How FinTech Govern AI

Shamil Malachiyev: Hello everyone and welcome to this week's episode of The AI Floor. My guest today is a Senior Director of AI and Emerging Technology Transformations at Best Egg, now part of Barclays Bank US, where she works directly with the leadership team and the executive board to set enterprise AI strategy. Before that, she launched a credit card platform that originated over 300,000 cards in 20 months and pushed underwriting precision up 40 percent with agentic AI. She's spent 25 years in engineering leadership across fintech, insurance, retail and telecom, and I'm pretty sure we'll have a lot to learn from her today. So please welcome to the studio Mounika Paruchuri. Hi, Mounika.

Mounika Paruchuri: Hi.

Shamil Malachiyev: Can you give us a little background on the current work you do and the experience that shaped you along the journey?

Mounika Paruchuri: Sure. As you introduced, I'm Senior Director of AI Transformation at Best Egg, under Barclays. It's called Barclays Best Egg now; we're getting used to the term. I've been with Best Egg for five years, and before that I was with USAA. My career spans about 24 years, primarily fintech, e-commerce and payments. I started on the AI journey along with everybody in 2024, and I found there was a gap in the company: everybody was doing AI, but there was no defined way to engage with it, and there were pockets of innovation happening. I saw a gap, that we needed to streamline it and put it under a strategy, and that's how I started in this role. I've just finished onboarding Databricks for the company, and we'll talk more about paved pathways as we go. That's a little about me.

Shamil Malachiyev: And you've been busy on the heavily regulated side of the tech industry as well.

Mounika Paruchuri: Yes. I've been working in regulated industries for ten years, and AI just enhanced the need for regulation. So I'm very much involved in AI governance and guardrails too.

Shamil Malachiyev: Doesn't that make your job a lot harder? The more regulation you have in a company regarding AI, the more attention you have to pay to compliance, legal and risk, because you have all the banking data going through your systems. Does that stop innovation, because people are more scared to try? How do you allow innovation to happen in a heavily regulated industry?

Mounika Paruchuri: Very good question. You're 100 percent right that governance and guardrails have become a higher priority in the strategy with AI models in place. People assume that any AI-based product launch is a rocket or something. Essentially it looks more like a kitchen remodel, for lack of a better story, because the inspectors are watching behind you, and you have to finish the remodel the right way. The model is only 20% of the output, and the drift that happens once you deliver and the loops that need to be happening to keep in check for the drift, the harness, and all of that which comes under the governance and guardrails is the most important part.

I want to split those two terms and talk about them. Governance is teaching people what to do, what not to do, how you set your pathways, your logging, your activity tracking, all of that. Guardrails are more about protecting beyond the human. We want our people not to tiptoe around. We want them to feel free and fully empowered to go and implement solutions. So the guardrails are the most important part an AI strategy leader gets involved with, because once they create those paved pathways, it's very easy for people to feel secure enough to go and implement.

Shamil Malachiyev: Can you talk more about the actual guardrails? Take a company with different departments. Your role is to make sure the finance department uses this AI model with a harness and certain tools, but these tools don't interact with HR systems, they don't have access. How do you set up all of these guardrails?

Mounika Paruchuri: Very good question. Before that, let me step back and simplify the language. What we call a harness in AI is essentially integrated unit testing, your runway testing, in older software terminology. We are constantly evaluating the decisions made by the AI to be fitting certain criteria. As you described, a harness means different things in different domains. A harness on the business side, for example an underwriting model, is also about our proprietary knowledge staying in-house and not getting leaked out. Whereas an HR system is more about PII; I don't want personal information getting into a model.

So there are different aspects of what we want to govern and guardrail against, but there are certain common pathways. Would we want PII to go into the underwriting model? No. That's a common path. So there are layers of common guardrails and harness testing, and there is specific testing and harnessing for each. Usually we do this using gateways. Whatever platform you pick, you have to have the guardrails, which are the gateways by which we protect them. That's why I separated it from governance. Governance is also a lot about auditing: SOC compliance, and now the NIST AI RMF. We're still getting used to it. But these are two different pillars of AI.

