From Prompt to Production: Hands-On AI Workflow Training with Stonebranch

Hello, everyone, and welcome to the fourth and final session of StoneBranch UserVerse online 2026. We're excited to wrap up our four part series with today's session from prompt to production, hands on AI workflow training with Stonebridge. Today is a hands on session where you'll see how AI workflow builder, AI integration builder, and Robie AI help your teams create, build integrations and troubleshoot automation faster while maintaining the control and reliability enterprises environments demand. My name is Lauren Tanzini from Stonebranch, and I will be your moderator today. Before we begin, let's cover a few things. This event is designed to be interactive between you and the speakers. If you have any questions, please add them into the q and a tab on your dashboard, and we will answer them at the end of the session. Additionally, in the right hand corner, you'll find some handouts related to this session. With all that said, it's my pleasure to introduce our Stonebridge speakers for today's session. Gwen Clay, chief product officer Nils Buer, vice president, solution management Hussein Golubescozoglu, senior solution engineer, and Len Seaman, solution engineer. Let's go ahead and get started. Quinn, take it away. Thanks, Lauren. So I think it's no secret that, now this is the fourth user verse online session for 2026. Actually, fourth and last. And that our... A lot of our focus here at Stonebranch is really around AI. It's not that we... We're not continuing to do a lot of other customer request enhancements, but that really the focus for the the major, things we're delivering on a release by release basis is really around making sure that we're ready to help you automate AI in the most reliable, secure, and governed way, moving forward. And this session's gonna be be, you know, nothing... No difference as well. If you look at our portfolio of AI products, it's kinda... Of AI capabilities. It's kinda split in two areas. Right? So there's the features that we are building within our platform itself, like Robie AI, like the workflow builder, And then there's the the broad spectrum of integrations we're building, which are allowing you to connect to the world of AI. All the, you know, all of the little, connections out there to the LLMs, the, the the agentic platforms, the AI gateways. Right? And all of this stuff which makes up allowing you to automate and orchestrate across the AI ecosystem. So we're gonna show you you you both today. First off, we're going to go through the AI integration builder where we're going to have Len take us through and talk about the integration builder. And then we're going to have Nils and Hussein really take us through some great examples of of orchestrating AI agents and then some of the other capabilities, you know, with SAP and the root cause analysis as well. So we're gonna start off looking at the AI integration builder. So take it away, Len. Thank you, Quinn. So as you've seen on the agenda, this two part series inside of this webinar because we are doing a life generation and enhancement of an integration during the demo, and we will come back later to show you the finished product. But first of all, the AI integration build us a new product, which is... That needs early testers. And if you are interested in becoming an early tester for for the early adoption program of the AI integration with that, you can check out the the pop up in the bottom right screen and become... If you're interested and check it out for yourself later on. But let's not... Just let longer be on slides. I will... I made a small example, a small scenario that may... Some customers may know. For example, you have multiple Linux or Windows tasks with specialized scripts for a specific scenario that you want to execute. For example, I have a Linux task here that does one specific Postgres query, and it has a script or command inside of it, and it executes one thing all of the time. And we don't have a specific task in the controller yet, so you just use the Linux task or Python script inside of the Linux task. But this can get tedious, for example, if you... For for the Postgres example, if you want to change database names or you want to change credential names and you need to use variables or anything else. So and you see there's no specialized integration for this yet on the integration app, so we can use the AI integration builder to generate this one. And this is what we're gonna do today in this demo. So I want to generate a Postgres integration that queries the database and creates a CSV from it that we can then use in the workflow to process it later on. So if you head over to the integration beta task, I already prefilled a task so we don't have to to waste some time on that. And if you head over to the details, we have our actions... Different actions. This is the generate integration. We fill it with the universe controller URL of which controller we want to deploy this one as well as the credentials for it in order to to execute on it, and the LLM key. For this use case, we will be using Cloud, but later on, we cannot... So you you can also see that you can use any LLM that you want to. We we both support Anthropic and the OpenAPI... OpenAI compatible models. I gave it the extension name PostgresSQL, and I have made an requirement script just