Building Reliable AI Skills in NetSuite

00:06Okay. Well, good morning everyone and welcome to today's LUXUniversity session. I am Brooke Gyepes and I'm the director of marketing here at Luxent and I'll be moderating our conversation today with Luke Hood. Today we're talking all things AI skills. So, what they are, how to use them, some real world use cases. And today, like I just mentioned, I'm joined by Luke Hood, our AI engineer here at Luxent, who's going to walk us through how we're training NetSuite AI to actually perform the tasks the same way every time, how you can check that it's working and trust the results um for your own day-to-day use.

00:40So, let's get started. So, here's kind of what we'll cover today. I'll I'll kick us off with some quick introductions just to Luke and his background, and then I'll hand it over to Luke to get to the get to the good stuff. He'll walk us through where AI shows up in NetSuite today, kind of break down the anatomy of a skill and kind of bring that to life with a little bit of a demo here. So after that, we'll open it up for questions. If you have questions as we move throughout the demo today, please put them in the Q&A panel and we'll do our best to answer them either live or at the end of the presentation.

01:12Um then or really quickly kind of before we jump into today's training, I just wanted to take a quick second on what LUXUniversity is all about. So, it's really about helping your team feel confident and empowered in NetSuite. Whether you're newer or looking to kind of level up your expertise, these sessions are really built around real world use cases with practical tips and best practices that you can always bring back to your day-to-day work. So, at the end of the presentation today, a survey will pop up.

01:41If you have any suggestions or requests for courses coming up, please put them in the in the um in the survey at the end of the at the end of the training here today. and we will we will do our best to incorporate them into our training program. So without further ado, I'd like to introduce Luke. He's got a great mix of backgrounds here, manufacturing, food and beverage operations, data science, all of the things. And he brings all of that into his AI work here at Luxent. So before he was in consulting, a fun fact, he was actually a professional brewer and data scientist.

02:14So he's really used to figuring things out and how that how to make them work better. On the NetSuite side, Luke is very deep in manufacturing, order management, cycle counting, integrations, and these days he's spending a lot of his time on AI. He has a master's in food science management from Washington State. And he has a certification in SuiteFoundation. And when he's not working, you'll find him cooking, got a live music show, or out hiking with his very cute dog, Murphy. All right, with that, I will pass it over to Luke to to get into the good stuff.

02:47Luke, take it away. All right. Thanks. Hi everybody. Thanks for joining today. Um yeah, today we're going to be talking about AI and NetSuite. Um in particular, we're going to be talking about skills and context engineering. Um so you can kind of learn how to utilize AI to give you the most consistent best results possible. Um when we're talking about AI in NetSuite, we're generally referring to like four different things. NetSuite Next which is kind of the new interface um that NetSuite has slowly started rolling out.

03:23Um that's where you'll find Ask Oracle which is kind of the chat assistant um where the large language model lives um and that we'll be interacting with um building skills with uh NetSuite also has a library of pre-built skills that you can utilize as well for that. Um beyond that they also have embedded AI um stuff like case summaries, text enhance, uh narrative insights, uh bank transaction matching. Um and beyond that as well, NetSuite allows you to bring your own AI generally either like Claude or ChatGPT.

04:02Um connecting to AI uh to your NetSuite instance through an AI connector uh an MCP connector. um that allows you to apply like a different large language model um but still get like the context from NetSuite. So enabling it to read your NetSuite data to answer your business questions. Um so we can move on to the next slide. Uh so we'll talk about using skills um within uh NetSuite.

04:39So as I was just saying there's kind of two avenues for doing so. There's Ask Oracle the embedded like AI agent within uh NetSuite um where you can apply skills too. Um and you can also do the same thing via the AI connector in uh Claude or ChatGPT. Um couple things to note with this. Um you'll be utilizing Oracle's model. um it will run as your role um in the session and the same thing goes for the AI AI MCP connector um it will always inherit like the role privileges that you've assigned to that and with that one in particular um nothing no role is like preset with those you have to assign those roles um explicitly um so you can't accidentally give access to that.

