Transforming Enterprise Delivery with AI, Cloud, and Agentic Systems

Hi, and welcome to the Data Science Salon podcast, which is now the AI Loves Data podcast. If you're returning, welcome back and thank you for being here again. And if you're listening for the first time, we're glad to have you here. Thank you for choosing the Data Science Now AI Loves Data podcast.

We know there are many choices out there, and I'm Anna Aniston, the founder and host of Formulated by the Home of AI Loves Data. And joining us today is Jojit Roy, a technology and program management leader with more than two decades of experience driving enterprise transformation, cloud modernization, and applied AI and machine learning initiatives across really interesting industries like insurance, financial services, and e-commerce. And he currently serves as a lead principal technical program manager at USAA, where he leads a multi-year modernization program focused on transforming legacy PNC insurance platforms with cloud-native architecture and machine learning-enabled capabilities.

And Joji brings a unique mix of program leadership, technical depth, and AI execution, and his work spans across cloud-native modernization, ML Ops, intelligent automation, and agentic AI. And he's also spoken publicly on how agentic systems can move beyond rule-based workflows to enable adaptive planning, contextual memory, and closed-lube decisioning at enterprise scale. And in this episode, we're going to dive into what it really takes to operationalize AI inside a large organization, from modernization, roadmaps, and delivery governance to real-time decisioning, MLOps, and the emerging role of now-hot, agentic AI in the enterprise systems.

It's such a pleasure to have you here with us today, Joji. Thank you for having me, Anna. I'm excited to be here. Let's go ahead and dive in.

All right. Well, let's start. Why don't you go ahead and share a little bit about your journey into technology leadership and what led you to focus on, you know, the intersection of AI and machine learning and cloud, specifically cloud modernization and the enterprise landscape? Yeah, absolutely.

So it's been more than like two decades for me. I started my career as a software developer and over time I moved into leadership roles, driving large scale enterprise programs. So my experience has spanned multiple domains. I spent close to nine years at TEL working on large e-commerce and digital platform organization.

In recent years, my focus has shifted more towards regulated industries like financial service and insurance, where systems are complex and the stakes are high. So across roles at TEL and now at USA, I have been leading large modernization initiatives. Some of these are actually multi-year programs, which have a significant scale and investments. So one pattern I keep seeing was the same challenge.

So legacy system, slow down innovation, fragmented data, and a clear gap between business intent and execution. So that pushed me into cloud modernizations, moving from a monolithic architecture to cloud native systems. But even then, decision making was still very manual. So that's where AI becomes a central of my work.

not just as like models, but as a system that can actually assist or automate decisions within like enterprise workflow. So more recently, my focus has evolved or shifted into Agent TKI. So I see autonomous agents as one of the most practical way to bring AI into real business workflow, or better say like in real world. So systems that can reason, adapt, and take actions.

All we need is this kind of systems, and Agent TKI offers that kind of force. So overall, it has been a progression from building systems to transforming platforms to now enabling intelligence, autonomous systems at scale. And that's my journey, Anna. No, that's a really cool journey.

Thank you so much for sharing, Jajit. And you've led large scale transformation programs across insurance, e-commerce and enterprise platforms. When you step into a legacy modernization effort, how do you decide where AI can actually create the most value when you go in? I look at the data readiness, not just whether data exists, but whether it's usable, consistent, and accessible in real time.

And that is because without that, even the best models don't add values. I focus on business impacts. And this, I believe, is the most important consideration. So I try to identify use cases that are directly tied to outcomes like revenue, like risk reductions or operational efficiency, etc.

So, for example, in insurance, that could be things like real time risk assessment or personalized quoting. Say in e-commerce, it could be dynamic pricing or recommendation systems. So overall, my approach is to start with decision points. So validate data readiness and then anchor everything to clear business value before introducing AI.

That's a great answer. You've been involved in modernizing a major PNC insurance platform while integrating machine learning and AI into areas like quoting and decisioning. What are the biggest challenges in bringing AI into core enterprise workflows like that? Because without disrupting delivery velocity or governance, because I'm sure there's a lot of regulations around this at these kinds of huge organizations.