lining out what essentially my Linux script does. It equate to CSV from a database name needing the port database connection and the credentials that we already have in the controller. And for a second action, to show you that you can have multiple actions, of course, we have a check connection as well. It's a simple it's a simple test connect action. I I also gave some line outs, and it's pretty in-depth, the script. Of course, the more information you put into the LLM model, the the closer it can get into into a working finished product in one shot. And then later on, once the generation... The the integration is generated, I want to, of course, have documentation as well. So I can click this checkbox for generating a documentation alongside it, and it will be delivered to my delivery method of choice. In this example, I chose GitHub, so there will be a GitHub rate created for the extension under the extension name and everything from the integration on towards documentation will be available there. We... The integration also features run tests, what what we showed you on Tuesday's session. We will not be doing this here. We will be doing this in part two manually to show you how to how to finish integration. You can easily just pop in everything, and then it will work. And, of course, the LLM provider will... Is below here. And if you want to choose an OpenAI compatible one like Olama or OpenAI as well or any of the local LLM models that you want to use that are available for tool calling, you can use this as well with this task. So let's kick it off with the launch. And while this is launching and running, I also want to go back to Tuesday's session essentially where we generated the GitHub's action... Actions integration, and I noticed that I was missing something. So I can use also the integration builder to click... To use the action enhanced integration to to update the integration directly from my universe controller instead of just regenerating it from you. For this, we have the exact same initial conditions. And for the extension name, we can query the controller directly and choose the extension from the extension list, GitHub actions for this example, and then give an enhancement description. So here, I gave an enhancement description for downloading run artifacts, essentially. So after a GitHub workflow has been done, it creates a a file and a ZIP, and we want to download this on the agent as well. This this could help us with verifying or moving the data to another place afterwards. So I will also launch this one. So we have both of them running at the same time the activity screen. We see both running. And while this is happening, I will head... Over to Nils where he shows you how to use AI and workflows. If I... So if I go back to the screen, and here you go, Nils. Yeah. Thanks a lot, Len. So let me quickly wrap up what we did in the last session. So the last user of our session, you saw how to create workflows in USC using the Ruby AI workflow builder by David. Today, I want to show you how you can use AI when running your workflows. The goal is to show how AI can help automate operations, simplify troubleshooting, and make your data pipeline smarter and more efficient. For this demo, I've created a sample pipeline that processes sales data and uses AI for different tasks. For example, you will see how AI can automatically create a ticket in ServiceNow, help us analyze and fix the errors, process sales data, and create a report about executed tasks. I will also show how to run an AI agent pipeline in Google as part of the workflow. Yeah. Let me start the demo by uploading some new sales data to SharePoint and then see how the AI works. So I just I just pop in my SharePoint. So this is my SharePoint. So I... From the SharePoint, I will upload the new the new sales data. And then we have a SharePoint monitor here that will identify the data in the folder. And as soon as it has identified it, it will... Should go to success and then kick off the next process. Now it has kicked off the next process, so it will download the data from SharePoint. It will now download the data to the SAP application server. And from there, we load the data into the new HANA data lake. So this new HANA data lake task is now quite sophisticated. So it supports not only loading the data into the... In... Into SAP HANA data lake. It also supports now the SQL engine, so the SQL engine. So you can do both things, uploading data plus all the SQL commands. And then what you see here... So after the data's loaded, we start an SAP ABAP job, and you see this ABAP job failed. So what I want to show you now is how AI can identify the phager and open the ticket and also analyze the phager. So as soon as there's a phager in our controller, it is... It generates internally an event. It's called a universal event, And this universal event, we can capture with a simple webhook that you set up. So I just go to administration, set up a webhook, give the webhook a name, then I just say which events I want to listen for. In this case, I want to listen for the event task status change. And if this has the information running problems, then I will start my AI task. Of course, these criteria can be very sophisticated. So you might not want that every single task generates an incentive ticket. So you can have here very complex rules, so maybe only for certain workflows