Let me step one more step back and explain. When I started on the AI journey, I started with the concept of a four-pillar enablement: education, enablement, governance and guardrails. Why is this important? I want to tell the story of a front desk. I've explained it to my team several times, and I think it's a beautiful analogy. There's a building, and you need security. Are we going to implement security per floor, with everybody determining their own? Or do we make the entire building secure, with a front desk, a security guard and a badge system at the front, uniform across the floors? Compare these two. If you enable it per floor, there might be advantages, but one floor has a badge system, one floor is manual, they just look at the person and let them in, and another has a QR code. Every system is different. Now imagine the confusion for an auditor entering the building and trying to understand what's going on with security.

Now imagine the paved pathway I talk about, the central desk, allowing people into the building right at the door. People don't have to figure out each floor. They walk in, they keep going, they know what they should and shouldn't do, where they're allowed. And that auditor is a very happy person. He knows exactly what's going on. Same for companies. It may sound simple, but if there's no central behaviour, central education, central enablement, and governance and guardrails, the confusion afterwards is massive. Then you have to pull the brakes, stop things, go back, reconfigure and enter again. That's how I'd explain the AI strategy overall.

Shamil Malachiyev: I love simple. Whenever I explain something to my employees, the first job is to create a picture that's simple to understand. When you really understand a topic, you should be able to put it into simple terms for other people to start the basic neural connections before building on top. In the US right now there's a huge battle: do we use open source and host it locally, or do we use Anthropic, OpenAI, one of those models? Then we have all these Chinese models coming up, cheaper and quite powerful. How do you make the decision? Token economics, but also security and convenience. Do you think the future is open source, or do Anthropic and OpenAI have their place in the enterprise world?

Mounika Paruchuri: I think the models are not the issue. What you're asking is absolutely correct, but I don't think the answer is a straightforward this or that. I'm a strong believer that one size never fits all, even in software. We cannot build one solution that's uniform across everything. I've seen many, many projects fail with that. Here it might be one switch on and off, but there it might need four steps: get in, check the plug, and then maybe there's a dimmer. It's never going to be one size fits all.

The model we use really depends on the use case. If I'm putting in a solution that answers help-desk queries with AI, because somebody doesn't want to click through four pages, do I need the most secure model hosted in my environment? No. I can use an open-source model, feed it a document and let the RAG search happen. On the other side, underwriting is absolutely proprietary information which I don't want leaked anywhere in the cloud. Am I going to put it in an open-source model? No, never. It would kill the business. What are we going to do? We deploy models within our VPC, most secure, private-linked, everything hosted by us, everything in my network, under security, everything encrypted. This is getting a little technical, but you get the point.

Will Claude be the answer? Will these wars do more harm than good? It doesn't matter, because if you build the systems right and you can plug a model in and out easily, you're fine. To me, 80 percent of the industry is not training models. Unless there's a specific need and a specific model you're training on your own data from scratch, and that's only 20 percent of use cases. 80% of the use cases are prompt-based, fine-tuning a model to work with the data, and that doesn't need so much elaborate thinking process on which model to pick. As long as you know how to swap, you can always do it. In the old days we called it refactoring. I think it's still called a refactor with AI. You just refactor to a new model. We don't have to worry so much about it.

Shamil Malachiyev: Do you feel a lot of difference when switching models? Is there a real felt difference? On a personal level, when you had Opus 4.8 and did some jobs with it, and then Fable 5 came out and you tried it for a couple of days, you found it really hard to go back to Opus 4.8, because you felt the relaxation of being able to trust the model to do a lot of work correctly. Do you feel that when you switch models within the company, within functions? Or is it all token economics: good enough is good enough across the layer, let's push it?

Mounika Paruchuri: You're touching the core. You're running a business at the end of the day. A human picking a model and interacting with it is very different from an enterprise running a model in a production system. You're absolutely right about unit economics. Is there a need for me to pick a new model? There's rapid improvement, but that's what the harness and the looping are about. We've built it and baked it into the system from the start. We're always evaluating the model against all the criteria, and there's a human checking in. As long as that's doing well, there can be prototyping on new models to check whether the output is worth switching the older model, and the unit economics of it. Personal productivity and personal model usage are very different from a model deployed in a production system and the unit economics of it. We're not going to switch models every day. It has to be a planned development, and it has to be proven worth the switch.