05:39Um with Ask Oracle there is um an Ask Oracle and NetSuite Next um preview like function that will give roles um access to those features. Um the data um Ask Oracle generally um utilizes reports um saved searches it'll look at the records um you have a couple other options with the AI connector and Claude ChatGPT um it can utilize reports and saved searches much like um ask Oracle but it can also run SuiteQL um generally if you're following uh Oracle articles uh skills packages for the AI connector.

06:26It um prefers like a hierarchy of those looking at reports as the source of the truth first then saved searches and then looking at the records individually and then using SuiteQL. Um there's different advantages to each of these um giving you like different reaches of data within NetSuite, but you do have to be careful and make sure you're not like misinterpreting the data. Um and there is um mutation of records available in some of these features. Um but like you generally have to enable that and like I said before uh those um preferences will always follow the uh the role applied.

07:14Um and then AI units um NetSuite is uh using a unit structure for distributing AI. Um, right now it's generally a thousand units per uh user and that applies to both sandbox and uh production environments. So if you're developing a sandbox um you can utilize those and your production user uh bank won't be impacted until you start using those in your production role. Um, NetSuite kind of estimates between 50 and 200 uh, units per uh, question asking Oracle.

07:54So, you do have to use it sparingly and uh, kind of be aware of that. Um, the pro side uh, with the AI connector, bringing your own AI service into it, um, you're not utilizing those AI units from Oracle. um you're on whatever your company's team plan is with either of your providers. So, you're not utilizing any of those AI units through that and then we'll move on. Um so, we're going to talk about uh what a skill is and essentially it's documentation that the large language model can read.

08:37Um, so it's in a file format called markdown. Um, the main thing is always going to be the name and the description up top or the first thing that AI reads. Um, with it, so like this is how AI picks the skill. You can also invoke the skill with directly within your environment usually through a slash command or I'll show you in a little bit how to uh call a skill directly um within Ask Oracle or NetSuite Next um beyond the skill if you think of it um this is kind of like a recipe book.

09:20So the context um that we're trying to gather from our NetSuite instance is kind of all the ingredients um going into a recipe. Um beyond that though, you need the instructions for how to utilize those ingredients to get like a consistent result to make like a consistent dish if we're following that uh recipe analogy. So, within it, we have a step-by-step instruction guide with explicit instructions of what to do when that skill is invoked. Other things that can be um included in this are like references.

09:53Um whether these be definitions, queries, or like report templates. Um you could also include assets like if you have a business logo or things like that. Um, and if you're creating skills within like uh Claude or ChatGPT, you can give like helper scripts and stuff like that. So when we're talking about um, you know, invoking skills to create repeatable reporting within NetSuite, scripts are a great asset because it kind of enforces guard rails.

10:30um it's going to follow the script um to a T and then uh it's going to execute that and deliver it in the same way kind of every time and we can move on. Um so what the description decides as I like mentioned before it's like one of the main um aspects of the anatomy of the script itself. um it's going to um decide uh when the script is used and what it returns.

11:08So like if you're vague in your instructions and your descriptions, uh in this example here, um the description is just helps with customer order analysis. Um it's not a great example because there's no trigger. Um it's not giving any suggestions on output, there's no boundaries. Um, and as Oracle has really nothing to match on. Um, you can still invoke the skill manually doing that, but it's never going to hook on to that skill and know what to do. Um, utilizing like keywords from like the description.

11:44Um, a stronger like use case would be um the following. uh used when the user asks how long order to cash sales take from sales order to shipment to invoice to payment. Uh which customer takes longest to pay and which stage is the bottleneck. And then it also gives the returns um returns median days per stage per customer uh compared with a prior period. Um it's going to call out that this is read only. That's kind of a guardrail.

12:21Um, just on top of everything else, like I said, it will follow the role-based privileges of the account and also the tools it has um, access to. Um, so you can monitor each of those as well. Um, it's also going to give you like a do not use for picking, packing, uh, or shipping timing. um it's directing you to another pre-built NS fulfillment performance insights skill um if you're performing any functions like that.

12:57Um so that's just kind of how the description um a good example of a weak one and how a good one can be stronger. Um and in summary of that it's when to use it uh what it returns and what not to do. Um if we move on to the next slide talk a little bit about context engineering. So the data the one of the best assets you your business has utilizing NetSuite is the structured like um data format behind NetSuite.

