Unlocking JD Edwards World Data with AI

Bill Onion
All right, good afternoon, everybody. We will get started here in just another minute. Appreciate you all showing up. Okay, well, it’s 1:00. go ahead and get started. So I appreciate everybody taking time out of their day today and spending some time with us. We’ve got some fun topics to talk about today. Really, today we’re talking about JD Edwards World Data, IBM iDB2 data, and really leveraging AI to access that data. So What we’re going to talk about is business users. Can they get access and can they get answers from JD Edwards World faster and easier without bypassing any IT controls or making major customizations to their system? There’s two companies that are presenting today, Bright Skies, myself, I’m with Bright Skies, and Eridani. We’ll get through introduction here in just a minute. But one really important thing that I want to share before we go any further. The presentation today is not about this is a cool use of cool technology, although it is very, very cool. But this is really a business opportunity to get access to your data, get access to that data faster. Before we get started, I just want to take a quick snapshot of what we’re going to see a little bit later. If you know JD Edwards at all, this is the advanced price and availability lookup. You can see here that we got Claude running, right? And the user at the top said, hey, look up a price for this item, for this customer. for this branch. And AI did a little bit of thinking there, and what it brought back was everything you would see in that advanced price and availability lookup, which is the unit price, some of the inventory availability, some charges that are in there, and finally a final price. Also, if you look at the bottom here, it references core charges and alternate branches. that information is important to this particular company and how they have JD Edwards configured, but ultimately, it’s returning exactly what’s in that. I want to point out, again, we’re not logged into JD Edwards, we’re not on a green screen, and this is asked using just normal English language. So that’s the kind of thing we’re going to take a look at today. But before we get there, Quick run through of the agenda. We’re going to do some introductions. I don’t want to bore anybody. We’re going to go through this fairly quickly. Then we’re going to talk a little bit of background about why this project started and really talk about the project, which is really a proof of concept. We’ll have some time to go through a demonstration of this. We’re going to have Aaron and his team talk about how this actually works. And then we’ll take some lessons learned, talk about some next steps, and certainly open it up for questions and see what anybody has any questions on and answer those questions for you. So before we get started, quick round of introductions. My name is Bill Onion, managing director of Bright Skies. My personal experience is long, long, long time ago I started on IBM i. and RPG and JD Edwards World, never made the jump to E1, but my experience really is on systems and configurations and integrations and architectures and those types of things. I’ll be speaking the most today, kind of emceeing everything and bringing other people in and out as we need. Nathan Giri is also on the call. He is a very, very talented solution architect for Bright Skies. He’s been on IBMI and JD Edwards for decades. He has experience both JD Edwards World as well as Enterprise One, orchestrations, Erika, Picket Technology. He’s probably touched it somewhere throughout his lifetime. And then a little bit about Bright Skies. Bright Skies as a company, we are an IBM partner for the IBMI platform. and an Oracle partner for JD Edwards platform, software platform, I should say, and just a lot of experience around modernization and integration and reporting and business consulting and that type of thing. Erin, I’ll throw it over to you. I’m sorry, Dan, I’ll throw it over to you for Eridani introductions.

Daniel Magid
Great, thanks, Bill. Yeah, so my name is Dan Magid, and I am one of the founders of Ariadna. And Ariadna was founded for the purpose of helping to bring new technology to IBM i users and making it easier to access data on the IBM i and to get to outside the IBM i to access data. We’ve been working with Bright Skies now for several years, doing just that for many, many customers. I’ve worked in the IBM i world and the JD Edwards Edwards World since the early 1980s. So I’ve been involved with this for a long time. I’m very familiar with it. And I’m really excited about the possibilities of AI in this environment and what it can do for you. Meanwhile, Aaron, who is quite a bit younger than I am, has been programming, I think, since he was in middle school. and has now become very, very skilled in the IBM i world, but also brings a lot of experience with new technology and in fact is way out in front of the move to AI. And so he’s going to talk a little bit about how the Ariadna pieces work in the Bright Skies environment for bringing you data from your IBM i.

