AI Hallucinations: The Shocking Truth Behind Your Data!

Table of Contents   

1. Introduction and Agenda
2. AI Risks 
3. Steps to Mitigate AI Risk
4. AI Preparation for Your Business
5. Conclusion

Transcript

Introduction and Agenda 

So hello everyone. I've done a few sessions so far, but if you do not know me, I am Drew Rob. I am the AI Advisor at ERP Suites. I've been with ERP Suites for about 5 years now working in various different roles. I started in the product space working with our JD Edwards products, Clarity, Scan Ability and Mobility. Then transitioned over into a data analytics, data consulting role and was on that team for around 2 years before transitioning into the AI team becoming the AI advisor. Along this path, I have helped with developing some of our PO CS, a lot of our digital assistants, embedded AI solutions that you will see this week if you've not have already seen or heard about. But my main role is actually, you know, educating customers about AI And really what we take is our AI journey. And the biggest thing, you know, we really talk to customers about is, is how to get integrated with AI, but also really just ask questions about, you know, what are the biggest risk and, and how can we help them through those risk and remedies so we can provide a good AI solution that can provide efficiency and maximize ROI in the business. But there's obviously a lot of questions that always come up. So I do love giving this presentation. I've given it to a, a, a few, a few times at user groups and a few things here and there. But this is probably the biggest, the biggest thing that comes up with implementing AI, especially if you're just getting started on your own AI journey and just getting involved the, you know, the technologies, the various tools and Emily, yeah, just getting started. So, yeah, just excited to give this presentation today. So it's a little bit about myself again. I've been the troll for about a year and a half, two years now. And I'm really looking forward to what AI has to bring us and hopefully mitigate some of your concerns today is is my biggest thing. So with that, I'll, I'll go ahead and get started with the agenda.

So just real quick agenda, not sure if this is going to take the whole 30 minutes and that's OK. I don't want to expand it out. I just want to give you all the necessary information you need to move forward and hopefully alleviate some some concerns as well. So first we'll start off with AI risk. Just go through, we have about eight of those that we think are are come up a lot, especially with questions from customers, various user groups. Now we'll go through the list of AI remedies or mitigation plans, what we like to call them to kind of go through each one and how you kind of handle each one. What you're going to see a lot when going through the remedies is there's a lot of various patterns that show up in different ways to kind of handle these sort of risk. So what we're going to do at the end there and that's .3 is, is really talk about, you know, 5 main points to really prepare your business for AI and really get you guys started with that journey. We do have a security track. You've probably seen some of those sessions throughout this week that will go more in depth, especially in the security pieces that I'll, I'll cover very high level today. But just know that this is just the getting started. This is, this is getting you guys going and understanding these at a, at a very high level.

AI Risks

Hallucination 

So, so we're gonna go through these and, and the first one's really hallucination. So you've probably seen this before and I and I have an example on the next, next slide here. It's just really inaccurate or, or, or misleading data or results you can get from, you know, an artificial intelligence model, one that provides a lot of errors. You know, bias training data can, can cause this really one thing you got to think about and especially another good example is is, you know, using Chachi BT and what not. Sometimes you know, the data it comes back with is just somewhat correct. Sometimes it's, it's always correct and sometimes it's not correct at all. It just really depends, especially with these open AI solutions here, what could happen as far as hallucinations. Because really what it does is when it doesn't know the answer, it provides a lot of assumptions. It's all assumptions. They, they could be correct or incorrect. It, it really just depends, but it's all about, you know, really picking a, an AI tool that, that mitigates the hallucinations and we'll go through that a little bit later.

Job Loss Due to Automation

And, and this is a lot of the business analyst roles, the ones who are doing the day-to-day work every single day. You know, there's a fact out there that throwing 3,000,000 jobs will be lost in the next five years or so. And a lot of those ones you can, you can think about like bank teller jobs, entry clerk. People are doing a lot of manual data entry. It's definitely a thing. But what people need to understand is it's actually, you know, AIS around enhanced jobs. And we'll get into that a little bit further later on as well.

Lack of Data Privacy

Using AI tools is, is very important, especially if you know, as AI tools continue to develop, right? It's not all one-size-fits-all. There might be, you know, an Oracle tool out there that's very good or a large language model like Cohere that we like to use here at ERP Suites. It's actually embedded in one of our digital assistants might not be so good at doing something as far as like data visualization of providing data insights for predictive analytics. So it just really depends, how can we kind of combine all these different AI solutions with your various data sources, with your various analytic tools, How can we bind them, combine them all together in the best way possible to minimize lack of privacy? Because that's definitely something that comes up when you start doing all these different integrations.