13:38It's kind of tried and true. Um, it's infinitely customizable. Um, and it's been around for years. It's constantly being developed. Uh, so we have access to this context um to, you know, get our business decisions from. So that's going to be like whether it be a saved search, a report, um just looking at records or like a SuiteQL query. Um we can retrieve that context.

14:13And you'll hear a lot about context in AI. Context meaning um the input like going into the prompt or into like the chat itself. All that adds up to tokens. So, a lot of times you'll hear that like too much context can be a bad thing, but really the important thing here is providing the right context to make that decision. So, not like flooding it with a bunch of unnecessary um information, but giving a real curated query that um is going to give it the right information. and then the skill which has the instructions on how to like parse, slice, dice and interpret um that information and what you get is kind of a repeatable result um with that.

15:05Uh we can go on to the next one. And this is just kind of summarizing um you know when we're developing skills especially in like uh creating business reporting in something like NetSuite you generally want like repeatable results every time you run that report. Um, so you don't want to give the model like the room to like kind of shuffle around be like, "Oh, let's look at it this way this run and a different way the next run."

15:42Um, so like to do that, you name the source. Um, generally you're um pointing towards a saved search or report um or like SuiteQL. And then as we get into it too, you'll see that um once we get into using the MCP connector, uh NetSuite has a script type called um custom tool sets. And essentially what that is is allows you to apply scripting and like SuiteQL logic um to a tool to retrieve data within that suite and then it's going to return it in a very like structured schema and format.

16:20Um the advantage of that is that the call and response is going to be the same every time. So we can kind of uh put that guard rail on um how the data is retrieved within NetSuite. Um define the terms um the business logic and the terminology that you're using for this. um you fix the scope. Um things like uh filter criteria you would see in a report, date range, subsidiary. Um if you're trying to like narrow in on a customer or something, uh you fix the shape as well.

16:59Um what data do you want to see? Um how do you want it returned um in like a table or like uh what columns are in there? Um how you want it ordered. um templates are useful for this part of it. Um and explicitly say uh what you're not looking for. Give examples of good and bad. Um give what you want to see, but also like give an example of like uh what you're not looking to see. Um in this too, um you're kind of guarding against hallucinations.

17:35Hallucinations occur when like AI kind of haywires and um tries to fill in um data. So by calling out um that you're not going to estimate a missing date, you're going to hold it accountable to that and just have it mark it as an exception and not like potentially give like bad data back. Um and then another useful trick is to kind of uh make the numbers balance. So kind of having AI check against itself using a different method.

18:12Um so like kind of reconciling its results at the end can help catch uh mistakes through it. And this is like particularly uh useful in like development of these skills. Let's move on to the next one. Um, and so we're just going to briefly talk over building and testing. Um, you know, in order to do this, you're going to have to get access um by asking your admin um to give you access to ask Oracle or um through the MCP uh connector within like Claude or ChatGPT.

18:54um you'll need access to that connector both in Claude and ChatGPT and they'll also need to download the MCP connector SuiteApp within uh NetSuite itself and then apply the permissions for that. Um the other thing to note on this is agent skills uh within NetSuite for ask oracle that record contains um uh kind of a permissions audience view. So when you develop a skill, it'll automatically like keep it private just to you, but you can expand uh that audience um throughout your organization as you develop skills um test them and do all that.

19:40Um and then you've got to do the work of writing the skill um and then testing it. Um generally like testing processes for skills is you test it with or without the skill. Um does the skill kind of add any like context or any like benefit to the prompt that is given. Uh you can ask it five ways um to ensure that it's properly reading those definitions. Um ask what it shouldn't answer. Um run against a known answer.

20:15um kind of a golden answer type technique and then you iterate through and you tighten and repeat that before kind of expanding your audience and distributing it. Now, I understand that sounds like a lot of work. Um, and it is, but luckily, um, AI providers like Anthropic ship a skill to make skills called Skill Creator. So, it'll kind of walk you through that process and do most of it for you. Um, so you can develop, uh, iterate through and hone these skills, um, to use it.

20:55So, it's a skill for making a skill, which is super convenient. Um, let's move on to the next one. Um, we're going to get into the demos here real quick. Um, I might just go over this slide. Um, just go over it. I've got two demo scenarios um to share with you guys. The first one will be um how to like invoke a skill within Ask Oracle and how to like how that can be used to like generate a report um and have it retrieved for its canvas.