Bill Onion
Great. Thanks, Dan. Appreciate that. Let’s keep moving here. So first, let’s talk about the current reality, right? First of all, in JD Edwards World, there are many, many, many tools that are used for reporting. And to be very clear, we’re not suggesting that you replace what you already have in place, right? But there’s tools out there like ReportsNow, ClickView, Power BI, Tableau, Insight. All of those are in place and they do a very good job and they have a place and a space in the reporting world. What I want to think about just for a minute is, what if a user has a new data requirement? Maybe there’s a marketing campaign that somebody wants to run for. inactive customers, or someone needs to adjust an existing report, or you just need to add a new field or add a calculation, or maybe the calculation is wrong, or they want to add or remove some filter. Well, in the current environment, right, business user says, I’ve got a change that I need to make, right? and they e-mail it to IT or maybe it’s a ticketing system or some process like that. Now, the developer who needs to work on that or would make that change, as we all know, they’re usually very, very busy. They have competing priorities and they’re pulled in about 17 different directions, especially if there’s a production support issue that’s going on. Eventually, they can work on that project or that request. And the RPG coding change or the SQL change, right, that’s usually very straightforward. It’s usually a very simple change that somebody needs, unless it’s a whole new report. But that change gets done. The report gets run, maybe goes out to an Excel, something like that. and gets back to the end user, then hopefully, imagine that the user then has another change to that, you know, then that whole cycle starts all over again. And what we see, and I think we all see, is there’s just some bottlenecks along the way. There’s competing priorities, there’s some slow turnaround times, et cetera. And so. The idea is around this is what if the user could simply ask JD Edwards? What if they could actually talk and have a conversation with JD Edwards? That’s the concept that we’re exploring today and taking a look at today and going to share with you today. So a little bit about the why, like why did this project start? So some You know, some questions that business users ask every week and are asked of business users every week include things like, hey, what’s the current inventory? Maybe for a certain region, or what’s the price for this customer? Or why didn’t this order ship? You know, again, imagine if a business wanted to ask, hey, which customers haven’t ordered recently? What’s kind of below that reorder point? Again, not difficult questions, just sometimes are difficult to answer quickly. Yep, 100%. JD Edwards screens have some of that data, right? You get some reports that have that. But navigating between those two can be tricky. And if you’re trying to combine a couple of things and changing your filters along the way, That’s always a little bit of a challenge. So again, instead of requesting a report with AI, the user simply asks AI and JD Edwards, and they have their answer. And longer term, the user can dig deeper into data into JD Edwards, again, trying to avoid that developer requests or the waiting in queues or just delays in getting those answers. So let’s move on a little bit. Let’s talk about that initial project, the project that we’re showing off today. This project was approached like a proof of concept. As a matter of fact, it is and was a proof of concept, but that was 100% on purpose, right? Scope was kept very, very simple. And really the goal was to get a couple of quick wins before overcomplicating. overcomplicating everything, which we know can always happen. So really, let’s talk about why this project was started. So like I mentioned on that earlier slide just a minute ago, you know, this company is also in the B2B space like that. And again, those most common questions that they’re receiving are, what’s my price? Do you have it in stock? and where’s my order? You know, longer term, the thought is, you know, using this technology and using this tool, can deeper questions be asked? Why didn’t this order ship? Hey, maybe which big customers in 2025 haven’t ordered anything in 2026? What high moving items are below their reorder point? Again, these are not difficult questions. right? They’re just difficult to answer quickly. You got multiple screens you got to get to, you got multiple reports, you got different filters, you probably got some Excel sheets in there, and you’re trying to work all this together and make it happen. Again, like I mentioned earlier, hey, what if the user could simply ask, right? That’s the idea for this project is let the user simply ask. And eventually, we’re going to get to, imagine this, what am I open orders? The easy question. Well, hold on a second. What if we exclude the Western region from that, right? Let’s also exclude the Southwestern region and then say, hey, of those orders that are left, what are the top three items on back order on those orders? You start to see here where you can start to build a conversation on your data and just continue the conversation to get to that end. So that’s what this project set this foundation for, just letting the users talk to JD Edwards. Now, we talked specifically about the company. They are located here in North America. They’re automotive aftermarket distributor. So there are lots and lots and lots and lots of those kinds of companies running on JD Edwards and in the United States in general. They are running JD Edwards World A9.4. On IBM iOS, they’re 7.4, I think, might be 7.5. Claude is their AI tool of choice. This could be GPT, this could be Copilot, doesn’t really matter. They just happen to want to use Claude as their tool of choice. Some project goals are, first, no changes to JD Edwards World. right? Next, keep IBM i and DB2 in place. Again, no changes to that. They really must maintain and leverage existing security. That’s both IBM i security as well as JD Edwards World Security. Think branch plan, company security, that type of thing, right? Another goal was to get easier access to JD Edwards data and information. And most importantly, introduce AI and use AI responsibly. That was a key one. So kind of a crawl, walk, run thing. So that’s what that is. And I just want to reiterate that most importantly, JD Edwards and IBM i and DB2 all should remain the same with really no changes to that.

