowledge I had.

I became part of the team who was managing the application design and all the automations. I went into consulting, managed operations, came back when were setting up digital and AI function. So the path I took was not something I had planned for and something I always believed that I go by the flow. But I will keep doing something which drives my curiosity and that keeps you relevant. Right. In today's time, even in the last, whatever 12 months, this whole agentic AI has come up. I can code agents now at least in six platforms and to me it's very relevant because when I'm consulting with the clients, I need to know how real the technology is for me to sell them the idea of transformation.

So I don't think there is one thing which kept me with exl, but I think the constant innovation in my role and constant change has driven me to almost think it's a new role. Right. It's a new world. What I have been talking to the clients for the last 12 months is something which I never did for last 25 years. So it's relatively to me a new.

11:19 - EXL's Cloud Migration: From On-Prem to All-In Public Cloud

Matt Pacheco
Role that's really exciting. And with the world changing and the landscape changing, I'm sure it'll constantly feel new to you every little bit. Let's talk before we dive into AI, I want to talk about the cloud piece because that's going to tie into what we talk about with the AI a little later. So from what I understand, EXL was largely on premise until about 2018. Can you describe what your cloud migration looked like and like what drove the decision to go all in on, from what I understand, public cloud.

Sumit Taneja
So exl, because we service a lot of regulated companies from insurance, healthcare and banking, our data security practices have always been extremely high. So if the client A requires 10 things and client B requires 12 things, every client gets 12 things, right? So it's always the highest. And were always very conservative on moving to cloud till about 2018 and again in 2018 we took a gradual step that let's start this journey. We are seeing a lot more applicability of AI because transformer model has already come up in 2016. 17 we had already started to do R and D on it. So actually our first foray into cloud was the R and D team got access to the AWS environment and they started to play around. Right. And then we created a more enterprise roadmap on what all how that pace would be.

But if you ask me, the acceleration happened with COVID That was the final almost nail in the coffin that now it needs to get done right. Otherwise it may would have dragged for maybe one or two years. But Covid, with all the areas, you're not sure whether all the data centers can be manned with all the distancing norms. It just drew that we had to do it faster, otherwise we would have taken a few more years.

Matt Pacheco
Yeah, I can imagine even the chip shortage during that time, I think there were 2022. There's a lot going on there. It seemed like a lot of companies at that point their covid was the catalyst for moving to the cloud and they seem to have not stopped. It seems to everyone's kind of marching that way though we still see a little hybrid cloud here there for certain workloads. But you've said in the past that if a client today were to ask you to deploy on premises, the Excel would say no. That's a pretty bold stance because some people still like to control their data or at least see their equipment. How did you get to that point? And then also how do clients react when you tell them.

Sumit Taneja
So? Surprisingly I will not say 100% but close to 98, 99% of the clients are really now comfortable with cloud. And this has happened between post Covid between 21 to 25. It's been a transition, but now I don't think it's ever a discussion. Even in 2122 we used to have this discussion on security, what's approved, what's not approved. So for example, one of the insurer in Australia was telling me that they took one year to approve AWS in 22, something which most companies would have approved like 10 years back. They took one year to just to approve AWS. So when we speak now to the clients, I don't think it's a challenge. But if you ask me why went all in.

So in 2020 onwards we started heavily working on AI solutions and when we started to leverage services from AWS, Azure, Google, we realized that you can be faster, you can try new things, be much more agile rather than getting caught into oh, I have bought in this GPU or this CPU and now I have to use those services. I could buy similar capability or even better in the cloud I'm much more agile, I can manage my cost much easily. I can open a environment and shut it down tomorrow if required to. I can't do all of that in a on prem environment. So just because went all in with AI and this is even before Genai came in, AI was more in ML and NLP and NLU kind of categories.

But even though that AI, which is now I would call it traditional AI, required intensive compute and that intensive compute was easily available in cloud rather than in the on prem data center. So we had for example a GPU even till 2022 in one of the data centers, which we of course finally shut it down. But we realized that we are not nimble and agile if we don't go all in cloud.