Shamil Malachiyev: When thinking long term, in software engineering, now that we're at Fable 5, I'm asking myself: this is sufficiently good. Do we need Fable 6, Fable 7? Probably not. It's already doing a fairly good job. Right now we've found models that do 90 percent of the job really well, so we don't need to push for the more advanced ones. Or maybe I'm mistaken. Maybe we're waiting for prices to slowly go down so we get more benefit. Or do you think we'll constantly be chasing the more powerful model that needs fewer guardrails, and switch every year? How do you plan for that long term?

Mounika Paruchuri: I look at it like a human learning a system. In AI, getting a little technical, there's supervised learning, unsupervised learning, all this terminology. The models are learning from themselves. I think it's just like a human. A human becomes stagnant if they're not learning something new or challenging themselves. I envision a time when AI corrects itself in thinking time. Previously, when we did software, it was static. We'd generate code, it would run in the system, and you knew exactly how much time it would take from start to finish. You could measure in milliseconds how long a process takes to respond. You can't do that anymore in AI. An AI model is now almost replacing a little of that human knowledge: how can I keep improving myself, how can I learn better?

So yes, we do need improved models that learn on top of themselves. If they become stagnant, we become stagnant as humans too, because every time an AI challenges me, I have to think higher than the AI. And where is all this learning coming from? From a human. We're supervising the systems, saying this answer is good, this answer is not good, redo this because it's not the right way. It's taking that feedback and relearning. The intelligence is learning. So of course we're going to need better models, and it's going to be an evolution. But that evolution is forcing every human to be higher than the next evaluation happening in the system. It's a continuous process.

Shamil Malachiyev: So it's never going to stop. We're going to have to keep learning, and learn to live with more and more powerful models. From your experience, when AI came into play everybody panicked: AI is going to do everything for everyone. Then we realised, maybe not everything. It's good at certain points, not very good at others. Looking at different functions, and I think it's important for a lot of people listening to hear the sobering truth: which functions require us to evolve, to step back and think, technology is in, what do I do differently now? Which functions, in your experience, are undergoing the biggest change and need to adapt?

Mounika Paruchuri: I heard your question in two or three parts. Let me first answer the psychology part. For any advancement in technology, there have always been the forefront, the middle and the resisters. When AI came out, people said, this is just another autocomplete, this isn't going to change my entire world. So it was somewhere between this is a chatbot and this is science fiction. That's the psychology, and it depends where you are in that range.

Then the second part. Now people are interested, so they have to rapidly learn and understand how to use AI. Everybody was fumbling at the start, including me. Today I have my failures. I also learned that what works for engineering pipelines would not work for AI. That was a learning. I have to create pathways so people can do better. That was a learning. As the steps went on and I was learning on the job, I realised there are certain things that should be left to AI, and certain things that can never be left to AI. That was the best part of being somewhere between the middle and the forerunners, which is where I'd categorise myself.

So I created a North Star for my company strategy. One was: 90 percent of code documentation, give it to AI. Why? Because it's already recorded, it's concrete data. It can read code, it knows exactly what it is, it can produce 90 to 100 percent of the documentation with accuracy. We don't need to recheck it. It's coming from concrete human data. The second directive I gave was testing. Again, there's concrete data, concrete code, concrete everything. Should we be spending time writing test cases, or should we be telling it what we're testing? Previously I had my developers spend so much time on code quality. I outsourced it to the AI, because again it comes down to what concrete data you're going off and how much drift is possible. It's testing code that's already in. AI is even going to find you the outliers the humans didn't think of, because they were already in the code and in the context and couldn't switch context enough. There you go: 70 to 80 percent of testing is by AI. And now coding. You see the theme. Anything coming from concrete data is always going to get you the best results.