or only for certain business service you want to open this ticket. So in this case, I say for every phaser, open... Run the Kong AI task. So now let me show you what the Kong AI task does. But this Kong AI task is different than the other task we had before. I think maybe you know we had already the LLM task for Bedrock and for Azure OpenAI, but those tasks were more like a prompt. So you were asking questions to the LLM, and then the LLM gave you answers. This task is more. This task, first of all, can call the LLM with standard standard prompt, but then it can also call tools. So it can can call tools from any number of MCP servers. So in this case, in my example, I connect to the GPT four dot one model. Of course, you can have more models. So you can have all the models which you are allowed to use by Gong, and then you choose the MCP server that you want to use. In this case, I want to connect to Dynatrace because I want to send... Also check if there was some... Something noticed about this error in Dynatrace because Dynatrace does logs and traces, so it might also have some more information on this SID pager. Then I send the stuff also to the MCP server from Stonebench. It's called the user verse because it's how I called it in com, but it's the standard MCP server from the controller. So this one checks for the log file of the task. And then I have ServiceNow, so ServiceNow MCT server, which I will use to open the tickets and attach all the information which I get that from Dynatrace and from the log file from the universal controller. So what is Kong actually? So Kong gives you... Kong acts as a central place to connect different... The different AI models, like, in this case, GPT or Cloak. Instead of connecting each provider separately, you can now send requests through to... Through Kong. This makes it easier to manage access and monitor units and cost and switch between the different AI models. Another usual feature is that you can see here that you can have multiple MCP servers allowed to interact with models and other systems. And one thing which we can also do here is the allowed tools because maybe you don't want that all tools are allowed to use. Maybe I'll show you the... Show you the tools. So these are the different tools. And if I don't select this option, any tool is allowed. But maybe I just want that this user is allowed to only... Let's, for example, view incidents, retrieve log files, and not able to create tasks. Then I can control this by here. So, like this, I can really control which tools can be used by this task. So... And finally, what you do is you have your system prompt that defines basically how your LLM will react. So in this case, I say what it should do is so it should investigate incidents to access stone stone bench data trace and service now towards. Then in the user prompt, I define what actually should be done. So I tell, for example, that it should create the ServiceNow ticket. In the ticket, we should have the task name, insert ID, search severity, Dynatrace findings, and so on. So everything... What's actually I want to be in the ticket and the information that should be created in ServiceNow. So so as you see, it's it's really like a normal chatbot, plus you can do all the tool calling. Okay. Let me quickly show you... Just give you a very quick view on the Kong AI task so that you know what Kong actually is. So this is a a Kong gateway. So we support both. We support the enterprise gateway, which is hosted by yourself, or we also support the SaaS gateway. This is the one that you see here now. So there is a so called AI gateway stone bench configured, and then it tells you which MCP servers you can, for example, connect to. So the stone with MCP, ServiceNow, Slack, and so on. And it also controls which models you can use. I'm... Although have these... I only have these two small models here to save some costs. So this is... If you have another AI gateway, then just tell us, and we are trying to also build something for that. We we choose Kong because Kong is, at the moment, one of the market leaders, and we have also some customers already using Kong. Okay. So now let me let me see if the ticket already has been created in ServiceNow. So I just go to ServiceNow. So... And you see, there is the incident.sapcom task incident fails. And you see there is... There's a pager. So SAP ABAP, a pager. ABAP list not found. So so what I will do now, I will now analyze further this error. Of course, there is already some information, but this information is mainly the information from the log files, and I want to really analyze it more in... More deeply, this issue. So what I will do is I will have configured, in this case, my CluedCode chatbot. So... And I I did this because I want to show you how you can interact also with the CluedMCP server and UAC. Maybe let me do it like this, and it's a bit more nice to see. So so with Clot, I can just connect to the MCP server Stonebench, same as I can do with, for example, SAP Jewel, and it works really nice because it has the same functionality, basically, than our Robie AI, which you saw on Tuesday in the user verse. So let me check. Why has the taskzapcom start build. That's up if he knows it. So now it should connect to our MCP server, and you see it already start the tool get tasks instances. And it says... It tells you, SAP rejected the variant on step u s