Combating Fraud Using AI – Techstrong AI Podcast EP53

Hello, and welcome to the Techstrong AI Podcast. I'm Amanda Ani, and with me today I have Gura Al. He is a data science manager at Thermo Fisher Scientific.

How are you doing? Yeah, I'm doing good, Amanda. Thanks for having me.

Happy to have you on the show. Can you share a little bit about your background and experience? Yeah, sure.

Uh, currently I'm working as a data science manager and, uh, I have sound, I'm a technical person having sound knowledge in A-I-M-L-A-W-S cybersecurity. Uh, I have several articles which have got published and, uh, I love hacking, uh, and participating as a judge in the hackathons. Uh, like currently in my role and previously in all the organizations, I have been responsible for developing innovative solutions for the organization, which can save, uh, like money as well as time for the whole team.

I have written several utilities, which, uh, are a combination of ai, uh, algorithms, deep learning classification regression, along with the automation, utilities, the solutions, what I try to build. They are mostly like, uh, open using open source technologies. So they go a long way, and, uh, that's what I love to do.

Wonderful. Well, you're the person to talk to then today about using AI to combat fraud. That's our topic for today.

Yeah, sure. Frauds are happening everywhere and, uh, AI is also moving very fast in each and every domain and in every, uh, different kind of, uh, industry. And so it's very nice to include AI to combat fraud.

I can cite few scenarios like how AI can be helpful. Yes, absolutely. I think that's my first question is where are some of those fraudulent activities being seen, and if you can give a little bit more detail about how to utilize AI to combat those.

Yeah, so, uh, like insurance organization, okay, so what happens, like people, they are trying to file a claim and, uh, this claim process is completely manual, whether it's healthcare domain or it's a phone insurance. Okay. What happens, like, uh, when the claim person, they try to call it odd timings when the, uh, like the employee, they want to leave, so they are in a hurry, and this guys, they try to make up a story, uh, uh, which, uh, quite emotionally these guys employ fees.

Okay, yeah. He's saying correct and they file a claim, they approve it. In this whole approval process, it's a company who has to bear the loss.

For example, like I give you an example, like we have, uh, several algorithms like Sound X, sound X is like, uh, uh, your, your names, they're matching, but you are not the real person who should be filing the claim and whether looking over it manually, it's actually quite difficult to figure it out. Oh, yeah. It's a fraudulent claim.

With ai, you can train your model with the past fraudulent claims, and when the new request is coming up, instead of having the manualize on it, AI model can give you an output with a confidence level. Yeah, this is like 70%. Yeah, this guy seems to be a fraud, so it'll give you some, uh, checkpoint.

Yeah, I need to pay attention to it. That's how AI is very helpful. So like, uh, another example is like we are receiving several emails like, uh, this insurance companies, they are receiving hundreds of emails on daily basis manually.

It's a very tedious task, but if you are generating, uh, any model, like a named entity recognization model, what it does, it reads the email content and it will give you the output. Okay, yeah, this email is about this, is this an issue or not? So over the voice, you can feed the model, the email, you can feed the model, the model will give you an output according to its confidence level, and, uh, you will get a checkpoint.

Yeah. Whether this is a fraudulent or this seems to be, uh, like, uh, a good, like we are, we are good to approve the claim. So that's how like AI helps, uh, uh, in claim process.

So, you know, we see this technology advancing very rapidly and while AI can be used, so we're seeing how it can be used to combat fraud, we're also seeing way more evolved and advanced fraudulent activities. So how do you, what advice do you have for leaders as to stay on top of this and the evolving the rapid changes to this technology? See, when we have rapid changes in the technology, the problem whether AI is like, uh, your model is being trained on some specific data and based on that data, it is giving you the response.

The problem happened. Like tomorrow there is some different use case, which model was not aware of it, and it may give a wrong output. The my advice is like, uh, you have written a model now it doesn't mean that your work is over.