Nathan Giri
Yep.

Bill Onion
So now let’s get on to the really fun stuff and let’s take a look at how AI can be used to access JD Edwards World data. So first, as I mentioned earlier, this project started with simple topics, right? Pricing, inventory, customer master. That’s really, really the data points that they want to take a look at. And the goal is for these questions to be able to be answered in one contiguous conversation or continuous conversation, meaning not through separate requests and not through separate report things, not even 3 conversations, but just one conversation and that can go. So Nathan, get ready. We’re about to throw this over to you so you can run through this, but I also want to be clear. We did use screen captures today rather than a live demo. If anybody’s done live demos before, we all know the challenges with that. But what you’re going to see here really shows the power of AI connecting the JD Edwards World and providing some really interesting answers for that. So with that, Nathan, I will throw it over to you.

Nathan Giri
Thanks, Bill. Now, let’s take a look at what does actually look like from users’ perspective. The key thing to keep in mind is that JD World, IBM i, and DB2 are still in the systems underneath. We are simply adding a conversational way to get to the information. So in this slide,

Bill Onion
Yeah.

Nathan Giri
I want to quickly show how AI is connected to the IBM i environment. This is the connection configuration. We have the MCP server URL and the access token. The MCP server act as the communication layer between the AI and the approved IBM i capabilities. The important thing here is that we are not giving AI direct access to the database. That’s a key thing. So we are connecting AI through the MCP layer, and only the tools that we have configured and approved are available to it. So IT still has full control over what AI can access. So AI doesn’t get open access to IBM iJDE. It gets access to the. specific tool we allow. So, next screen. So this is how the screen shows the MCP server that AI is using to communicate with IBM i. So you can see the client IBM i connection configured here. So again, the important point is controlled connection. AI doesn’t directly connect to JD tables or IDB2. The MCP server acts as a communication layer between AI and the approval. Approved IBMI capabilities. Now, let’s see what AA can actually do with the connection. So, next screen. So now we are inside AI, and you can see that it recognized the tools and have been made available to IBM I connections. For example, we have a customer lookup, pricing, JD data, and some other IBM I related development tools. The important thing is these are not just AI tools. they are connected to the business environment. So we are giving AI business capabilities, not just giving AI access to data. Next. So this screen shows the actual tool list in more detail. Here we can see some of the actual tools that have been exposed. For example, we can retrieve customers from IBM i, perform a price lookup, or look up pricing for a specific customer account. We also have tools that can work with other IBMI process. And also it gives a sample prompt for each tools. So, let’s move from configuration and actually ask a business question. Next, next. Now we get to the part that is interesting for the business user. So instead of opening JD and navigating through multiple screens and looking for the right customer and item, the user simply asks a question in normal language. So in this example, I’m asking for the price for a specific item for a specific customer and branch. AI uses the approved JD pricing tool and returns the information. And notice that we are giving more than just the price. We have the unit price, availability, and other pricing letter details. So the user gets the information in the conversation way. So user didn’t have to know which JD program or file contains this information. They