The Truth About AI Adoption In Regulated Industries

Added: 2026-05-31

[00:00:00]Today we're talking about the massive divide between AI in the move fast and break things world and AI in regulated industries like banking. We're going to look at why deploying a large language model for a retailer is fundamentally different from doing it for a financial institution where a hallucination isn't just a glitch, it's a regulatory violation. If you've ever wondered why banks seem slower to adopt the latest AI features, or if you are an engineer trying to push innovation through a wall of compliance, you know exactly what I'm talking about. In this video, I'm going to break down the unique challenges of data engineering and AI engineering in the financial sector. We'll look at the friction between legacy systems and modern AI, the absolute necessity of human in the loop validation and why data quality is the single biggest bottleneck to adoption. Then we will hear directly from practitioners on the ground Anisha and Maxon about how their roles have shifted in this new era. This applies to anyone building systems in healthcare, insurance, government or finance. any domain where accuracy isn't just optional. Deploying AI in these environments is huge for three reasons.

[00:01:18]One, the conflict between probabilistic AI and deterministic regulation. Banks run in deterministic rules. If I transfer $100, exactly $100 must arrive.

[00:01:31]AI models are probabilistic. They guess the next best token. When you mix the two, you risk catastrophic errors like AI misinterpreting a currency code or hallucinating a regulation. You can't trust the model to figure it out. You need very strict guard rails. Two, the legacy data problem. In a tech startup, your data might be 5 years old. In a bank, it could be 50. You have silos, mainframes, and duplicated metadata across dozens of systems. As we will discuss, you can't build a reliable AI agent if it doesn't know which field is the single source of truth. Three, the human in the loop is the most important.

[00:02:14]In unregulated industries, we try to automate the human away. In banking, the human is the regulatory safety net. The goal isn't to replace the engineer. It is to amplify them while keeping them liable for the output. Let me share some context before we dive into the interview. When we talk about AI in regulated industries, we aren't just talking about a smarter chatbot. We're talking about fundamentally changing how we parse terabytes of sensitive transaction data. The challenge is the governance of that capability. You have to deal with PII, personally identifiable information, huge data gravity, and regulators who demand explainability. You cannot explain a blackbox neural network decision to a federal regulator easily. So how do you actually get the value out of AI in this environment? You focus on developer efficiency like using tools like GitHub copilot to write boilerplate code and you use AI as a refinement layer rather than a decision layer. Quick shout out to this video's sponsor, Omnisend. If you're shipping a product, a course, or a newsletter, the hard part is usually not the build. It's what happens after someone shows interest. Most of that intent just evaporates because there is no system catching it. Omnisend is that system. It's email and SMS in one platform. You capture someone when they sign up. Then automations follow up based on what they actually did, opened, clicked, bought, went quiet. instead of you doing it manually. Think event driven workflows for your audience.

[00:03:57]Infrastructure already handled. What makes it worth a look now? They just cut SMS pricing to start at 0.007 per message. That direction is unusual for a SAS. Everyone else raises prices yearly. Lower cost means you can actually use SMS instead of just rationing it in the same flow as your email. If you're already on another tool, their team moves your automations, templates, and lists for you in about five days, free, and people switching pay around 35% cheaper. For context, email averages $79 back for every $1 spent. They're $4.8 out of five on Shopify across 150,000 brands with 24/7 human support on every plan. Check out Omnisand with my link in the description.

[00:04:52]Now, to dig deeper into what this looks like dayto-day, I sat down with senior data engineers Anisha and Maxon. Here is what they had to say about the reality of AI in the financial sector. What are uh some of the things that that changed in how you do data engineering before the AI became so prevalent and after it happened? One advantage what I'm seeing is like the automation of programming uh like okay so the amount of time we need to actually take to produce something uh uh uh to produce something produce a code or develop an application test it that has reduced a load once you have that so that that is definitely helping us to deliver value outcomes faster and sooner that is definitely one key value which we are actually seeing about that but again like everyone is saying like you have mentioned you need to of that personal human oversight on top of that to actually validate that okay what this more I'm using copal github copilot uh in my development environment so once that produce that okay so I need to go through that extra layer of validation as with the human intelligence and with our knowledge to make sure that okay it is actually giving us the correct output before we actually productionize that that is a kind of uh what we are doing but definitely adding value in terms of uh how fast okay so typically for to give an example uh an application which I used to uh which used to take like 3 to 4 weeks previously uh to develop the application, test it and make it production ready. Uh it the development part actually happens within hours like minutes or hours. Okay. And the testing part probably may take a bit longer because we need to go through the dehuman iteration and validation like over that one. But that 4 weeks may reduce for four 4 days in uh in terms of uh producing value outcomes. So that's that's the benefit which we are seeing.