c connect as a declined. The step runs up above list list d. SAP returned list d. Above the list does not exist, and jobs variant list is invalid. So what I will do is netfix change the step to list p. So it already knows what the... How to fix it because from the past from the past, it knows that this issue appears more often. Because when I promoted this task from the development environment to the production environment, I forgot to rename the variant. So then what I just do, I just say change change the variant to this p. And should I directly rerun? Let's let's directly run and rerun. We run. Of course, I can only do the stuff where my user, the one that I'm using here include, has the access to. And you see it's it's very quick. Right? You see it immediately started. So it corrected the the other program. It has now the list the list p as step. So if you go to the steps here, then you see it has now this p. So I just changed it. Yeah. So the... So once it's... Now let me go back to the float. Now it's already successful. Now it is successful. So let me close the ticket again. So close the... Or maybe first, let's check. Check. See the log log output. Let's first check if the log output was correct of this task. That looks good. Yeah. So I already see job finished. So everything was fine now. So now I can close the ticket. Close the tickets to close the incident Didn't in service now. Intent is now resolved. So if I go to service now and press the... It's showing up for refresh, it says resolved. So it worked perfect. So what you've seen is ticket was automatically generated. Log files were retrieved by the... Through the different MCP servers from serve... From Dynatrace and from Stone branch, ticket was opened, and now then I showed you with ClotheCode. Of course, instead of ClotheCode, I could have also directly did... Done it with Ruby AI, then just with the analyze functionality, but I wanted to show you all the connection with the MCP server. Okay. So now some other new... Cool new stuff that we have. So what I did, the file which I uploaded to the SAP server is now here passed with the standard Linux task and then sent to a new kind of LLM task. So I think you know that we have the standard LLM task that can call Bedrock or OpenAI models from, like, GPT GPT five dot five, for example. So this is this is a standard task which we have for a long time. But now we have another one which uses decision LLMs. So since decision LLMs have the nice advantage that they are very fast and very cheap from the number of tokens. So we... Len and I did a lot of tests, and we never came over 0.001¢ per execution. So it's it's it's really cheap. So what the decision does, it just decides on certain things. So in this case, I asked the decision LLM if if I look at it. Maybe I'll look at here, then it's a bit nicer to explain. So the decision LLM are special LLM types. So the the most known in the market at the moment is the one from Jeff, but there are also other decision elements out by now. But we said we should build this prototype with the first one for Jeff. So instead of writing in a prompt, we have all of these JSON structure. And the JSON structure, you ask what you want to do. In this case, I want to check if the sales information contains personally identified information. And if yes, Jeff is going to vote it. So it says, is this true with a with a very high possibility? Then I would say exit this task with an exit code 20. If I say it's a low confidence, so there might be some data in it, then we do a 10. And if it's for sure no person data's in, I return a zero. So so it does not return any text. It just returns an information about the result here. So this Jeff model, if you look at the Jeff model, it says it went to success, but it's went with an exit code 20. So there was personal data in the sales in the sales data, which is normal. So the... This is why I then start the AI masking task. For this AI masking task, that's the one we we had already for a longer time. That's just the standard LLM task that starts, in this case, g... That use the GPT form mini model where you can use any model where you have access rights to in your Azure OpenAI. So here, have Azure OpenAI credentials, and these are the models which, basically, my user is allowed to use because our administrator gave me these models to use, and he's looking very much on the money. So you see I have only access to the cheap models here, but they they work fine. So then also for the... It's the same thing. I have a system prompt where I define how the how the how the element should react. So in this case, I'm telling him what to do, that it should anonymize the data and how to anonymize data so it put this just in clear text. Yeah. And then the analyst data is saved on the disk. And you've you've seen it already has analyzed the data, but if I just look at the output, you will see, though, the data is the the data is already anonymized. So it's... The number is 000, and the street is just a redirect street. So some... Yeah. Some some arbitrary street. Yeah. So so in this case, I want that... Now I'm able to send this data to the Google AI training pipeline because all the anonymized data is gone. In in in case Jeff says, oh, everything... There was no anonymized data, then I would directly send it to the Google AI pipeline. There's one last step. I want to approve the