just asked the question. Business question actually. Next. So now let’s look at another example. I’m asking A to look up an address book record, in this case address book number one. So you can see that A is asking for permission before executing the tool. This is another important part of the governance model. The A doesn’t silently execute every action. The user can allow the action or deny it. For this example, we allowed it. So this is a permission and approval mechanism between the AA request and the actual system call. Next. So now we approved it and we got the reset back from JD. AI found the address book record and presents the information in much easier format. Notice that this is not a report that somebody created enough time. The user asked the question, the AI interpreted the question, called the appropriate Tool and return the JD information. But here’s what it gets even more interesting. Next, next. So this is something that I can, I think this is very important. A, doesn’t always have the guess. If the question is ambiguous, it can ask a follow-up question. In this case, I said, show me the address book records for ABC Brand. A, it recognized that it needed more information. So it asked me how I want to search and which environment I wanted to use. So I simply answered yes and continued. This is more like having a conversation with somebody who understands you were JD in alignment. Next. So now A returns a large result. Now we have another practical situation. The query returned like 4000 records. Obviously showing 4000 records in a chat window isn’t very useful. So A recognize that result. It’s too large and gives us a simple, gives us a sample. And now it gives us two choices. Either we can narrow the query, maybe only for customers or employees or specific business unit, or we can export full reserve to Excel. So the point is that the AI isn’t replacing Excel RPA tool here. It’s actually helping us get the JD data into format the business user can work with. Next. So I can simply say export to Excel here. A creates the spreadsheet with JD data, and now the user has the information in Excel and can continue working with it. In this example, we have all 4000 address book records with the relevant columns. So, think of what just happened. We started with a simple business questing. We didn’t create any new JD report, or we didn’t create any new JD RPG program. We didn’t manually extract the data. We simply asked the question, got the answer, and exported the information. That’s the real idea behind this proof of concept. Making the information that already exists in JD. much easier and faster for the business user to access. So that’s the user experience we are looking at. Bill, Dr. Bill.

Bill Onion
Very good. Thanks, Nathan. We appreciate you walking us through that. That is some pretty cool use of technology there. Now, right, now let’s talk about how does, what’s the magic behind this, right? Like it was a very straightforward, very easy to use front end. Anybody who’s used AI has done queries like that before. Now I’m going to turn it over to Aaron, and he’s going to explain to us what that magic kind of looks like. So Aaron, next off to you.

Aaron
Yeah, thanks for that intro. So, okay, great. This is awesome. So I’m getting so excited watching this stuff because, you know, I love getting calls from people when, you know, when I get to hear what they’re doing with these AI technologies and how they’re building out their integration. So that’s so cool. Thank you. Thanks, David, for that. So All right, what I want to talk to you all about now is how the how the piping works for this to actually make this possible, because what what we’ve gone through here is very cool, as Bill said at the beginning, but it is so much more than you know than than than what it may seem, you know, just when you look at one of these screenshots, right? What you’re actually doing here.

Bill Onion
I know.