Yeah. So this is probably one of the hardest part of enterprise AI So from my role perspective I closely involved in program where AI capabilities are being introduced into core workflows So what I have seen is that the biggest challenge is balancing speed with governance. So in regulated environment like USA, governance is not optional. You have to get very close to zero risk tolerance with explainability, auditability, and strong building controls, particularly in regulated industry.

And that obviously slows things down. And then the second challenge is data, right? So most legal systems have fragmented or inconsistent data. So a lot of the efforts goes into data refinement before AI can actually be effective, right, or kicks in.

So the third challenge is integration into real workflow. And this is probably the most important one, I believe. So AI cannot sit as a separate layer. It has to be embedded into process like quoting or underwriting, right?

And it has to work in real time without disrupting delivery. So and finally, the operational model. So you need tight coordination between engineering, data science and business teams, along with strong ML ops or ML operation practices to keep things stable in a production. So overall, if you ask me, it is really a balancing act.

You are introducing intelligence into critical workflow while still actually maintaining delivery velocity and a strong governance. So that is important. No, that's a great answer as well. And you've spoken about Agenda KI and you just spoke actually at our conference in Austin too.

and you talk about it as a shift beyond static rule-based automation. For listeners who are still kind of figuring out, which I think a lot of people are, what it means in practice, how would you define Agenda KI and specifically an enterprise setting? And where do you see the strongest real world use cases today? Yeah, that's a great question.

So because the term Agenda KI gets used a lot, but it is not clear, all is clear to everyone, right? so the way I define agent AKI is an enterprise setting is this instead of system that just generate insights or predictions you have systems that can actually take actions that is agent AKI for you say an agent can understand context make decisions within defined boundaries or guardrails and execute tasks as part of the business workflow so that's an agent AKI so it's really a shift from passive AI to more goal-driven autonomous systems right so one interesting pattern I see many organizations start their AI journey at the front end things like you know chatbot or conversation conversational AI right both those areas often don't have enough structured or reliable data because it is very new right so the impact can be limited where agency care really starts to show strong value is in the back-end workforce, right?

Where we have rich historical and structured data. For example, in insurance, areas like claims processing on underwriting, in e-commerce, things like inventory management, pricing, or supply chain decisions. They have tons of data, years of years of data. So these are multi-step processes with clear decisions points and strong data foundation, which makes them ideal for agent-driven automation.

So hence, in practice, I see agenting AI being most effective, not just as a front-end interface, but as an execution layer embedded more deeply within enterprise systems. And this is where it can actually drive measurable business outcomes. No, that makes a ton of sense. And one thing that stands out in your work is the blend of agile execution with deep technical systems thinking.

How do you translate a long-term enterprise roadmap into actually executable milestones that engineering product and all the other folks in the company can actually align around? Yeah, this is really where execution actually matters in a disciplined way. And I approach it by breaking the problem into layers. So at the top, you have done, you know, long-term vision and business outcomes.

that could be things like revenue growth risk reductions or you know platform organizations all these things and the first steps is translating that into a clear roadmap while well-defined milestones are there i usually anchor this around outcomes not just like features so everyone understand what success looks like it also keeps the conversation more strategy from there we break it down into executable units like epics right in jira say like epics features and then into sprint level work items this is where frameworks like the scaled agile framework or safe really help in you know aligning multiple teams the big part of this job is managing dependencies across different teams right in your program so we make those uh visible early during planning and continuously track them through program level sync up the other important piece is alignment so engineering product and business teams all things differently so communication has to be tailored right say for business stakeholders it is about outcomes and timelines then for engineering it is about architecture and feasibility and for product it's about like you know priorities and customer impact correct so having worked across development architecture product and program leadership I naturally act as a glue here.