You need to retrain your model on the new upcoming data because the hackers, they are also intelligent. They, they are also getting a smarter, so you also need to be, uh, like you are not retrain yourself like, uh, uh, with new updated data and uh, uh, like, uh, come up with a diverse data so that your model is intelligent enough to understand any new upcoming requests. So my advice is keep retraining the model.

Are there any challenges or limitations to these AI based fraud protection tools? And if so, what are they and how can organizations mitigate those? So I give you the, uh, example from generative ai.

People are liking it. It's, uh, it works on LLM model. You just, uh, provide your input, uh, in English plain language, and it gives you a response back.

Okay? The problem here is, uh, with this gen AI versus other models, this gene AI model is not giving away confidence level every model. Like, uh, when they, uh, give you an output, there should be a confidence that okay, model is saying, yeah, I'm 70% confident this answer is correct.

Whether gen, ai, uh, the models, this confidence level is missing, they're still in process of implementing it. So this is one lacking feature. Like you cannot directly deploy the gen AI output to the production because that model itself is not very confident.

You don't know what is that percent with respect to other AI ml models like you have for regression classification, where we are predicting something, it gives you a feature of, uh, uh, like a confidence level. So that helps. So that's why when you are talking about the fraud, it can only give you a, uh, like a sanity check, yes, whether this can be a fraud or this cannot be a fraud, but that does not mean that you avoid the manual activity at all.

Frauds are happening all over the place, so please retrain your model again and again so that you can get a better output. Absolutely. So do business leaders need to consider any laws or regulations when it comes to utilizing AI for this?

Yeah, so like, uh, every company, they are coming up with their own RAI, uh, like responsible ai, uh, framework. So there are rules, like you cannot, uh, because if you are start feeding like your organization personal data, then that is not correct. So you have to abide with the laws, what your organization is giving.

You have, if you are building a model internal, uh, model based on the, uh, like company's data. So it, that model should be consumed internally only. It should not be exposed to the outside world.

So this is just one very simple example, but yes, there are rules which business leaders they need to take a security was there is there will be there. So security is still there. AI is something new, which has to go along with the security.

So that is a very important aspect and we have to follow like what uh, organization is coming up with in the rules and regulations. All right. Is there any issues with skill level?

AI is a pretty, um, well, it's been around a while, but the more recent version of AI is a pretty new technology for most people. So are you seeing that there's a skill level gap? And what advice do you have for business leaders to improve that area?

Yeah, so definitely there is a skill level. Everyone knows what is something is their ai ai, but they don't know like exactly how it works or how we can, uh, pitch in how we can use it. So the, luckily there are so many UI platforms like chair, GBT, you don't need to know ai, you just provide your input, it will give you the output.

So that is all ai. So my advice is like, uh, if, uh, like if there is a technical person and he or she wants to learn ai, you don't have to, uh, like, uh, in, uh, like include AI in your code and every phase, no. You just go with the flow and if, uh, in your project it'll involves ai, you think like when you are, uh, trying to implement something, you Google it and Google is saying, yeah, so this is a prediction case.

Use ai, otherwise just stick to your regular work. For the business leaders, like we have process business leaders, they're mostly involved in the process, like how we can make it better. AI is helpful to them in making the process being more innovative or providing some, uh, use cases so they can use AI for learning how they can improve the existing process.

But that does not mean they have to implement it for guidance, for knowledge, for coming up with new ideas. AI is there if you find something better. Yes, go ahead, use ai.

Wonderful. Well, if there was one key takeaway you could leave our audience with today, what would that be? My key takeaway is like AI is very good, but it is an artificial assistant.

It does not have its own brain. So even though you are using it, you are getting some output. Please, uh, uh, do do a thorough manual review so that you can, uh, look over the answer and then you can implement it in the production.

Just don't directly rely that, okay, it's AI output, it is a hundred percent correct, we are done. No, it is not that. So it's, uh, it's an assistant.

That is my advice. All right, wonderful. Well, thank you so much for coming on the show and sharing your insights with us today.

Yeah, it's a pleasure. Thank you. Thank You.

Great. And thank you to our audience. Stay tuned.

There's more.