Aaron
is you’re bringing intelligence into your data, right? You’re giving your systems the ability to act for themselves and to proactively deal with new situations that you may have never encountered or never anticipated. right, which gets very, very powerful. And I can geek out about this for a lot longer, but I’m going to stay focused. So one of the key things I want to cover here is the security layer here, because That, honestly, right now in the AI industry, that’s largely what the whole industry is trying to figure out right now, right? There are news reports again, you know, kind of every day about another agent breaking out of its sandbox, you know, and going and hacking somebody else’s system, you know, or doing something. So the question is, how do you make sure that this structure is working, that it is governed, that it’s secure? because that is what you have here. And if you’re going to expose real business data, you’re gonna need to make sure that you can rely on those AIs, right? So how do we actually do that? Because that is what’s in place here. So I wanna first break up this flow, right? The flow of this process is. A user starts off with a question or a prompt into the application that goes into your AI, right? Your AI is a flexible, what we call a non-deterministic system, right? It’s able to make decisions. It can be creative. it can answer differently on two different attempts of the same inputs, right? So from there, the AI is then going to leverage your MCP tools. and the capabilities that you give it to. perform governed actions on your system. So there’s this connection point between the AI, which is non-deterministic, which is creative and flexible, and your core systems. And that’s the point where we have to be very careful about security, right? And that’s really where we’re going to be focused on. And the analogy that I want to give everybody to think about how we think about this security is, I might get some weird looks for this, but I really do believe it. AI at this point. is a lot like a person in a lot of ways. And we need to keep that in mind when we’re thinking about security. What do I mean by that? What I mean by that is they can be persuaded. And to give you a concrete example, right, if you’ve got valuables, you put them in the safe, you maybe have a person standing in front of it, you know, controlling access into those goods, into those valuable things. But you don’t just want to have a person because if a person has a rule book and you know, somebody asks, can I get in and can I pull things out of the safe? And the person says no. That’s the right answer. That’s what we want. But depending on the inputs, that answer might change, right? Somebody slips that person $100 bill, the answer might change, right? And we need to be watching that with our AIs too. So what we do with the AIs is we make sure that they have a governed pipeline for accessing data. that is a non-AI process. The AI is leveraging these tools, these MCP servers, that allow it to access data in a secure way. And what that means is you’re circumventing the whole question about can I, about the AI. and its reliability. The AI is your assistant, but it can never do anything without the appropriate governance and the appropriate authorizations. That’s the core of how you keep these systems secure. That’s really what it’s all about. So Bill, go on to the next slide and I’ll talk about the architecture here.

Bill Onion
Sure. Yeah.

Aaron
So fundamentally, the way that we do this, right, is the AI is going through a governed tool. Again, it’s a lot like a person. You don’t just give a person, hopefully, random administrator access to all of your databases, right? You give them web apps, you know, especially, you know, like, you know, people all throughout your organizations, right? You give them an application, right? They’re going to your 5250 screens, they’re going to your web apps, they’re going through whatever those processes are, and those processes are governing their access. There are things they simply cannot do. AI follows the exact same requirement, right? So in here, what Ariadna is providing in this application, what Ariadna Connect does, is it provides the governed bridge between the AI and the core business systems, right? And so what’s happening is… Bright Skies is working with our tools to build out those pipelines so that an AI assistant can make a request of the integration layer, and the integration layer is then accessing the systems on its behalf.

Bill Onion
The.

Aaron
And that’s similar to giving an app to a person or a particular governed process to an independent actor. That allows the AI to make whatever requests it wants. And the governed layer is tracking everything it’s doing. It’s securing every action and authenticating the AI at every turn, and it’s only authorized to do specific actions on specific parts of the core systems, right? And that’s what keeps it secure, right? No matter how, no matter what the AI is asking to do, it’s always going through this layer. We’ll go ahead to the next slide.

Bill Onion
Sure.

Aaron
And that. That basically comes down to a couple of key principles. And I just want to close out with these critical principles, right? Number one, as I said at the beginning, the AI never has direct access to your database or your core systems, right? We’re not just, you know, granting it a, you know, a SQL connection to the system and just saying, do whatever you want, right? That, right, AIs, you know, they’ll come up with ideas, they’ll be creative, they’ll be flexible, and sometimes you really don’t want that, right? So it’s not doing that, and that is a critical point. It’s also not directly interfacing with database tables through its process. It’s going through a high level. business layer. It’s not saying I want to read this database table and run this query for me. It’s saying I want this particular type of data. I want the accounts in the system. And it’s up to the governance layer to figure out how to get that data for the AI. And critically, the governance layer is also not just getting generic access to the system. The governance layer itself has to go through the IBM i and the JD security layers to actually access information. And, on top of all of that… The integration layer is tracking, logging, and authorizing every request that’s going through it. So what that builds out for you is a whole pipeline that says no matter what the AI is asking for, and the human user behind it, no matter what they’re asking for, the AI gets to figure out. how to place its request. The integration layer figures out how to execute on that request, and the security layers, which are all non-AI, so they’re not vulnerable to that persuasion that AIs can be vulnerable to, they are responsible for determining what you’re allowed to do and whether or not your request can actually be allowed through. And that is how you create a secure layer to get the power of AI in a safe way over your core systems. That’s really the core of the whole strategy. So I hope that’s helpful. We pass that over back to Bill to keep going.