AI Bias

That's that's another big one. It's when your data is really, you know, it's skewed in a certain way to get results that you want, but aren't actually correct. And that, and that's the biggest thing is like making sure your data is, you know, complete and trained. So, so your results are indeed not inaccurate, right? And, and Justin doesn't just produce the miles don't just produce flat out wrong results. It is another big thing that that definitely comes up, right.

Lack of Data Governance and Compliance

And again, I, I mentioned it before, this is really huge inside of organizations, right? Big for like data breaches when you start to use things. I mentioned ChatGPT, but other large language models out there, other third party systems that aren't integrated in your, in your four walls. It could be a potential, a huge concern and a potential big risk as well. So doing it the right way is very important.

Loss of Human Intelligence

And this is an interesting one that we really thought about, but it's something that came up because it stands out in the sense that, you know, we all add some way, shape or form you ChatGPT or or some form of AI tool every single day, whether it be to help you create presentations or help you write code. The game back on top, it's, it's all about not losing your sense of human intelligence and just really going out there and, and still making their strategic decisions and having AI help you in a sense, but not take over your entire job because again, it's not always going to be 100% correct. And that's the biggest thing. It's just understand that AI can help your job, but not fully take it over.

Steps to Mitigate AI Risk

Steps to Consider for Eliminating Hallucinations

So, so moving forward and again, I'll kind of go through each one. I kind of hit on a few of my talking points earlier, but just really steps to consider for each of these, right? So steps to consider when to eliminate those hallucinations, those incorrect data sources. It's really restrict restricting your data set to reliable and verified sources. You know, it's whether your data sources are in JDE, in some other Oracle database or in a customer service system. You know, Salesforce is 1 I can think of off the top of my head. And if you need to bring in data like economic data is another one we use for predictability in building out some models. Just make sure it's built in, in a secure data lake when you're reaching out to it, when you're, when you're connecting these models to them, just making sure the data is reliable.

Continuous Data Validation and Data Cleaning

Another one big one is continuous data validation, data cleaning, right? The one thing we really talk about when we build out AI solutions and where we like to start is is definitely on master data management. It's just very important, right, for an AI solution to have good data to train on and have good master data management. I remember sitting in an Oracle session probably like a few months ago now, maybe it was not even longer, it might have been six months. The main tagline was data is gold and that, and that's really tried and true when you're, when you're building out any, any AI solution, whether it be a digital assistant or a document understanding tool or, or really any predictive tool that that we have here at ERP suites. It's, it's all about starting with that data preparation phase.

Human in the Loop

And this is a huge one. It's integrating that human integration loop, making sure that whoever is building out the process, who, whatever human is involved in building out an AI solution is involved in, in that entire solution. We want subject matter experts there when building up any tool because they are the ones who understand the process, they'll understand the data, they'll understand how it works. And, and that's really just the best way to, to build out an AI solution. I'll probably get into what I'm going to talk about later with, with the loss of jobs. It's, it's really important to understand that, you know, we want people to be educated and be a part of building out AI solutions. And it's not, again, not a loss of jobs, it's, it's enhancing jobs in the end. So those are the steps for Illumination and we're rolling into this point as well.

Employee Education on AI

It's really getting employees educated on necessary knowledge and skills of AI. It's not just the tools and technologies, it's how they really use them, right? And if they really want to, they can be involved in the whole model creation process as well. It, it just all really depends on how, how much customers, how much people really want to be involved. But it's all about promoting usability. It's not just, you know, implementing an AI solution just to implement an AI solution, you, you want people actually using it and promoting it in your business, right? And then again, just effectively communicating to employees how again, business, you know, the business processes, business improvements, helping them make more strategic business decisions. They're not doing the automated and mundane task anymore. They're being efficient with their work. And that's, and that's really important as well.

Gradual Implementation of AI Solutions

And again, just, you know, implementing AI solutions gradually over time. And I've mentioned it a few times this week already, but it's really the crawl, walk, run approach that we love to say. And then phase one, phase two, phase three as well, you know, start with one specific flow, start with one specific process with one specific data source or a small subset of data to really get your feet in the water with AI, right. And then this will help in turn with everybody in your business top down from from the executives, middle management all the way down to to the business users to really start adopting AI in the right way and really getting people comfortable as well. So, and another term we like to use here at ERP Suites is definitely just a quick win, something easy, something simple, something to start adopting AI so your competitors can see you have it right. And it gives you a big market advantage as well, competitive advantage in the market. So that's also very important as well. It's continuously coming back, just imploring to employees, this will help you in the end. And again, it's, it's all the right AI solutions as well. Like during my AI 101 session yesterday, I talked a lot about what we really see in the enterprise's strong AI solutions. Not everything's a good AI solution right now. There's still a lot of development out there. So the important thing is really understand what are the right solutions. How can we get the quick win? How can we get everybody on board? It's super important with the next one and I probably have a few of these points, but really lack of data privacy, especially with the ERP suites. We, we work hand in hand with a lot of OCI services and, and they, they have a good authentication between OCI and JD Edwards where we house most of the data. And with that it, it becomes, you know, good end to end data encryption. All our CI services are encrypted from that rest and that in transit. We can talk about, you know, data security as far as roles and environments inside of JD Edwards as well.