If there's data that needs to be judged, put through a transformer model or a visual model that translates it, extracts the information, we need a human there to check that the right judgment is happening. We're not going to judge every one; there's a harness. But you get the point that we're going to judge the quality. So to answer where I would use AI: where I know there's quality of data, concrete data, and a result from it. We're not trying to replace human judgment and intelligence, but trying to make it easy for humans to do the right judgment and use their capabilities for better outcomes rather than the mundane and the automation.

Shamil Malachiyev: That's why all the companies are running around with data, data, we need to prepare our data, standardise it, go from paper to scanned copies, OCR, structured data. Which brings us to Databricks, one of the latest initiatives you've run. Can you tell us about the thinking behind it, and your experience actually integrating it?

Mounika Paruchuri: Databricks has evolved into a leading platform for AI, and I recognised that because of the automatic enablement of various AI features and models. It creates the paved pathway I was talking about. In the industry right now there are cloud providers, and of course we use Amazon a lot: Bedrock, Bedrock guardrails and pathways. Then I started to put in Databricks, and there's Snowflake Cortex, Snowflake AI intelligence. But all of these tools, and I'm not just going to talk about Databricks, what do they bring to the table? They make prototyping faster. Here I want to introduce two metrics: cost to value and cost to market. I started using these about a year ago. A prototype takes some time to develop, and that's the cost to value. Cost to market is the prototype deployed in production at production-grade scale, and the unit economics of tokens. How does that scale? I had to create two metrics for judging a prototype. First, prove it. Don't build an entire engineering pipeline for a POC.

Databricks enabled that faster for me. There are features in Amazon Bedrock and Q, which Amazon has now come up with, that make it easier. But we wanted to deploy a lot of data science models in-house, and Databricks made that easy, hence we chose Databricks. It took away the hard work of setting up pipelines, CI/CD pipelines, just to prototype an idea. I went from two weeks to two days for a prototype. That's a massive unlock. It also created paved pathways of governance and guardrails. I created guidance documents so people can come in, prototype, do step one, two, three, four. Then: now you tell me how the unit economics are going to work in production, let's measure against this one, two, three, four. Now let's talk about how the looping is going to happen, because AI development is never done. It's a constant loop. You put a prompt in, check the harness, loop it back, check your quality. So how are we going to loop this, and these are the steps you provide on the quality. Then let's talk as a team about whether this prototype makes sense in production.

So it went from two weeks to get a pipeline in, one more week for the prototype, three weeks, then unit economics decision-making, another two to three weeks with the harness and looping. We cut project time from three months to a month. It made so much sense to deploy Databricks as our platform. But Databricks is a leading one, not the only one. What I want to emphasise to leaders is not the tool but the strategy behind it. What you're looking at is how easily your teams can do a prototype. What is your cost to value? Then your cost to market is how easily they understand what they need to do to set up the harness, the gateways, the guardrails, governance, and loop back into human judgment. Then you have production pathways, paved pathways to production for AI.

Shamil Malachiyev: Two further questions. First, the human quality gate after the loop completes. In maturity, what does that look like? Do you have a predetermined document that should be generated, or a piece of software, and somebody sits down and checks it's still as good as it should be? How often do you do the quality checks?

Mounika Paruchuri: This is going to be part of every project. I'll give the example of the RAG bot, not a chatbot, a RAG bot, for the help desk. That's going to be user feedback. You can put buttons on the UI and let users tell you whether they like it. Simple. You run a query every week and see if things are derailing and the value is gone. But that's not a customer-facing bot. It's an internal customer, and you're monitoring it, but it's not the highest priority where a human has to sit and watch.

Shamil Malachiyev: Everybody's putting thumbs down.

Mounika Paruchuri: Now take a customer feedback prompt engine, looking at customers' feedback on the responses we give back. If I'm hearing customers come back and complain a lot more, of course there are buttons, but there's also feedback text, and we're summarising it. That's something I'll watch more closely, and I want a human in there every day, watching, analysing, seeing something derail in real time. The minute they see something derail, they engage the development team again. The people engaged in it are the agents or customer service, who are close to the problem. It's not one and done. The people looking at the feedback are separate from the people working on it, so there's context of judgment happening, and we're not blind to everything because one human is doing everything end to end. That's how I'd go about it, and that's how we've implemented it.