Bill Onion
Great. Thanks, Aaron. That was really helpful. And that really explains that layer between just AI and JD. And it’s just, it’s literally not the wild, wild west, although it could be if not done properly. So appreciate that. So let’s talk a little about lessons learned. So going in this project, as you would expect, there were some assumptions about where AI could provide value and where it couldn’t and what was good and what wasn’t. And Like we know about all assumptions, some were right and some were not. So I’ll just share a couple things here. So first of all, none of these questions that I’m even, that I shared earlier that people are asking, none of them are hard. You know, getting the answers manually, you know, can have many steps. And that makes it hard to get that answer. AI just makes it easier to get that answer. It makes it easier to get there faster. That’s frankly what that what that what that’s there for. So, and 2nd is, I mentioned it earlier, like this, this is not intended to replace your existing reporting system. There’s a place they’ve been in place for a long time. There’s value to all that, but this really enhances that. You know, think about it, like think about having a really smart BA and be able to ask that BA, Aaron mentioned a question, AI is a person, right? So think about that as a person, a really smart BA who has super fast access to all of your data. You can ask it a question and then go get the answer for you. You know, it’s different than Reports Now or ClickView, Insight, et cetera, Tableau, whatever you’re running there. But that’s just really the best way to kind of think about that. You know, experience matters, right? This is less about the tech because it’s cool tech, but it’s more about, you know, how well somebody understands JD Edwards and JD Edwards data and they can ask them better. questions. That’s really what that’s all about. As Aaron explained, you know, the API framework, I mean, it gives you that access to the data. It just makes it easier to ask simple questions. And it really keeps the person out of what I call AI trouble, because we all know we can definitely get into some trouble there. I also recommend starting small, single question, maybe 2. get the system working and the infrastructure right, you can always build it and overcomplicate it from there, that’s for sure. And the reality is JD Edwards has a lot of data and a lot of information. It’s just getting to it timely, that’s kind of always been the challenge. Like how quickly can you turn requests and how many conflicting priorities you have. that type of thing. So a couple more slides and we’ll go to Q&A. Just a quick thought here. You know, a project like this really requires really two different skill sets. One I think is JD Edwards World Expertise. That’s where Bright Skies fits in. That’s kind of a core competency for us. And then integration layer, and that’s where our friends from Ariadna fit in so nicely, just building that API framework. You know, kind of modern connectivity, whatever buzzword you want to use. Really, APS security, I think, is the best kind of thing there, and then those two skill sets combined provide a nice, complete solution for doing something like this. And then, you know, for me, when I was thinking about this… I really kind of come away with one kind of overarching thought on this and that, you know, those of us who worked in JD Edwards for more than a minute, right? We know JD Edwards already has the data and the answers to your questions are in there somewhere, right? So the opportunity isn’t creating more data, it’s making the data you already have. and the data you already mostly trust, or you trust, just making it significantly easier to access that data. So final thought before we get to QA. Once we get done here, I would invite everybody to, you know, if you’re at all interested in this, you know, let’s schedule what we’re going to call an AI discovery workshop, right? Let’s take a look at you know, maybe where your users are struggling to access some data or maybe identify a couple of potential use cases or find out where some existing reporting has some challenges, that type of thing. Love to schedule some time with you and kick those ideas around and see what could make sense for you and for that. So happy to schedule time for those conversations, whatever makes sense for you. in your JD reporting data. And with that, I’m going to open it up for questions. So Aaron and Nathan, I think you guys will have the questions the most. But if anybody has any questions, certainly happy to take those now. You can throw them in chat. I think people could come off mute now if I’m not mistaken. Matt, you’ll let us know if that’s. that’s possible or not, but any questions anybody has on how this worked, what we’ve done, that type of thing. Tim, I see you. You got a question there? Go ahead, Tim.