21:40And then the second one will kind of branch out of NetSuite a little bit using Claude. I have a skill built in there and see how we can add even more context and instruction and uh refine uh different reports and grow that. So like in that demo we'll see um how the skill is used. Claude's going to pick that skill. Um I've also added to that a custom tool and um if we recall from earlier a custom tool is just going to be like uh example of like script or like SuiteQL um so a curated search that you're putting within NetSuite um and then returned in a very like specific schema um that the AI model can then um kind of interpret.

22:31So first it's going to call the skill, then it's going to call the custom tool to retrieve the data. It's going to change that um to preserve that. Um and then um I've expanded beyond this to uh like parameters that are going to kind of parse, slice, dice, um do a couple calculations within it. Um that's utilizing Python um a Python scripts uh that Claude will use to do all that put it into a format that can be applied to a template um that is kind of um saved in that skill package as like an asset and template.

23:16Um, but before we do that, um, I guess we'll cut over and I'll give you the little demo within NetSuite. And it's essentially going to be the same process. Um, let me just set this up real quick to share my screen. All right. And here we have the um, old uh, NetSuite interface that we all know and love. Sorry, I've got some stuff popping up on here that you probably can't see, but it's blocking me. Um, so this is the classic kind of NetSuite interface.

23:55Um, NetSuite uh has kind of started rolling out features within NetSuite Next. Um, so there's a couple settings that we have to like go through in order to figure that out. If we go to setup company and scroll down, you'll see this AI portion has kind of been um added recently. NetSuite Next access is what we're looking for. We can see in this account um I've kind of added the optional um setting for all of the roles we have like within here.

24:36So, all of these can flip-flop back and forth between the classic view and the new Next preview. Um, if we like check mark it, we have the option of changing this and then you would just save up here to apply those preferences um to that role. Since this is already done though, we don't have to do that. The portal for getting into NetSuite Next can be found um just how you would switch between roles or uh instances within NetSuite. So you'll see this little banner up here for NetSuite Next.

25:15When we click this, it'll load up and it will bring us to the NetSuite Next interface. Um it's just kind of a different structure uh way you interact with NetSuite, but it's still got the same like NetSuite architecture um behind it. All the processes are still the same. It's just this new updated look. Um Ask Oracle the assistant is located in the bottom right corner down here.

25:48So this search bar would behave um closer to um the global search that we're generally like familiar with at the top center. Um so just don't confuse those two. This is going to pop out the sidebar um or this. So when you're trying to invoke uh skill or add context in NetSuite Next, we see this little chat box down here. If you click on this plus sign, we have a couple options. You can either attach like exported data from like another source.

26:29Um, you can apply a skill, which is what we'll be doing today, or we can also apply context by adding it like a record type or like a saved search or report or something. Um we'll click here apply a skill and we can see that um NetSuite has uh a number of skills that they've kind of prepackaged for us. Um different use cases spanning different functions within it. They all have this NS prefix. Um for our demo today, I created this Lux order-to-cash lag skill.

27:05you would just click this and you would see this little button up here. Um, you can remove it by removing that skill and then you would uh type in a question there. Um, since this is the demo, sometimes these take a little bit to run and especially since this is kind of a preview version of it. So, I've kind of uh through like TV magic, I've pre-run one of these so we can take a look at it um without having to wait around for it. Um so asking like a simple prompt like where is our order to cash cycle stalling and then giving it the context of like uh compare sales orders from June through August 2026 with March through May and then show the results in a canvas.

27:57So specifying how we want the data return. So it's going to like go through a number of steps um to retrieve it and analyze that data and then the result like we asked for is a canvas. Um these can be um exported in PDF. Um there's some um interactivity through them. We can scroll over um highlight stuff like that. And these are also hyperlinks to the specific record that we're looking at.

28:30Um, so the skill itself kind of outlined what data we want to look at, how to retrieve it, and then how to process it. Um, and then fed it into kind of a template. Um, and created this report for us. um that in order to do this, we would have probably had to run several saved searches or reports and then process it, kind of dump it out into Excel, process it and interpret it that way.