You know, for I'll give you one example. You know, we have a financial digital system, we have a manufacturing digital system here at ERP Suites. Actually, I, I might even go a little bit further. Let's let's use this example. So what about, you know, someone working in manufacturing, right, one person's in planning and one or one's in procurement or something. And and you know, we, we will have JD Edwards security strictly around certain people who can access certain process flows. It can go even deeper than that. It's not just manufacturing or finance. It goes into who has access to certain roles and responsibilities. What is their day-to-day lives? How do they operate that? It's really important as well. And I think that's, you know, that's actually an access control is is definitely one of the sessions. It might already happen actually, if it escapes me, but that's a really good session that we're also giving this week at AI week. And again, and I kind of harped on this earlier, only collect data that is needed for the AI solution. Really start small. You know, don't don't just put all of your data in one basket, especially if you're moving it right. We can a lot of the solutions are hybrid, right? They can be a data can be on Prem and you know, a lot of our services are in the cloud. But if you're moving the services to the cloud or you're moving on Prem or vice versa, just make sure to start with that little subset of data because that will also help to mitigate data privacy in the end as well. And then again, going back to that bias, right? So just making sure data sources, you know, we're feeding the right data into data models before training them. That's also just very important, right?

For example, like we're working on a, which actually you have this afternoon, it's another session if we're working on a document understanding sales order and we're testing that, right? Let's say we send in 50 PDFs that look the exact same from like 5 different customers, and then you send in two that are handwritten from completely different customers. Obviously you're training it to read the other 50. And whenever there's one PD FS come in, but the handwritten ones will be incorrect, right? It's all about, you know, having level training, right? When you send in your data and you start to train it as well as just in the end just master data management, it comes back to that as well. We'll all help to eliminate bias. So again, it's it, it requires that human in the loop again as well. Just it, it being a part of this entire process. And again, as that, as that last point says, refreshing and retraining models as they go on, you know, continuous improvement, continuous maintenance through AI solutions. You don't just build out an AI solution and then just kind of leave it be and use it, right? We want to build on to these, we want to make them better, but that all starts to really improve data and, and good data and continuous training. So that's very important as well. When when you're going to eliminate bias from any AI solution.

Data Governance and Security 

And then lastly here it's really just about, you know, steps for strong data governance. This is huge. Like this is what we hear from any anyone in the business have a good code of ethics, right, clear roles and responsibilities in your business who will be dealing with what AI solution really getting involved. You know, you have a security, I'm sure you have security plans out there, responsibilities, architecture, right? Being able to seamlessly integrate that with AI security is huge as well. And just having steps towards that, right? And, and just doing regulatory reports, you know, data privacy is also involved in this and enroll based security is huge as well. And we talked about that a little earlier when we were talking about data privacy, but it's just in internal reviews, right? Keeping a log of everything is important. We like to say that a lot of our AI tools keep keep history logs and that's it's, it's true. Definitely use those. See how they're working. This will also help the train models as well. I know I'm getting a little off topic, but it's also very important in this space as well. So a very strong data governance, data security. Again, we have a whole security track where you can learn more. If they, if the tracks have already happened, you can go listen back to those recordings as well, which are always which, which are always very helpful.

AI Preparation for Your Business

Conclusion

And then lastly, just how can we help to talk with us? We'll be a blueprint. We got, we got a booth down there out you'd be down the blueprint as well. Multiple user groups will have sessions. I'm sure later this year on, on AI might see you guys out there as well. So looking forward to that. And lastly, we have a video podcast. It's actually out on YouTube, it's at ERP suites as well. It's on YouTube and Spotify. I've made a few appearances on there. It's not just AI, it's it's, it's application managed services, it's our advising practice. It's it's Configurator, it's, it's really anything. So get a look at that podcast as well that we're, we just started also good information as well as, you know, visit our website as well, you know, ERP suites.com, a lot of demo videos that we're showing this week on AI and, and other things. So, and, and if you want also just real quick plug, I don't have it on this slide. We do have an AI starter guide out on our website as well. So if you need access to that, that's really everything AI journey to get you started wherever you may be on your AI journey. So, and with that, I that's everything. Any questions? Don't see any questions. OK, well, you guys have a good rest of your AI week. I hope this session was beneficial and I'm sure I'll see you on on a few more, but have a good rest of your day as well. Thank you.