Tim Callari
In terms of the APIs, do you have to tell the AI how it reacts to each API? Like what, in terms of, hey, this is what this API means and this is how it links to another one. And also in terms of the fields that come back, do you also tell AI what each field means?

Nathan Giri
Help.

Tim Callari
In terms of like in the example, I saw something like category two, for instance. I mean, do you tell it ahead of time, by the way, each API, what each field represents and a description and all that?

Bill Onion
Mhm. I’m going to toss it over to Aaron and Nathan. You guys got an answer for Tim on that one?

Nathan Giri
So I will start with my answer and Aaron can I think add more. So the answer to that is there are a few ways you can solve that issue. So basically AI knows which tools to call. So basically it sends you a natural language question to LLR model. LLR model will figure it out. what API tools are available, and then based on the tools available, meaning what APIs are available for that MCP server, and based on that, it determine which API to call and what parameters to call. So, and then sends back to that request back to the agent, then the agent would call agent would instruct the MCP server to call appropriate API using the appropriate parameters. So that’s how it works. And also, in order to fix this, in order to make that as a permanent, you can have a skills and or in cloud we call it skills. You can set up a scale saying, hey, AI, if it is a category code, use these are available category codes. So AI knows which category codes, how to, you know, cross-reference it converted and what parameters to use it.

Bill Onion
Great.

Nathan Giri
Uh, Adam, do you want to add that?

Aaron
Yeah, I think that was really good, yeah.

Nathan Giri
Yeah.

Bill Onion
All right, guys, we got a couple of other questions in here. Next one is, when an AI agent calls a tool, whose identity does it run under in JD Edwards? Is it a service account or the end user? And how does JD role security carry through on that?

Nathan Giri
So I’ll answer that question. So when AI makes the call, so the system, so we know that who calls that, meaning we capture the, you know, the person who makes the call’s e-mail ID, in this case actually. So it’s coming through metadata. We capture that e-mail ID, our user ID, And then, you know, inside the we crack check with the IBMI profile and and and then we know who’s you know what IBMI profile that connected to, and then and then you see appropriate JD or JD profile and determine whether he whether he or she has access to the. in a business function or data.

Bill Onion
Great. Thanks, Nathan. Next one is, is security governance entirely reliant on the configuration of IBM slash JD, or can any of that be managed through Ariadna?

Aaron
Oh, sorry, I just actually answered that one in the chat. No, but that’s great. No, so I’ll just read off that answer. So, and you guys, please feel free to add anything if you want. But Aerodyne Connect provides a comprehensive security layer. You can absolutely secure systems directly in the integration layer. Many users also choose

Bill Onion
Oh, I see that now. Yep, sure, yep. Yeah. Yeah, great.

Aaron
to lean on the security of their core systems. It really depends on the requirements for your specific case. So, I mean, that’s really a, you know, a major place where, you know, where the Bright Skies team comes in, where they can help you out with that is making those architecture decisions about where, you know, where is the right place for me to do this. based on my specific requirements. That is a question to ask.

Daniel Magid
But Aaron, you can actually front end the IBM I security with a whole different security layer if you want to do things like OAuth, Okta, multi-factor authentication. You can do all of that at the Ariadna Connect layer before it gets to the IBM I.

Aaron
Yes, absolutely. And, you know, we have users who completely buy, you know, who actually do not have real security on their core systems, and that’s actually part of the problem they want to solve. And so they put up a comprehensive layer in front of it using Ariadna Connect. And there are also users who have a really comprehensive system on the back end, and they say, no, you know what, just pass it through. I trust my systems. and everything in between, and that’s totally fine.

Bill Onion
Yeah. Yeah. Next question for you guys. How are you guarding against prompt injection through the data that the agent reads?