16:43 - Agentic AI in the Real World: Claims Letters & Invoice Auditing

Matt Pacheco
So you mentioned the word AI. We have to talk a little bit about AI because I know that's a big part of what you do. So thank you for that setup with the cloud. So agentic AI, let's start there. Some might say it's a buzzword at the moment. You hear it a lot of businesses are implementing it and getting familiar with it. But you've already built and deployed agentic systems for clients. Can you walk us through some of those projects, how you did that and how those work?

Sumit Taneja
Before I get into a couple of examples, I'll also want to say, look, there is a little bit of hype and which companies are calling everything as agentic, which may not be as true, but the technology has lot of potential, but it needs a lot of work on design, it needs a lot of work on operationalization, change management. I don't think all the problems which we are going to face in the next few years has been fully thought through or solved, but the promise is there. I think hype is a little too high as compared to where the real value is being delivered for clients. So for one of the clients, they send claims correspondence to all the customers, whether the claim is accepted, denied or partially given. And in the US it's a very regulated environment for sending those letters.

Because when you are sending a claims denial letter, which is probably the most negative a customer can hear, it needs to be backed by sufficient element of details. There are a lot of state specific regulation which needs to be referred in the letter, etc. Etc. Now this client of ours sends about a million letters a year and they have hundreds of people doing this on an ongoing basis. And client said, look, if we want to really test the AI, let's test in this use case and see if it really adds value. And can we completely do this autonomously? Right. As much as possible? Obviously there will be some exceptions. So we said let's take this as a challenge. And they said today a human takes about three to five minutes to draft the letter and then send it to the print.

They said, let's take a target of doing this within 60 seconds done by the Autonomous agent. Maybe there is a little bit of approval before you send it for the print, but let's create a letter generation agent within 60 seconds. Sorry. So when we got this requirement for 60 seconds from the customer, we had two things in mind. One, we have to follow all the guidelines compliance. Second, we have to do it fast enough. But the complication is you have to get information from multiple systems. So what happened on a claim, a summary of that, what are the amounts customer was claiming and what was discussed? What are the invoices we have got? Are those valid? So all of that needs to happen before you draft the letter, right?

So because all of that is feeding it, so we had to integrate to all of the systems, get the relevant information, create a repository of all of that data and then use different agents to create. So we created a few agents. One was to generate the letter. Second one was to check the quality. Third one was trained on specific state compliance regulations. Fourth one was it will be formatted in a way which is relevant for that client's formatting guidelines and then it will come to a human for the final review. So we created these agents we initially started creating on Microsoft. We used Autogen as one of their agentic frameworks. But we realized were taking about close to 200 seconds which did not match the 60 second requirement from the client. We investigated.

We figured out that the agents within the Autogen for the ecosystem exchange a lot of pleasantries between them when they hand off from one agent to the other. And it's unnecessary time lost in those pleasantries. But that's how it is, it's designed. We can't change. That's part of Autogen. So then we piloted on Crew AI and were able to get it less than 50 seconds. Same design, same largely same code base, but using a different framework. We were able to get to that outcome. Now, for most people, this may look a very simple use case to generate a letter, but believe me, a claims denial letter for insurance company to send in an autonomous fashion is not an easy decision. Because if that letter gets challenged, they can get sued, they can get, they can be taken to the court.

It's a very critical part of the claims lifecycle. But for another client we were trying to look at, can you look at payment of all invoices and see if there are overpayments or underpayments? So the process was very simple. There are suppliers who have set contracts will have SLAs, payment discounts, etc. Etc. And they will send all the invoices and the insurance company had to pay. We realized that this process, while it was manual, you can automate it. But the reality is that there is a lot of value which is being left on the table because nobody is auditing the invoice against the contract. And that can be a area where. So we said, let's create a agent to extract all the contractual terms. There'll be one agent which will extract the all the details in the invoice.

It will match and tell you what amount of money we should be. We are overpaying or underpaying. We found within the first six months a million and a half got saved because were able to figure out if they pay 10 days earlier, they'll get a 2% discount. This as because they did not meet the SLA, they cannot charge me the full amount, they can only charge me 95%, etc. Etc. But again, this process was running for years. Nobody touched it because there are millions of invoices which comes in. Nobody looks at contract and invoices and matches and sends it because the putting a human workforce behind it is just too costly. But with AI, we could just build these two agents and get the work done very easily. And we still have exceptions.

So we have, I think we still have about 15, 20% exceptions which the. We are not able to decide whether it's the right thing or a wrong thing, which the human reviewer will look at it, decide and then process it. But large part of the process is now automated.