[00:06:43]>> Thanks Max. So I've been a data engineer and now I'm working as a senior data engineer in one of the uh financial sector in Australia. So what I have seen in the last 10 years because I started my career as an application developer before and then I joined the data team because data was something that everyone used to talk about but there was no mention of AI at that point in time. We all learned biology right and we knew how human brain worked. It's I think the similar kind of analogy uh which we can apply in the AI because it's artificial intelligence but end of the day it's like how the humans thinks the agents are thinking alike. So that's basically the AI that's my understanding again there are multiple definitions all around the world. Now how it changed the data engineering thing is basically when I uh studied 10 years back um I had to put a lot of effort into identifying what from where I would learn stuff right there were not many books or not many stuff on Google um that I could really check and learn from right so I had to for example if I think about Python or Java there was oops concept right people can learn it in now 5 minutes right but I had to spend like 2 months to learn what abstraction is what polymorphism is right so I think that had helped a lot so it's basically multiplier effort uh effect or basically an amplification so what had been done or learned within like 6 months you can do it now in 6 days if you have the aptitude for it right if you actually want to know what's happening around you can learn quickly so that's I think the good thing but in terms of the data engineering so I feel um because I am I am in financial sector so in finance I I think it takes a lot of time probably to implement AI perfectly because it's huge data that we are talking about and there are a lot of business concepts as well.

[00:08:31]So I think what's um good is every day I'm using copilot I'm using chat GPD to spec especially um pritify my emails or um pritify what I tell people. I want to I don't want to be aggressive I I want to be more polite. That would help a lot. The AI are helping a lot there. um especially if I'm stuck sometimes I think as a developer we all face challenges someday that okay my brain is not functioning properly how would I do a simple date parse right sometimes it's a very basic thing but we can't do it right as a human but with AI just because it's a machine it can do it quickly right so I want to do a date parse function how do I write it if my brain was not working I wouldn't have been productive for 6 hours but if you see in 6 seconds AI could do that so that's been really helpful so if you know what you are looking for and if you have good ideas if you are creative enough and if you put that like proper prompting then I think AI would do amazing but if you are like lost or do not know where to start from that's where I think AI would struggle as well but I think yeah overall I think we have seen a lot of benefits having AI in data engineering especially doing basic SQL but yeah where it lags currently is the performance optimization especially if you're dealing with like pabytes of data but overall it has increased the efficiency of all the developers.

[00:09:53]>> The I treat the prompting and I as long as you have some idea like you know what you what you want outcome is. So if you do a proper prompting so normally people like some sort of uh programming background and the engineering background. Okay. So they know what the outcome what they are expected. So if you do prompting and once you don't like on first time okay you might not be getting the absolute accurate result. So you do iterative prompting that's what I have learned okay over the few for past few months spent also using uh AI uh leopilot so you do it prompting on that one so that will make the correct refine the result and okay so uh so typically what one mistake what I have seen people doing is okay they prompt it get a result then start working on that manually so if you can avoid that one and do it prompting on that one so you will get to that result much faster than what uh you do in in in that you bring in the manual intervention. So that really helps actually and it can deliver the value faster.

[00:10:51]>> One more thing is um because I've learned C, R, Java, Python, few other programming languages but I think AI does a lot better when it comes to like known languages. So if you like Python, Java, but when it comes to like R or something which is not mainstream, it doesn't do well. So you need to know the language by yourself to make sure whatever it's producing it's the right thing, right? So yeah, that's another thing.

[00:11:17]>> That's a good one. Not many people work in heavily regulated industries. So what are your thoughts on the unique challenges that data engineering teams or data science teams are experiencing when they work at companies that do banking? you know uh what are some of the things that you have to be conscious of that other teams who work in retail like they don't have to think about >> yeah straight away one thing which I can think of is uh the adoption of the tooling and the the security and the guardrails which actually the banks the financial especially the highly regulated sector like financial sector has to go in so the adoption may be a bit slow but within uh a regulated environment when that adoption happens actually things are getting faster but adoption of this tooling and uh especially uh going around like putting that guard rails to stay within the regulations. Uh that is one thing one challenge which I see uh which is slowing down a little bit of uh in in terms of adoption for the financial sectors I believe.

[00:12:19]>> Yeah. So I think I've been working in the high impact process for almost a year or more than that. So we work in like all the banking processes like bereavement, hardship, collection. So it's all related to the non-financial risk sector. We do get a lot of u regulatory issues coming from ASICAPRA right but what I've seen there is basically there are a lot of it's huge right the banking sectors or any financial sectors and we are dealing with a lot of sensitive information compared to the other retail industries right so there um especially we need to incorporate a lot of data policy checks because and you need to know what data you are getting it's right that's the most difficult part in any financial sector We do get a lot of um discrete data right here and there. So if I want to know customer details, there is no one single source of truth. I want to know customer's date of death, customers date of notification. I see multiple systems in a bank. There are like lot of data. So people are predicting people's death. Sometimes the date of notification is usually prior to the date of death which is not possible right. So I think you need to identify the business rules identify all the systems in a bank properly before you do any regulatory stuff because when we submit something to the regulators they do not know what we are submitting right we as an employee of the banks or any financial sector need to be 100% sure the data that you are producing that would be valid and it comes to AI as well right because AI is basically a a layer a very superficial layer I would say because it trusts the data that we are producing so if you do not produce a right data on right time then it would be difficult and the and the I think the consequences would be catastrophic as well right because you do not know what you're producing for a simple example I think people talked about in last session centraliz session as well that the currency conversion right people thought okay but AI thought it's US dollars but it's AUD right so these consequences you wouldn't know at the very beginning so you have to be very proactive in preparing the data and making sure the data quality is almost accurate. So especially with the ASIC and approx.