Shamil Malachiyev: A lot of people will remember the very early days of ChatGPT, when you'd get two variations of the same answer and it would ask you to pick which one you liked better. I just checked Claude, and it still has good response, bad response at the bottom of every answer. These are the triggers they use to see whether the new model performs better. I think that's easier for people to relate to.

Mounika Paruchuri: That's the essential part. Going back to your question about why models keep evolving: they collect all that data and improve. The same as what we're doing for our enterprise projects. We're collecting the data, monitoring the emotion the customer puts in the chat, and recalibrating what our answers are and what should happen in that situation. The prompt gets enhanced to reflect that. That's the looping.

Shamil Malachiyev: On a related topic: you mentioned two weeks to prototype has become two days. I feel that with my teams as well. Everything is accelerating so much, I feel time is moving seven times faster. Within one day you get something that half a year ago took a week. But the mental capacity, the biology of the body, doesn't evolve that fast, to contain that much information and make that many decisions within two days. When you're moving seven times faster, how do you avoid getting drained by the end of the day and burnt out by the number of decisions? With two weeks, you think about it, you let your brain do the work, map the neurons. Two weeks is a great time to get your head around a full concept. Within two days, a full prototype? What life hacks do you use to stay sane?

Mounika Paruchuri: The burnout is real and you see it everywhere in the industry. Everybody is either in love with AI when they start, and hating AI by the time they get to the end of a project. It's a natural cycle. There are days when I'm just so burned out, so tired, because that loop failed and I've been at it for many hours, and I'm like, to hell with it, I'm walking away now. Let me explain with a story. Previously, if I was creating a document, I'd open a blank document and start putting my mental information in. We'd do Google searches, validate. What would take me a week is now a prompt I put into the AI. I tell it what I know, I say go look at this, go look at that, and in 20 minutes I have three paragraphs of dump that happened at me from the AI, and I have to go through all of it. Is it burning my neural pathways? Yes, 100 percent. At the end of the day, am I tired? Yes.

It comes back to the fact that AI is unlocking productivity. But at the tail end, when that productivity is unlocked, how are we managing that time unlock for the human to do better? A human isn't a machine that can keep going. So the fact that we have to step away and do more physical activities, things that take away the exhaustion, is more important than before. And I think that should be the human story we write: AI unlocked my productivity, so now I'm spending less time at my computer and enjoying my evenings with my family. I really hope that's where we lead, instead of AI replaced 20 percent of the workforce and I just need the other 80 percent to do the job. That's probably not the best way to write our history.

Shamil Malachiyev: As long as companies stay profitable. Look at Europe: a year ago I was hearing about four-hour work days, and I wondered what people would do with the extra days. Now I can see that if we increase productivity within some functions three or four times, maybe it makes sense to work four days but still do three times as much. It's a bit worrying when you hear technical leaders describe what the future might look like. I think Elon Musk was saying we'll have lots of commodities, everything at very cheap prices because of competition, so you don't need to make much money, with robots producing 24/7. What's your view of how this looks in 10 or 15 years?

Mounika Paruchuri: The world has always been shifting around us, and it will continue to shift. There's positivity and negativity that can come out of it. What we identify with, and what we do about it, is what matters. If I'm the leader and I have six people on a project, and they can do that work in four hours, the next four hours I'm going to ask them to do something creative, something they want to do, something they'll do because they love it. This could be a bigger unlock for companies and projects. When I worked at Southwest Airlines, I loved the culture, and there was something called ten-percent projects. Ten percent of the time, people can do whatever they want, even within the day. I had a colleague who was building airplanes in his garage, and he'd use that ten percent to build.

Shamil Malachiyev: There must have been a big garage.