25:07 - Governing AI in Regulated Industries: Trust, Traceability & Guardrails

Matt Pacheco
That's really interesting. When you're using AI in some of these like heavily regulated processes, how do you think and how do you approach the governance of the AI, the security, the safety in those deployments? Because again, that's, that's an agent is someone new essentially looking at that data and having access to that data. How do you ensure that security and safety.

Sumit Taneja
There's so a few things is extremely important. So I'll talk about two or three concepts. So one is how do you train the agents, which is the knowledge transfer. Second, how do you create trust in the whole design? And third, while the gen AI or agenti is very probabilistic, how do you still create a deterministic cage around it, right? So that you know it has to operate within the boundaries. Let me talk about each all these three. So first, we do a lot of work in capturing all the tacit knowledge which sits with the people who have been doing some of these processes for many years, right? Because typically every client has SOPs, procedures, documents, they are only for the happy path. That's not the path which they usually take when they do the work on an ongoing basis.

So while we consider all the SOPs and procedures, we look at observing those agents on a continuous basis. So we deploy software which will mine how they are doing the work for say last three months, right. And extract that information and then train the agents on. That's the first thing is can you be very deliberate about transferring the knowledge from the people who have been doing this work for 5, 10, 15, 20 years to this new AI agent? And can you be deliberate about it? Second, once you have done that, how do you create trust that so for example, in this claims denial, because once the claim adjuster had made the denial decision, it has to be trusting the AI to send the right letter. Because if the claim adjuster doesn't trust it, they are going to go and manually create the letter.

Because they believe they cannot do wrong thing for the customer, they should do it themselves. So you have to create the trust. So that means when, for example, when we create the letter, we give all the what all data sources they considered why did they write the letter? The agent wrote the letter. So you could almost see the complete history of what the agent is taking an action, why that agent is taking the action. Second, even when we are building the agent, you have to involve the claims adjuster so that it's. We are considering on how they think about when they are deciding to send the letter to the customer. Right. And we are taking all the input. So the trust by design is a complex thing. There's a human element which is to take them through this journey.

But then there is a element of giving all the traceability so that there is a trust which gets created when it gets operationalized. And the third element which is how do you create the determinism around all this probabilistic model? I know it may sound complex or confusing, but we know that the letter can be generated in a manner which may not be what the customer needs, because it can hallucinate, it can give you information which is not supposed to give. And that is where there are guardrails which needs to be established, which is it checks. This is the type of letters for this is the type of claim which usually goes. There is a quality check agent which checks whether these are all the elements which are required out there or not. Does it validate to the customer claim?

Is it the date of claim policy? All of that information is validated. So you create an element of certainty that this letter is accurate to best of my knowledge, and that guardrails then ensures that the letter which is going out. So all of this is still very is being developed or learned by most companies, including exl. So we have learned this by doing this for customers over the last two and a half years. Otherwise we would have just said, okay, just give all these letters. I'm sure the AI will generate a letter. But we realize that unless you do this very in a very disciplined fashion of transferring the knowledge, building the trust in the whole UI design in people's mind and then ensuring that there are enough guardrails that nothing sort of goes wrong. I'll give you this.

There's a paper which was written called Goat paper on red teaming about maybe about 15 months back, which were different ways you can get a wrong answer or answer which the agent is not supposed to give. But the whole idea of that paper was if you test, which is adversarial testing, you will be able to build guardrails which will prevent it not to happen in real production. So now adversarial testing is now a practice within the agent AI deployment cycle, which is not there in any software life cycle. In software life cycle, we actually want to test what function and features should work. In AI testing, we also test what IT should not do. And that's a very relatively new concept. It's not done in typical sort of software testing.

Matt Pacheco
That's really cool and really interesting. So you talked a little bit about convincing the human agents that this is legitimate and this is correct and having the agents doing the job the right way. But thinking about the organization as a whole, you have to convince a whole lot of people to take this leap because that's a big risk for them. Especially when it comes to something so important that like you said, if it's wrong, it can result in lawsuits. How do you convince an organization from the IT leadership all the way up? Because that's probably a board level convincing that's happening as well. How do you go about convincing them that this is the right move and this would work for your organization?