thing because I saw that everything was anonymized, so I used the the approvals step in my universal portal. So this is very nice. You can do this from your mobile, and then I can say, what's this? Approval required. So I say approve, confirm, and then you see it's already approved. So this goes all in real time. Yeah. Now the Google AI agent pipeline is running. This pipeline has the advantage. This task has the advantage that it shows you every step that is happening when starting a Google AI pipeline. So if I retrieve these outputs, you see it has already... It's still pending the preprocessing, but then it will do the training primary model and the shadow model. So I need to wait a bit because it's it's all real. So let me check if it's not running. Otherwise, I'll do it at the end. No. So it's it's still processing. So it's still waiting. Okay. So do not debate on this. Instead of Google AI pipeline, you can also start all the other different tools, for example, as a data factory, and and Databricks, as a as Synapse, Informatica. So all the different pipeline providers, I choose the Google AI agent one because that's pretty new one, and it's really related to AI. But all the other pipelines, we we have also task for. Okay. So one last thing I want to show you is the Kong AI task. So we started... Or we processed all these SAP information here, and at the end, we created another Kong AI task. And this Kong AI task is also nice because what it does... I told this Kong AI task, please select retrieve all the task instances, the log files from them, from all of them which have been run-in this process, and then create me a report with the task name, task instance, how long they ran, if they have been restarted, what the status is, and then send all these... Send this whole report to, in this case, the Slack. So what I need for this, I need for this account task, and I need for the two MCP servers, the one from from Stonebridge again because I need to retrieve the log files, and the one from Slack to distribute the report to Slack. And, of course, I need to have a model. In this case, I use again g p t four. It's a cheap one, but it's fine for this task. So this task has already been finished. So if I look at my Slack now, so it should have generated... There's a new that's German. Sorry. No. So as you can see, this is the latest that I got. So it says workflow and oh, let's say this part, the number of tasks exit... Oh, that's... Sorry. That's not the one. So here it is. Okay. So it says executed successfully, so it tells me which... This file was copied, then the HANA data lake. So the files that are copied to HANA data lake is here. Then the SAP, and here's important that the attempt count rerun two is also there. So it also shows you that this task, because that's the one that we... With Claude had rerun was there. So the whole report is available then in Slack. And, I mean, what I wanted to show you mainly in this task is that you can do standards chat talking plus tool calling on any number of MCPs that you can put in this thing. So the last thing I want to show this is still work running. Ah, yeah. Now it's running. You see? Now the training model is running, and the... Also the people's end dataset was done. So you can really follow any step that's happening in the Google AI pipeline with the with the task. Yeah. With that said, let me hand over to Hussein, who will show you now how Ruby can learn from user actions. So that's another nice thing that you can also use Roby in all these contexts. Hi, everyone. So today, I will show you something. I think this is really small and enhancement, but it is... I believe it's it's useful because... And we we use Robi to analyze the the task for us, but, also, we need to teach Robi how to analyze our task. So we we need to tell how we want to resolve the issues. So in in this case, like, I will do a very simple demo. So I have a I have a task here. That task is running an incorrect command. Instead of l s, it's running an l s s. So normally, we expect Robbie to say that it failed because it's... There's a typo there. But maybe I have a I have a a task called l s s, but the the path was not correct, or maybe the parameter was not correct, or or there can be some issues. So if I see the failed task here, if I do an analyze now here, so the... Robbie will say that it failed because of this one. This is because this this is all the LLMs will will tell you the same thing. It says that, oh, you have a typo here. So it should be like this because this is the most expected solution for that one. But for my case, I have a task under the demo folder. So in this case, probably, I did a typo in the in the path. So let's say that I have a application like this. So let's save it. And then if I rerun it now... So in this case, instead of Robby, I did the fix. So I want Robbie to learn it. And next time it fails, I want Robbie to tell me the correct way to fix it. So now it's successfully completed. But if I go to the notes now... So this is the mechanism I created. So I created a a webhook that checks if a task failed and and afterward is successfully completed and with... A rerun. We know this by checking the attempt number. So if the attempt number is bigger than one, and if the status is success, it means that it failed and someone did something, and then it... It's it's successfully ended at the end. So