So I spend a lot of time connecting business and technical teams to ensure alignments, clear communication, and smooth execution. And finally, I rely heavily on matrices and KPIs, or key performance indicators. For example, things like predictability, velocity, volatility, and delivery health. So to continuously adjust and keep execution on track, we continuously maintain these matrices monitor these matrices and KPS So overall I think it is about taking a high level version and vision and breaking it into structured layers and then driving alignment and execution through a discipline approaches and continuous feedback If you're enjoying my conversation with Joy G, you cannot miss our next event.

It's all about using Gen AI in finance, insurance, and banking. It's going to take place on May 13th in New York at S&P Global Headquarters. Make sure to use promo code ALDNYCPODCAST, all uppercase, to save 30% off your registration. Again, it's ALDNYCPodcast, all uppercase, and you save 30% off your registration.

And I can't wait to see you there on May 13th. Let's talk a little bit about cloud-native architectures, because you've worked with those event-driven systems and a lot of different machine learning and AI pipelines. What are some of the technical building blocks you believe are essential for organizations that want to support real-time AI decisions at not just, you know, mostly at scale? So to support real-time AI decisions at scale, you really need a strong foundation across a few key layers.

In these layers, the first is data. So it is the most important one, I believe. So you need reliable real-time data pipelines. That means streaming architecture, event-driven systems, and making sure data is available clean and consistent.

The second is the model layer itself. So what we often call MLOps or machine learning operations, right? It is not just about building models, but putting them into real production use. Things like model versioning, real-time interfaces, monitoring, and feedback loops become in this layer.

And these are critical for this layer as well. So the third is the platform layer. So you need a cloud-ready foundation that can actually scale, like containers, Kubernetes, and API-driven services. that allow different components to interact flawlessly in this platform there.

Then comes integration. So AI decision has to be embedded in business workflow, right? So APIs, orchestration layers, or even base triggers are the key to take real-time decisions here. And finally, governance and observability.

Very important one, governance and observability. We need strong monitoring, logging, and controls to ensure decisions and traceability. And that can be explained and reliable to the production. So overall, it's actually really a combination of data model, platform integration, and governance working together to enable real-time scalable decisions.

No, that's a great answer. Thank you. And let's talk a little bit about MLOps because it's often discussed very conceptually. But actually, when it comes to putting it into operations inside complex enterprise, it gets really messy.

What have you learned about feature engineering, inference workflows, observability, and continuous improvement that organizations often underestimate? Because there's just a lot of different details you have to think about. A good question and a very valid one. So one of the biggest things organizations underestimate is actually feature engineering.

So it is not just about creating feature once. It's about maintaining consistency between training and inference. So in many cases, the data used to train the model is very different from what is available in real time. So unless you have strong data pipelines and feature management in place, model performance drops quickly in production once you deploy.

So it drift. The second area is inference workflow. In enterprise systems, latency and reliability truly matters. So you cannot have a model that works well offline, but slows down in real time.

So inference has to be closely integrated, optimized, and resilient. And the third is observability. It's very important. A lot of teams deploy models, but do not have a proper monitoring in place.

So you need visibility into model performance, and then data drift, and then decision outcome, biasness, and not just like system matrices. but these things should be measured as part of the observability. And then finally, continuous improvement. So models are not static.

They degrade over time as data and business condition changes. So you need feedback loop, like retraining pipelines, and then governance around when and how models are updated or retrained. So NetNet, the challenge is not building the model, but building the system around it. So that's what most organizations underestimate and do not get quite well.

No, that's a great answer as well. Thank you so much for explaining that. And we know that, you know, you're in really high, especially with insurance and finance, you are in really high regulated industries and trust. And we talked a little bit about governance earlier.

How do you really think about this governance and risk and the human oversight when deploying these systems? Because I'm sure there's a lot of things that could get missed. In regulated industries, trust is absolutely critical. It's a must.

In many cases, it's even more important than speed. The way I think about it is across three layers. First is governance, risk, and human oversight. From a governance perspective, you need strong controls built into the system from the start.

That includes explainability, auditability, and clear decision trustability. So every air-driven decision should be something you can go back and understand. From a risk standpoint, it is about defining clear boundaries. So AI systems should not operate without constants, right?