Sumit Taneja
So at least this is where hype has helped a little bit. Most boards or CXOs are actually wanting to get this done. Now what it may still differ is how fast they want to go at it. Do they want to go first or they want to be fast followers? Those are obviously different strategies, what kind of governance practices they want to ensure before they. They go full on. So that I think is varied across clients, but Believe me, two years back or two and a half years back we still had to discuss how AI can help you and it's most transformative. I don't think those discussions are happening today. Today it's, we know we have to do it, let's figure out how to get this mobilized.

I think at least at the board and CXO levels, I think when you start to go a layer below then you start to have more challenges. Because unless they have a mandate, which is whether it's a mandate on saving cost or it's a mandate of deploying AI because they want better customer outcomes, that's a different because then it drives different behaviors. But if it's only so I know an insurance company which did not have this one of these large objectives which is either customer or cost. They just wanted to implement AI and believe me, they are nowhere because they signed a contract with OpenAI as an enterprise AI partner, they got the licenses, the email goes to the rest of the company that we need to implement AI across and people have zero idea on how to use it.

We know people have been trying to just upload information and then figuring out oh, I'm getting wrong answers, I can't trust this. So unless you have you drive the whole organization with a single minded objectives or a couple of big objectives and then use AI as one of the levers, then you can still get to the outcome. But if you're only trying to implement AI because AI is the coolest thing, I don't think you will land anywhere. You will only land in a few million of license cost which I have seen in quite a few organization, zero roi. But there is a license cost and.

35:36 - Why Most Enterprises Aren't AI-Ready: The 3-Horizon Framework

Matt Pacheco
That's just one of the issues of rushing into AI. There's also the piece of AI readiness. So let's talk a little bit about that. As you said in the past, the most common problem you see with enterprise clients is that they are not AI ready when it comes to their data. Can you explain what that means and why it's such a widespread issue?

Sumit Taneja
And especially we work with relatively top 10 top 20 banks and insurers and healthcare companies. Now these companies have been here for 50, 70 and some are even 100 years. So they have all kinds of systems whether they have homegrown or they were acquired with acquisitions. But there is a plethora of admin systems which can be green screen to. One of our clients has FoxPro. One of the clients has a system which is running on all legacy languages and some of these systems don't even support APIs, you just can't even integrate. And all of this data is now hidden into these systems. And unless for the enterprise, I can have access to that data and I can even use AI in the most real way. So when we talk about implementing AI, we actually talk in three horizons.

Horizon one where I'm doing summarization, I'm doing classification, I'm for example that letter generation, that's all Horizon one because I'm trying to make one part of the process more efficient. This can be done in most organizations even if they don't have a lot of the data state correct. Horizon 2 starts to become more complex because I'm trying to now look at one function more end to end. So complete claims. So I need to now figure out whether all the claims data is accessible and I can use that to almost embed AI in the complete workflow. Horizon 3 is where things get relatively complex, which is what we will call as first principle or AI native workflows.

The Horizon 3 is where things become much more complex because what you need to do now is make your function or process very AI native and it should almost follow first principles means I will now need to invest in a data layer which is a coherent. It is real time accessing all data across systems. So that intelligence layer need to be built whether it's knowledge grass, whether it's a semantic layer and then only you can get to that. I would say almost like a nirvana of how you will implement AI. So and what we are seeing is people are started in Horizon 1, they are slowly moving to Horizon 2. Horizon 3 is being done by more startups because they have no legacy, they can just move, they can build knowledge graph from day one. They don't need to worry about multiple systems.

But that's not true for large enterprises and how they start to treat data. So almost data should be almost treated as a asset now. And this data includes unstructured data. This includes structured data. I think structured data. Most clients have some degree of control. Unstructured data is where they have. So over the last 15, 20 years people now have emails, they have chat, they have voice, they may have some photographs, they may have some videos. All of this data is very good data, but it needs to be managed so that you can extract the intelligence out of that data. And that is most organization, at least I have not seen. They're ready to really consume all of that data. We have seen.

If you see databricks is doing very well because that's where clients are now using Databricks as one of the providers to modernize their whole data ecosystem. They're one of the stocks which is not going down while others are.

Matt Pacheco
So hyperscalers like AWS are now publicly celebrating partners like Databricks and Snowflake as specialists they used to resist. What does that shift tell you about the cloud market in general and the AR market in general?