[00:14:29]>> Yeah, thanks. Thanks for that spot on Andesa and I just wanted to add one more to that one. So just try to remember like um one main thing which uh will block uh or come in in the way is actually like metadata is everything and in a large organization like a financial organization like bank which we have which we are working or in a huge organization which which has a lot of legacy there there are lot of systems which the data will be there but the metadata will not be available. So I think to run fast with AI I think that's the first step which in my opinion in my view which everyone should be starting because we as long as we start creating like collecting that metadata consolidating that and if we can actually get the unified metadata store that's when probably one um we can go faster with AI. Mhm.

[00:15:19]>> Yeah. And other thing is another point which I should touch upon the same information available in multiple systems. So that is actually going to create a bit of confusion and chaos like okay so reversals. So if you go to actual uh actual implementation in a large frame this is pretty common. So once that consolidation happens okay so it is uh a common system okay where like if you go with the engineering pattern if you go with the standard architecture like okay sourcing staging layering or consumption and kind of thing okay so that so that's a pattern which everyone follows but that pattern is replicated in multiple areas of the of a huge large organization that can create confusion or confusion and bit of chaos as well.

[00:16:04]So if you can actually reduce that one probably that is also going to help is my uh my thingy.

[00:16:09]>> Yeah, I think it's spot on as well. You mentioned a lot about being more limited in your choice of tools having a lag as the industry is moving forward. Do you see in your teams or uh in in your environment the leadership does it uh support and understand where things are going or do you have to do a lot of convincing yourself? Personally in the area which I am working in I am seeing actually a lot of adoption lot of uh encouragement to use tools and all uh so individual independently in a development environment we can actually try out all these things or in from our local system we can clust with the test datas and stuff like that but when it comes to productionizing the actual payload that is when we will see challenges so so that is where I think leaders are actually on top of that where especially in the organization which we are actually coming from uh there's a bit of encouragement in in to go in that direction. Um hopefully you'll get there one day.

[00:17:07]>> I think that's a very tricky question.

[00:17:09]Um because in our area um where we work as a central data team helping retail and the commercial customers I've seen a lot of adoption in the leaders leadership like they do um appreciate if you do something new if you try something new. But the problem as Maxon told right productionizing it and doing it with the real data because in the financial sectors you're dealing with the customer numbers, account numbers, right? These are all sensitive data and sometimes I feeling any banking sector because I have worked in retail because they are very uh like cool in in that sort of stuff right okay you want to use that you use it you want to use those you can use it only if you have the budget for it but in bank even though you have like huge budget sometimes people are sometimes restrictive but I think in leadership what I would like to see is it's not just using the tool right it's more of how to use it in a right way that I think it's still lacking in in few of the leadership not everyone but I think in bank even though the adaptability is less compared to the retail sectors I guess in banking people should be more open even though you need to think about the securities and all the stuff but you need to make sure the way you're doing it's right I think um the engineers would help the leader or um maybe negotiate with the leaders in a lot better way because they are the person who is doing the ground work rather than leadership ship dictating what needs to be done. I think there should be a work or collaboration between the leadership and the engineers who are working on the ground and they would they should influence the leaders what should we use to make things better and right if you have one advice to somebody for data engineers who may be watching this what would be your one advice um I think there are a lot of materials now you can go and read and learn that's basically gathering knowledge but when it comes to wisdom you need to know how to use it in the right way. You need to know the use case before you apply something because we don't want to apply something because sometimes because if I think myself 5 years or 10 years back in this um environment I would be like oh my god I have got all the information now I am the superior right I wouldn't think about what else should I know to make things better but I think you will get a lot of information which is good and cool you would learn everything but you need to know things to make sure the things that you're doing it's right.

[00:19:40]>> Yeah, I think that's >> I would like to say the one advice uh my manager recently has given me. So if you have haven't started doing this start it today. So if you like uh any of the agencies available for you in your organization for us okay so he asked me do you have copilot in your VS code I said no so do it today start from today.

[00:20:02]So that changed don't wait that's one uh advice actually that changed okay so the way I was actually looking at things and my confidence level gain day by day from there. So that's one advice I would like to say and second thing is probably trust in yourself because you have built a lot of knowledge which is like the same same thing which you mentioned in your talk today morning as well. So that is needed still needed with all the the things which you are actually doing that human intelligence is needed also all the knowledge which you have built over the years as a data engineer that is still needed for us to go forward. So trust in yourself use more more of the get more hands-on that would be great.

[00:20:40]>> Make sure to like this video and subscribe to the channel. Thanks for watching and I will see you in the next