It should not run wide You need guardrails for which decisions can be automated which requires validation and where the system should escalate right That human oversight becomes important Especially for high important decisions you need a human in the loop model So basically it is not to slow things down, but to ensure accountability and build trust in the system or around the system. So another key aspect is continuous monitoring. monitoring. So you need to watch the model drift bias and unexpected behaviors and have mechanism to intervene quickly when something goes wrong.

We just talked about in previous in a question. So overall it is not just about deploying AI, it is about deploying it with responsibility and it is about making sure it operates within defined boundaries with visibility, control and human accountability. That's a really great answer. Thank you.

And you've led very large distributed teams across product engineering, architecture, and program management, what are some of the leadership habits and practices that help keep these kind of distributed teams, you know, moving without losing clarity or accountability? Yes, a very valid question. So leading large-scale transformation program really comes down to a few consistent practices. So the first is clarity of purpose, I would say.

So if every team should understand not just what they build but why they build it and why it matters right so when people connect to outcomes alignment becomes much easier and the second is no structured execution so i uh rely heavily on planning guidance like we are planning right uh regular sync ups and clear milestones so that creates a rhythm where teams know what expected and when that is expected The third is transparency. So I make sure progress, risk and dependencies are always visible to everyone.

So when teams can see what's happening across the program, it reduces surprises and eventually improves coordination. And the fourth here, I will talk about ownership. So I try to push decisions making as close to the teams as possible while still maintaining accountability at the program level. So that actually, you know, balance and it is important for both speed and control.

And then finally, continuous feedback is very important. So through demos, demos to product managers, matrices and retrospectives, we keep adjusting as we go. So instead of waiting until the end to realize that something actually off the track. So overall, in my opinion, it is about creating clarity, establish execution discipline, and building a culture of transparency and ownership across the program.

At a leadership level, I strongly believe in a certain leadership mindset, right? It's very important for me. My role is to support the teams, remove obstacles, and create an environment where they can actually perform at their best. That's a great answer.

Thank you so much for sharing that. And my last question for you is looking ahead, and I know that we're moving really, really quickly these days. What are you most excited about as the next phase kind of enterprise AI, especially around, you know, agent to agent and intelligent automation and also decision intelligence? What are you most excited about?

Yeah, so what excited me most is the shift from AI being focused on generating insights to actually driving actions. For a long time, AI has been used to generate insights or recommendations. But the next phase is where systems can actually take actions as part of business workflow. And that is where agentic AI systems become really interesting.

So you are moving towards systems that can reason, make decisions within defined boundaries and execute tasks end to end. So I think this will have a big impact on areas like operations, risk management and customer experiences, where there are a lot of multi-step process and decisions points. Also, at the same time, I see decision intelligence becoming more integrated into everyday systems, not as a separate layer, but embedded into how business works. Overall, I strongly believe next phase of enterprise AI is as a combination of great autonomy, deeper integration into workflow and strong guardrails to ensure trust and control.

And I think the organizations that get that balance is right will be the one that truly unlock the value to enter per se. No, that's a really great answer. And I agree with you. And thank you so much for joining us, Jojit.

I really appreciate your perspective on how to bridge strategy delivery and technical innovation in a way that actually drives measurable business outcomes. and your experience leading large-scale modernization while also embedding AI into real enterprise systems offers a valuable roadmap for any organization trying to move from experimentation to operational impact. And for our listeners, if you're working on AI transformation, cloud modernization, MLOps, or Argentix systems, this is an episode you'll want to bookmark and share with your team.

Thank you so much for tuning in, and we'll see you on the next episode of AI Loves Data Podcast. If you enjoyed my conversation with Joy Geet, you will really enjoy our next event coming up in New York at S&P Global Headquarters on May 13th. Please use code ALDNYCPODCAST, all uppercase, to save 30% off your registration. Again, it's ALDNYCPodcast, all uppercase, to save 30% off your registration.

Look forward to seeing you there.