Sumit Taneja
I think I would say it's just the maturity of the ecosystem and how the hyperscaler want to play this long game. They want to be the core infrastructure layer that people are using and building on. Because if they are the core layer, nobody can take them away. And for that, if they have to give a little bit of control to players like a databricks and Snowflake and let them shine, I think it's okay. I think the second most important, at least in my view is AWS has X resources, right? They have to put those resources to something which they will extract the maximum value if the data modernization and that element, if there are companies like Databricks and Snowflake doing better stuff because that's bread and butter for them, let them become the specialists in this area.

Let AWS worry about other things which are becoming more important. So for example, the whole agent ecosystem, AWS was a little behind, but they had to catch up to match up to what was happening in the market. I think they had to divert all the resources on the bedrock, the agent core and the whole ecosystem rather than worrying about their redshift or data lake, which they also have. And I'm sure they will continue to mature it, but they can only invest in as many number of services. They can't do everything in house. So I think it's almost like any other ecosystem, whether it was Apple's ecosystem, whether it was Android's ecosystem, whether it is now these cloud providers, the rest of the almost like app players will always have a play.

And I'm sure the infrastructure players or the base core players, they're going to allow it because that's the model. Because they want to earn both ways, right? So they're going to earn with the infrastructure. And the more data bricks now get consumed in AWS, it's still AWS's money, right? So they win both ways.

Matt Pacheco
So next, let's look ahead at the next few years to see what your thoughts are on some of the trends, some advice, stuff like that. It's my favorite part of the episode where we talk about some of the fun stuff. So you've identified the management of CURATION and curation of data for AI as an underappreciated space. In the past when we've spoken, where do you see that going in the next three to five years? And is there a winner in this space that you potentially see standing above the others?

Sumit Taneja
This, this whole data for AI actually is going through a transformation on their own because now you can use the AI for modernizing data, which was not possible before whole agent Aki became norm in the last six to nine months. Now there are agents which Google has, which databricks has, which Azure has, which actually helps you modernize and get your data platforms modernized. If it used to take three years, it can be done in nine months. Right? So the pace of that change is getting faster. So I would believe, I don't know, I'm sure there might be some new players which will come in, but Palantir is doing well, Databricks is doing well. Snowflake has started to come up the curve as well.

I'm sure Microsoft and aws, their data ecosystems, they may not be paying attention today, but in two years down the line they may also start to build. But I would see any of these platforms, the easier they make to modernize all the data which is hidden in the legacy systems, the more work they are able to do in that area. Those are the company which will become the winners. Because the reality is all top companies have at least 200 systems where the data is hidden. And if you have to extract data into one layer, which is almost real time, that still needs work. I think today most of the work which is happening is still getting to data breaks, but it's a day old data, it's a night time interchange of data.

It needs to get to real time and then that's still a work in progress. I don't think the tech or the companies have reached there, but it'll reach and I'm sure, I don't know, it may happen tomorrow, you don't know.

45:09 - Navigating AI Obsolescence & Staying Current in a Fast-Moving Market

Matt Pacheco
And in there you mentioned. Thank you for that answer. In there you mentioned the pace of change. So a lot of CIOs and IT leadership are kind of caught between this kind of not wanting to be left behind and not wanting to bet on something that becomes obsolete. How do you navigate some of that tension with some of the clients you work with?

Sumit Taneja
And, and obviously that this is one of the reason which is holding off some of the clients to go very aggressive. I'll tell you, the first year when we saw all the gen hype, there were many clients who Wanted to scale in 20, 25 and then came agent AI. And then they said, hold on, let's look at what this is. And now this year we've already seen few innovations like openclaw, claudebot and Claude Cowork and all of that. That starts to create more doubts that we need to maybe hold till things stabilize. What we tell the clients is there are two very important elements. One, when we are building something, keep as much as modularity in your design that you can swap in and swap out. Second, this is still a area that you need to get in to learn.

You cannot wait for three years and then figure out your learnings. This is almost like a generational shift, right? So for your teams to get comfortable, you really need to start and start doing things in production. That's where you start to learn. Because otherwise people will still think this is just a software play, there's a new software, it'll change the world. AI is not like that. When AI gets deployed on day one, it's not perfect, but by day 90 it can get perfect. But there is a journey to it. And a lot of the IT teams are not comfortable with that right now. At least there's no mental model. Their mental model has always been you do a development, you do a UAT and bang, you may have some bugs, but the software works. But that's not how AI works.