the task I created is reading the the audit logs, sees... Checks the... All the changes, and also put this information inside the inside the notes of the task automatically. So if... This is like a small script. I I create an extension for that one. I can share it. But you can write a simple script like that for that one, or you can come and manually write your troubleshooting steps in the notes section of the task. So in this case, next time the task fails, Robbie will know if I launch again. So it failed again. So now the difference is now I have the notes how to... How I troubleshoot in the previous execution. So now if I do the analyze now, Robbie will read the the notes. And then now the recommendation is instead of changing it to l s minus l, so it says that you need to do this. Because in the note section, also, it describes some some places here. It says that in the note section, I see that the previous executions, and this was the solution. So now, Ruby gives you the the suggestion like this. So if I do it now like this, this will also be, added to the notes. So I said that it failed two times, and this is the... Also, the solution was same. So if I go here, I see that it's successfully completed now after rerun. In the notes section, I have the notes again. So it says that, like, a... Now it says that it's two times it failed like this. So after some time, if you have the similar errors or that kind of thing... So my task, I I also said that give me some recommendation. Because this time... And, like, if you still... If you are having the same error multiple times, so maybe it's a good time to fix the task definition. So it also gives you an... Another notes here for the the AI recommendations. And all of that kind of notes, you can see them in the report section. So you can have one report to show you the... All the troubleshooting steps. So it was running in the system. So it learned also. You see here, like, it it learned from... For the SAP task, cong task. So there were some errors. Someone fixed it, so it put the notes how they fix it. So the system will automatically learn from this kind of manual manual fixes. And also the same thing, like, can have another report for the recommendation. So based on this recommendation, you may update your tasks. Like, for example, in this case, it says that update the task definition, update the comments from this one to that one. So if you go to the task definition, if you change the the the the the comment as recommended, so we expect the task will not fail. And how I implement that one is, like, it's, like... First, I create the webhook. So the webhook I created is, like, a learn from user actions. That's a name here. So the thing I'm I'm looking is, like, a task status change. So I'm looking, like, whenever the task status change, I want to have an event. So I am looking if the attempt number is greater than one, which means that it failed and there's at least one rerun. And then I say that if the status becomes success and if it is not a simulation, so we don't want to learn from the simulations. And also, I add a a business service here. So... Because I I want to exclude this task from this this webhook. So I said that's like a... And I I have a business service added to that task. So if this task fails, we will not learn. And the task we are running with the webhook is... This is an extension I created. So, again, you can create a a small script for it, or if you want, I can also share this one. And, like, this is doing such thing. It's reading the the audit logs. So if I go to the details of this task, for example, I retrieve the output. It's reading the task definition. It's reading the audit logs. It readings the the output and everything. And from that, we try to understand what was the first error, what was the the actions from the audit I get this information, what was the actions done by a user, and I collect this information. I put them inside the... Under the same task. So if you have an access to that task, you will also see that the the the notes of the task. So we are trying to not expose anything that the user cannot see. Also, in this task, I put, like, a... This task is using, like, a a Anthropic, like, Haiku, very cheap model there to do the the the summarize, the actions because it's something really simple. But for example, you don't have Anthropic or you don't... You want to use something else. So in this case, I add another option here. This option is just collecting the data, give you a JSON file. And then you can give this JSON file to any LLM and have the the summarize, and then you can update the task, the definition for that one. But this task is is simple as is doing both for you. But the idea here is Robi can learn. So instead of... Like, you can write your own documentation, but this system, when you enable this webhooks, it will automatically generate a documentation for you under the task definition. And the Robby will learn from this this documentation or these experiences. And the last thing is, like, as Niels also showed, I also did... I I really like this one. So you can connect to MCP servers from the cloud desktop, and you can do the same thing. Like, you can say, I have the task here and launch it. And it's the same thing. Like, when you launch a task, for example, here, so it will fail. Check the the status of the instance. So it says that it it