Aaron
That’s a good point. Do you guys want me to take that? I can talk about that. Okay. All right, sure. So there are a couple of patterns actually for that. Correct me if I’m wrong, Bright Sky’s team. I think that that’s actually out of scope for this particular integration, because this particular integration is direct user sessions, right? So I think it’s…

Nathan Giri
Yeah, go ahead. Yeah.

Bill Onion
Yeah, like, I assume that was yours, Aaron. Yeah.

Aaron
primarily focused on that. But there’s two points here. One is in a direct user session where you have a human user live with the AI. The governance layer is providing a lot of, is helping you a lot with that because it’s kind of where I was getting at it earlier is it doesn’t actually, even if the AI. suffers from a prompt injection. No, a prompt injection is when you basically hijack an AI with malicious instructions. You think of it kind of like, it’s not the exact same thing, but in my taking valuables from a safe example, it’s kind of like slipping the guy a $100 bill, you know, under the table and saying, yeah, can you just let me in? right? It’s a way of getting around security systems and getting AIs to do things they’re not supposed to do. But this is why the governance layer is so important. Because if the AI is not authorized, actually let’s just use the safe example. If the guy doesn’t have the combination to the safe in the 1st place, doesn’t matter how much you pay him, right? He can’t get in. Sorry, right? And that’s why, that’s one reason why these governance layers are so important. Now, beyond that, for autonomous agents and those kinds of things, we can talk about guardian patterns and those kinds of things that you can set up with your agents. But I think that’s a longer discussion that we can all talk about in one of those AI sessions that Bill was talking about if you guys want to.

Bill Onion
Right. Cool, yep, I appreciate that. Right. Quick commentary, yes, we will send out the slides to everybody who’s registered and in attendance, and we will make this recording available as well. So give us a day or two, we will have that available. Next question, would it be okay just to have the API structure set up? restrict the API calls to read-only calls, then port it to the AI with a harness. Seems like that is the best way to get around security setup, especially just for super users. Thoughts on that, Aaron?

Aaron
Yeah, yeah, I mean, I mean, you certainly could. You know, it’s not all that different, I think, from what was done in this particular project. It’s just whether you want to use specifically an MCP tool or, you know, or just set it up in your agent harness, that you could certainly do either one. I just would, generally speaking, we want to be very careful about the security. You know, our policy internally at Ariadna, where we use a lot of autonomous agents here, is we make sure there’s a governance layer for everything, even for the super users, because we’re concerned about prompt injection. Sometimes you could have a super user who really, they’re a real AI power user. They’re using it every day. They’re maxing out their token limits constantly, you know, and they really know what they’re doing, but their agent could still. be the victim of prompt injection and it could end up acting on their behalf to do something it’s not supposed to do. And we have to be very careful about that. So we always recommend having the governance layer and it’s always in place for our teams, you know, because again, even for a super user, your agent could get hijacked. So it is a very important point. But yeah, you could, yeah, you could do it. You could do it with a harness.

Bill Onion
Great. Anybody else? Any other questions?

Aaron
Yeah.

Bill Onion
Couple of questions around pricing. Obviously, we’re not going to get into pricing today. Anything else? Anybody else? Now is the time. Got some really smart people on the call right now. Questions going once. Questions going twice. All right. I think we’ve covered it. So Nathan, Aaron, appreciate your time today. Dan, appreciate you being on the call. Everybody, thanks for attending today. Like I said, we will provide slides for everybody. So those will get sent out in the next day or two. We are going to have the recording available as well. Everybody wants to look at that. And I encourage anybody who wants to go deeper, happy to discussion some time and just talk. AI and how this works or how we could make it work in your environment or anything like that. That being said, I appreciate it. Thanks everybody. Have a great day.

Daniel Magid
Thanks!

Aaron
Thank you all.

Laura Ubelhor
Bye, take care.

Nathan Giri
Thank you.

Mitch Hoffman
Thank you.

Matt Laskosky
Hey, Bill, was there something else you wanted to do?

Bill Onion
No, I think I’m good. I think I’m good. All right. Thanks, Matt. I’ll talk to you later.