AI has a learning curve.

Matt Pacheco
Sticking with that theme and the pace of how things are moving, how do you guys stay on top of the newest releases in Agentic, AI cloud? How do you stay on top of all this and then on top of that transfer, translate it to all of your clients? Because you're constantly having these conversations, you're planning strategy and things are literally changing by the week in some instances. How do you, how do you stay on top of that?

Sumit Taneja
Honestly, you honest answer would be you can't, but you try as much as you can. So obviously what we encourage all our teams is to obviously keep looking at some of the great avenues online. You have, for example, where this guy Peter released Open Claw, which was the original name was claudebot. You could, you could see very when that got released, right, what was happening. You could follow all of that trend, but you need to know and then you need to have ability. In Excel, what we have done is we have created Sandbox and obviously we have an R and D team that they quickly get access. So within couple of days we got them access to claudebot, we got them access to Claude Cowork.

They were able to do play around and they could give more practical stuff to us rather than us just reading what's online. You also see what our teams have done and then you can talk to the client very confidently because we know how it works, we've seen it, our teams, if tomorrow we want to implement some of that, we have confidence.

Matt Pacheco
Very nice. Just two more questions. Agentic AI, what's your prediction? How do you see that changing over the next three to five years?

Sumit Taneja
I think this technology, I'm sure there might be sort of further enhancements to the nomenclature itself. But beyond that I think what we are going to see is maturity of autonomous systems. Now autonomous systems will be in how the businesses are run or how our cars are driven or how our meals are cooked. But there will be more and more autonomous systems which will start to come in our lives. And I think we need to get comfortable on how to extract more value and then keep doing more work. So I'll give you this small example. Not I don't know how if it's the best example, but the number of proposals and pitches we used to make with say a team of five, three years back, just hypothetically was five per week.

Today and this is real, we are doing 50 per week with the same five people. So what I'm also trying to say is that it's not that now I can manage many more client conversations, many more stuff. I can be faster within 24 hours, turnaround stuff and do things. So you will see everything getting fast tracked. Because that's what I don't. While I'm sure there'll be some reduction on some headcount here and there, but you will see economic activity slowly progressing much faster so that people while the pie will start to expand. So your total market if and the if U.S. economy is 30 trillion, that'll become 50 trillion. So your total pie will increase and so your play will also increase. So I would only say that autonomy and how things are going to become autonomous is a given.

I think in every industry the journey and the path will be slower or faster. I'm sure that we are yet to see, but it's going to happen. If you expect your insurance claims to be processed in few seconds, it will happen or it may happen in five years or 10 years, I don't know. But for example, we are already seeing express claims are done online. You don't need to wait. But if you have a injury claim, if right now it takes 30 days,

Matt Pacheco
That may take 24 hours, very interesting, exciting new world. And that's a reoccurring theme we've heard a lot in these conversations, is that the AI agents and the tools themselves are going to augment people. We hear that often. So really interesting last question. What advice would you give to any CEO who's overwhelmed by everything happening in AI and a cloud today?

Sumit Taneja
One piece of advice, I think. Get your hands dirty. Get your teams to start implementing, but have outcomes. What do you want to drive? So you want to drive better customer experience? Go for it. You want to drive better cost ratios? Go for it. Because if you have the larger business objective, then your teams will be aligned to get this done and not treat this as a experiment. But I would never tell them to wait for this technology to stabilize. It's important they start getting to work.

Matt Pacheco
Well, I wanted to thank you, Sumit, for being on the show today. It was a pleasure speaking with you and learning a whole lot from you. So thank you for being on Cloud Current.

Sumit Taneja
Thanks Matt, for having me.

Matt Pacheco
Excellent. And for our listeners, thanks for listening in. This was a fun conversation talking about AI readiness for cloud data and agentic AI. Just so many cool topics that we'll cover in more future episodes as well, because it's an exciting new world. So we appreciate you listening in and we look forward to telling you more stories in the future. You can find us anywhere you get your podcasts. Have a great day. We'll see you. See you again soon.