it has failed, and then it it gives you some information. So sometimes it automatically run it. Otherwise, you can say that, can you analyze? If you say if you say like this... So in this case, it will run the analyze now feature. And then it give you the the output of the analyze now. And in this output, also, it will read the notes. And then based on the note, it will give you the the correct suggestion here. I don't know what what I'm doing here. Okay. So it's it's doing its its own analysis. So I I wanted to use the Robi. But, again, like, in this in this case, like, you can also have this information. You can say rerun it from here or or that kind of thing. Think this is cool. This this... Because this can do more than what you can do with Robby, and and you can use any any LLM to communicate with our MCP server. That's all I have. So I will pass back to Lan to see the the results of the the extension builder. So welcome back to part two of the integration builder. We will now check the active activity instances, and sadly, as every demo goes, something has to go wrong, of course. One task fades, sadly, the enhanced one for the actions. We can take a look at this one first. In the output, it is... It's described to what exactly happened. So it tried to get the extension from the controller. It downloaded correctly. It tried to run an enhanced, but then failed because it didn't do the enhanced correctly. So instead of pasting... Uploading a wrong or malformed integration, it will just stop the the the pipeline at that point, and then the the launch has failed. So here we have also a successful example of a failure in this case. We see the download... The the extension successfully did everything just that the... That it did some wrong changes, and that it couldn't fix it within the five time limit that we have put into the integration. So we need to take a look at the integration and see what went wrong behind the scenes. But more importantly, the the Postgres SQL extension went into success, which is which is very good. If we check the output of that one, we see the different steps it took. It auto completed the URL. It creates the the rate before it, and it ran through all different stages of our agent skills that we have. It also gave the different in and out amount of tokens that we used and also created checkpoints in between that if at some point the integration should stop or you you cancel it or you fast finish it, if it has the checkpoint safe on GitHub or inside of the local agent, it will then... Can pick up from there and then continue with that. We see here that the the dependencies have been correctly placed and created and then an upload into the system. We see a skip in the testing because we we will run our tests ourselves, And then it will create the documentation for it and also saved it inside of the GitHub that I mentioned. Below, we see the token usage also split by all the different stages, which model it used, which provider, and the input and output token. We see here it used around 2,700,000 tokens. And I think below at the integration, we see it took a pipeline duration of eleven minutes to create this extension. So without you having to touch any kind of code or learn the the the integration framework, you can directly generate something from your from your scripts that you have already. I will once again head over to GitHub where we... Where I'm in the road... Repo. We see it's private because I set it to private, and we see everything that was created within the integration, we see the dot cloud folder, and we see the integration. If we click on it, we we have the full integration here listed with all its components and also a docs folder, which then create... Gives us a a user guide that was made with the user guide skills that is similar or closest to the one... Once the... For the integration of documentation. So it will create a very close example to what we have on the integration app. For any of the existing tasks, the the LLM will then try to do something similar. We see here we have the overview, the prerequisites. Even the task configuration is being shown very nicely here at an example walk through, for example. For example, scenario one, verify connection, that is just a simple one that we added, and also scenario two, which we will do live in a second where we will carry a CSV file from database analysis. So... Yeah. So let's head over to the integration, and I need to update it first here. There is some Postgres tasks, and we can see query from database. Let's just create the thing with an agent that we choose and choose to create a CSV. It will put the information of some... Of the database. We create the app DB, and we can see here we can use the credentials. I have some post credentials on the controller already, so we will not use the SSL mode. And for the SQL query, I also created something post... That's just called SQL. Yes. SQL query. So this is just a simple inline query that we had from the beginning. Then if we... Of the... I... Of of of course, because we need to create a CSV, also need to be dot CSV. So we will create a file for this and then launch the task, which should be running into successful. There you go. And the task went successful. We see output files has been created at the the the direct... The directory and the the row count and column count, of course. And I could