29:01But like with a properly defined skill and package, um we have this like repeatable format that we can make for this particular result. Um this is great and this lives within NetSuite. Um the next demo I'm going to show is kind of how to do the same within Claude. Um so I've created a skill within Claude and very similarly um to what we saw the prompting in uh the chat box with within here in Claude.

29:46Um this is our initial message. So in here um instead of the plus sign will still add a skill um we have a different list of skills here um that we can invoke um and with Claude you can also invoke them through a slash command like this and start typing and it'll start looking those up for you and it'll also give you the um the description associated with that skill. So we can see when we um input this prompt, the first thing that Claude did was went and pulled up uh the skill file for it.

30:28So we have a detailed description in here. Um we're outlining the goal. Um this skill has like several other files within it. I have a report template HTML. Um, I've got a couple assets um, outlining um, just like business assets like a logo. Um, and then I've also got the Python script and that's going to validate the trace. Um, it's doing all the arithmetic like associated with it. It's validating the results and kind of uh, formatting in a way that can be fed into that template.

31:08Um beyond that we also have the definitions in the references. So every term that's kind of involved in using the report. Um we kind of define and um state how they should be used. Then we have a step-by-step process um just taking Claude through everything we need to do to build this report um and then rules associated with it. So we can see that like I'm sure many of us have interacted with Claude or even used the MCP connector for NetSuite and kind of experience where it's Claude starts searching around and trying to decipher what you're asking for it.

31:52Now it has all that correct context and how to do it. So, if you've ever looked, you can blow these down, and it's just going to show you all the steps it took. And I won't get too into this because it's kind of like dense, but we can see that it's like going pretty direct towards this final output. So, when we open this up, make this a little bit bigger, we can see this like very structured report. Um this particular example is a lot trace.

32:28So the custom tool I use applies some recursive logic and um searches um inventory uh assignment numbers within NetSuite to figure out like where um certain ingredients were used in like a manufacturing. So we have this like way flashier um report built within like a couple of minutes. Um it's fully interactive. Um we've gone through and tested it multiple times with Claude um to kind of optimize that script and make sure everything looks good.

33:03Um we have details of shipments, production orders. these will still hyperlink out to um that particular record within NetSuite or like transaction um the type of trace it is um and all these like definitions that were kind of defined within our skill itself beyond that uh NetSuite also like or I'm sorry Claude also identified anomalies uh with this um so going beyond it um since this is like a demo data set.

33:39Um, in this account, we have some weird looking data within it. Um, these builds all occurred after these item fulfillments. So, Claude was um, smart enough to note those as anomalies and flag them to look it out. Also gives us this nifty preview of, you know, where everything is going. And then um this is kind of that final check, that reconciliation um where it's looking at this in two different ways to make sure it's all balanced.

34:12This is just kind of Claude checking itself against Claude. Um so that's kind of the two demo examples I have for this. Um I guess I will stop sharing and give the screen back to you, Brooke. All right, that kind of brings us to these five things to take away with. And kind of the goal of this webinar um is trying to get those like wheels turning about your specific like business processes and questions, what data is important and stuff.

34:50Um find like one question that your team asks every week or like a pain point in your process where like say you have a difficult um report that requires a lot of human like um you know extraction and manipulation between multiple data sources to kind of um whittle down and create like um usable business data. Um, so go through that. Um, and just remember like the description um is what gets the skill picked beyond that slash command or directly invoking it.

35:26So like spend some time um writing uh in-depth like description um based on those rules we went over today. And um you're going to need both for um repeatable results. you're going to need um the same data sources going into it, meaning the context that you go into it um and then the same rules and instructions that define those terms and everything that goes into it. So, you know, like the detailed skill that we went over before. Um you can also just check the agent skill permissions this week just to like familiarize yourself um as an admin in NetSuite with how those work.

36:07I would also recommend even before creating your own skills in it, go through and look at um the Ask Oracle preview, invoke some of the skills that NetSuite just kind of like packages up and provides you um as is because there's a lot of great stuff in there and um just poking around will kind of lead to more questions. And then um just familiarize yourself with the um token usage, the AI usage within NetSuite. Um just so you're familiarizing yourself with how that whole system works.