Mounika Paruchuri: And I had another colleague doing some kind of tweaking on plumbing. I've never seen any company embrace it like that, and that's why there's so much raving about Southwest Airlines. But you see how they're doing innovative, creative work, which actually makes them better at the job. It's very hard to measure that sort of productivity. When I took that to my team and told them they could do whatever they wanted for ten percent, though I told them to stick to software, not garages and airplanes, they produced amazing results. Somebody automated the entire Selenium testing, and I'm talking five years ago, pre-AI. Somebody solved how to help internal customers look up information sitting in multiple systems. What I'm getting at is that what Elon Musk and everybody is talking about has a good part to the story: automation is going to get better and the mundane things are going to be done better. So a human is unlocked, in time and energy, to do more creative work. We're going to keep innovating, because that's what humans do. If you have a paycheck and time, and all you have to do is something to keep yourself happy, we're going to unlock so much.

Shamil Malachiyev: So you think we'll have a lot more hackathons, basically?

Mounika Paruchuri: 100 percent. The jobs are going to get very interesting. It won't apply to every industry the same way, maybe not a shop or a call centre. But as much as innovation is happening, the story can go two ways, one negative and one positive. What we as leaders want to enhance, what we put our time and energy into, is going to decide the better part of the story.

Shamil Malachiyev: So we should aim for the utopian version rather than reading all the dystopian ones. Say somebody is in their job right now, and their manager or CEO says: everybody's adopting AI and getting three times the productivity. We heard the podcast with Mounika, and she says everybody should prototype within two days and deliver a project within a month. We want that at our company too. Show me some ROI. What would you advise people in positions of responsibility, with teams, who have to deliver on that? You mentioned the four pillars. Is there something extra to fight the resistance? Promote AI months within the company?

Mounika Paruchuri: The ROI unlock in AI is a problem because people are trying to insert AI into places where it doesn't belong. There's huge unlock, and I want to explain with another story. When Steve Jobs was first creating a hard disk for the Apple, he assigned a few designers to check how to make it better and simpler. They were working off a huge document, 30 pages, and designed a solution that was five steps. They were so proud, they walked into the meeting, and he listened for 15 minutes and said, no, we're not going to do this. He went to the board, drew a box, and said, I want something where people can pull this onto this box and they're done.

You see the strategy. He wasn't trying to take something as it is and reduce it. You have to rethink. So why is the ROI unlock so hard in AI? People want to take AI and insert it into something that's already complex. It's not going to work that way. We have to rethink the entire project, the entire process, the steps, in terms of AI. If a human was looking at call-centre feedback and just saying yes or no, maybe you can still insert it. But there might be complicated steps where you need to redo the entire process and do the mapping first, before you ever build it in AI. That unlock is very hard for many, many people, and that's what's happening in the industry.

And to any CEO I walk into right now, the biggest unlock of ROI is protecting yourself against cybersecurity. With Fable and everything, there's a lot of chance your systems are going to get attacked. That's an easy ROI I can sell to any CEO: we need to protect our systems, so let's build these models that check in, better firewalls, better bot detection. That's an easy ROI. But if we're trying to insert AI into some system and make it automated and reduce the burden, that's different. This is not RPA. This is not robotic process automation. This is an intelligence. So you have to rethink that entire process in terms of an intelligence sitting in it. That's what people are just now wrapping their heads around.

Shamil Malachiyev: A year ago we got ISO 27001 certification, and right now we're going through SOC 2. When talking to auditors, I ask them about all these things: we have an open AI model, just take it from Hugging Face, it doesn't make sense. We're trying to protect against something that can look at the system, see all twenty loopholes and inject itself. Is the whole security space going to evolve into something completely different? I don't have an answer. Maybe you have some guidance.

Mounika Paruchuri: I don't have an answer either, because how deep are you going to look into your cybersecurity? I might be using a vendor who's using another vendor, and the vulnerability sits in their code or their system. How deep do you go? It's a very hard space right now. I think that space of AI is going to be the biggest unlock of ROI right now, because we have to protect our systems. That's the nightmare keeping all the CEOs up at night, to be honest. But I think there's also enough regulation kicking in just for that, because we have to regulate some parts of AI in terms of the damage it could produce if we don't put in the guardrails. So I don't have an answer. A lot of enterprises are trying to do the best firewalls, the bot detection, and it's constantly evolving. Are we under threat? Yes. Are we under shutdown? No. You're going to sit on the vulnerability and the legal language provided by your vendors, and it's a he-did, she-did situation. But that's what it is right now.

Shamil Malachiyev: To me it sometimes seems like fifteen years ago, when everybody was hosting on-prem, then everybody switched to cloud computing. Now I'm seeing more and more people talk about hosting on-prem, with the new initiative from OpenAI and Anthropic about forward-deployed engineers within companies. You wouldn't need forward-deployed engineers if you weren't planning something on-prem. It seems like everybody is leaning towards their own AI hubs, separate rooms in the data centre. Is that the world we're going towards, and should people be thinking about where they'll host all those models, and getting people to live next to the kill switch?

Mounika Paruchuri: In highly regulated industries, yes, it's already happening. Look at any big bank. I saw this trend a year and a half ago and made a comment. There was this funny thing when OpenClaw came out and everybody was buying the Mac minis. I was laughing: okay, we're back to on-prem now. Somebody said that's not so true, and I said, okay. But I saw it happen, and it is happening in how proprietary models are being hosted for the highly regulated industries. I mentioned that even our proprietary models are cut away from access for anybody, even though they're in the cloud, on our VPC, on our premises. Big banks have big pockets. They can deploy total data centres. So I agree with you that the trend is going to be people wanting more control. If they're developing proprietary models from scratch, supervised or unsupervised learning, they won't want that exposed. So on-prem is going to happen, but it might not be on-prem as in physical hardware. Funny story: at my first job I stepped on something and pulled out a whole server for my startup.

Shamil Malachiyev: No.

Mounika Paruchuri: We're not going back to that. We're definitely doing some mix between cloud-hosted systems that are in your control, in your proprietary shell, secure from external hacks and attacks. That's where we're going.

Shamil Malachiyev: That goes to what you were saying about buildings. You put one good security on top of the whole building, and I think that's what Amazon is doing: taking very good security and putting it at the door, so everybody doesn't have to worry about securing something themselves.

Mounika Paruchuri: That's how Amazon was even with cloud. Forget AI. That's what Amazon was very good at, security that goes into every bit of information encrypted in and out. Most of the cloud providers.

Shamil Malachiyev: As a final touch, anything you want to advise the listeners? What should they be doing to secure a good understanding and a good place in the future being built right now?

Mounika Paruchuri: Humans are going to evolve with technology. For our generation, if I start talking about steam engines and electricity, it won't make sense. So let's talk about going from retail to online e-commerce, the internet. That's more relatable. If you were somebody in 2000 looking at that, and you're in 2025 looking at AI, the pace is even higher in the current years. The pace in 2000 was a little softer; you had time to react. I mention that because we have to evolve as humans. There are resisters who'll see all the negativity coming out of AI and say it's just another autocomplete, it's reading off the data, it can only produce what another human already documented, so how is it different? And there are people who'll do everything and anything in AI. In between there's a huge range, and for most people I'd want them to be somewhere in that range leading towards the front, because this is going to alter our lives. Getting on the bandwagon and understanding what it can and cannot do for us is going to be very wise learning to stay current.

And it's not about age. I've heard a lot that younger folk are better because they come with the learning of AI and understand everything faster, and seniors are in the resister category. No, that's so untrue. It really depends on how well a person can learn and how dynamic they are about change. That's what it boils down to. Change has always been the constant thing in our lives, and our neural pathways are always built against it. So on that spectrum I'd advise people to be towards the leading part, so they can be ahead and understand where they'll be at a disadvantage with an AI product or AI enhancement coming into their lives, versus where they'll be at an advantage. We all have to learn it, just like we learned to work with computers. That's my advice: don't resist, but you don't have to be the forerunner and put everything in AI. You can be in the middle, observe, learn, stay up to date. This is going to change our lives. Get mentally acclimatised that this is going to change. We have to learn, we have to be smarter than the system. AI is pushing a human, whether they want it or not, to be smarter than it.

Shamil Malachiyev: Well, you've heard it. There's no stopping it. Time to adapt, time to change and evolve. Mounika, I want to thank you so much for joining us in the studio today. Thank you.

Mounika Paruchuri: My pleasure, Shamil. Thank you very much for having me here.