36:44Um and then be aware that you can also bring your own AI tools and expand the capabilities using the AI connector within NetSuite. And um I don't think if I mentioned before if you aren't familiar with that AI connector MCP, it's just a SuiteApp that you download and it is free. Um so all you have to do is pay for your AI service um through your business. Um and then just bring us a question that your team asks every week. Um and with that, I'll turn it back to you, Brooke. >> Thanks, Luke. Um, I'm going to open it up for Q&A, but just really quick quickly before, um, just to kind of hit on the MCP connector, we did do a LUXUniversity a while back when the MCP connector first came out on how to connect that.

37:32So, um, if anybody is interested or needs help, you could find that on our YouTube page or on our our Luxent website if you needed to reference back to that or reach out to your account manager to send you that video if you'd like some more help. Or our team is also very very, you know, open to helping you. Um, and like like Luke mentioned, you know, these skills are something that, you know, certainly can be built internally or can be, you know, use the Luxent team to leverage, but something that we've we've seen is, you know, thinking about what is something that your team is asking quite often or or always having a challenge with or or something that, you know, you can't get to super quickly or easily through NetSuite.

38:11Um, a skill is often a great answer to that. So, with that, I will open it up for questions. I do see one question in here. So, first I'm going to ask this question and see what else we have coming through. Um, how hard is it to create a skill? Is it something that's very technical process or is it something that um can be created as a report in NetSuite and then ask AI to analyze it or create summaries and graphs? >> Yeah, that's a great way to look at it. Um, essentially like you do want to have like as detailed of a flow path that you can go through to create that report.

38:47Um, but like if you have a report already um, used and the context that you're pulling from NetSuite already um, that's a great place to start um, in building that skill. And um I do want to remind you about the Skill Creator skill within um something like Claude that will kind of walk you through that and help you through that process um by like prompting you uh what are the steps we're going to take, how can we make this more repeatable and then kind of doing that bench testing of improving the skill.

39:21Now like you may have to like iterate through that a couple times. So, like you may go through that process and then get into using that skill and it's not quite where you want it, but you can always restart that process um by like going through that Skill Creator skill again and refine that skill um to get it exactly how you want every time um before you distribute it out to your team. >> And Luke, is that something that we could help with as well? Like if you know someone, you know, teams can get so far and maybe we can help them complete the skill task or Okay, >> I'm sure we can help with that.

40:03>> All right, so we're open for any other questions as any other ones come through here. Um I'm seeing a couple come up. So can um or is there an an audit trail on any of these skills or AI? >> Uh yes, I believe uh the tool calls within NetSuite. um do show up in script logs. So there is an audit trail um with it and I believe just going off memory but when you develop skills um there is uh audit trails of who updates those skills within that suite as well.

40:42Um kind of similar to how you would make like a saved search or something you can see what fields were changed like through the system notes and stuff like that. Okay, great. And what about you know moving things from sandbox to production? Is that this question doesn't have a ton of context here so but just moving things from sandbox to production is that you know something that requires tokens or is it something that is can be done with AI? >> You know I'd have to check.

41:19Um the skill is kind of um it works as like a record and then stores your skill those .md files within your file cabinet and kind of makes that connection there. Um similarly to how like scripts and script deployments work. Um I believe there is a copy to account that you can utilize to move from sandbox over. skills can also be moved um through NetSuite's SDF which is a little bit more technical but also um the script record itself is literally just naming the script and then uploading either like the zip file with all of the your SKILL.md your folders for your assets your references stuff like that so it's really just uploading a file um and then creating that uh skill record.

42:15So, it's not super labor intensive either. Um but there's a couple ways to go about doing it. >> Okay, great. That looks like it's the end of our questions here, so I'll wrap up really quickly. So, thank you everyone for joining us today. We really appreciate having you guys on these LUXUniversity calls. As I mentioned at the top of the call here, uh there will be a survey that pops up after, you know, we close out this presentation. We do ask, you know, for feedback on on how the presentation went, you know, if there's any other future topics that you would like to learn about and then if you would like to set up some time to kind of learn a little bit more about skills or how we can apply this specifically to your team's NetSuite instance or or data sets.

43:01So, if you do have any questions, please let us know and we would love to hear from you. Thank you everyone. Have a great day.