now go back to my workflow and replace the the... In in a sense, deprecated Linux task because it does the same thing with the integration that I can easily filter through the integrations page. I don't need to search through a mountain of Linux task. If I have a lot of different specialized learning tasks, then I don't need to search every single one of them. I can just go through the integration now, and it's it's it's nicely named after what what it should. So, yeah, sadly, I I cannot show you the the enhanced for the GitHub actions, but this is the ability to create any extensions from this integration builder. And I... Once again, for those who are not there at the beginning, I want to make... Remind you that you can also try this out, the integration build up, by being an early test of it. Of course, we will fix any bugs that come along the way, and, of course, the bugs that happened today, those will be fixed before we will send out the first version. But with that, I will also give it back to Quinn. Okay. So I think we've seen quite a quite a bit here today from the AI integration builder all the way through how you can very effortlessly connect to a broad range of solutions and be able to interact with AI. If you look at things like, for instance, the the contest that we looked at where, you know, being able to use AI to connect to multiple servers, UAC, to get the right response and right information necessary would have required an awful lot of heavy lifting just a few years ago. And today, it's just basically a couple of well defined prompts. So, you know, it's it's quite remarkable how AI is is making integration simply easier, less expensive, and better than than ever before. So we have some time for some questions. So let's... Lauren, why don't you feed us the first question? Yes. Absolutely. Alright. Can we use our own locally hosted models? Yes. But but maybe the question should be on which product. So for instance, for Robie AI, right now, we're we're focused on using, you know, Azure, OpenAI and a specific version of that. I think it might be five dot one or five dot five. I forget which one. For some of the other integrations, let's let's turn it over to to Nils for... To to answer that. Yeah. So for the integration builder, we have just added that we can also use local models like OLAMA because we learned that not everybody's allowed to use, for example, clothes. Don't know you can use also codecs. And and also for the com task, because it goes through the ConAI gateway, you can use, by the... Any model which has been assigned by you by the Kong administrator. Excellent. K. Alright. Next question. The change Claude did was on the instance. Can it also change the real definitions? So who did the definition change? Was it you, Nils? Yeah. I did see in... Instance. I changed the instance. So Yeah. And the definition change will come. This is not in this first version, but it... It's something which is on the... Which is, for sure, come the next next release. Okay. K. Can you explain what are the benefits of going through an AI gateway? Should I answer this one? Okay. Yeah. Yeah. So there are actually several benefits, which we basically learned also from our customers. So they said the the company wants control who is using which LLM models and also control which... How many tokens are used and also which MCP servers you can connect to. So all this you can control via the... In this case, in the demos with the conga guy... Gateway, but any other gateways support similar functionality. However, we have also some functionalities already built into our controller. So for example, you can monitor the token usage, for example, through an observability dashboard or also all the AI tasks that you saw today... Saw today show, for example, the output of the token. Great. Next question. You showed using Claude to interact with UAC MCP server. What are the limitations? I think the limitations are that the tools which are provided by the current UAC MCP server. However, it supports the functionality that the same time change that Ruby can also do. Maybe not as inter... Interacting as as connected to the controller, but generally, tool that is called by the Robe AI, and you can also call it from the external MCP server through, for example,. But there's no real... No real limitation, I would say. Maybe, Usain, you can add. Yeah. It's it's... I think it's really better to use MCP server because the... We also generated guidelines or guidelines with the with the tools. So if you use MCP server, so it will do the validation and everything for you. And it's really not limited to Clot. No? So what I've not shown today, you can also use it, for example, from SAP Joule. No? So there you can also use our MCP server, and it works nicely. And, Hussein, during your demo, we did get a couple comments that they would like for you to share the integration that you wrote. Yeah. We can. Perfect. Alright. That's all the questions that we have for now. Thank you, Gwen, Hussein, Nils, and Len for showing the StoneBranch's latest AI innovations that are making enterprise automation smarter, faster, and more efficient. Thank you everyone for joining us today and throughout StoneBranch UserVerse online. As we wrap up this four part series, we hope you've gained insights into the future of enterprise automation and how StoneBranch can bring AI driven and traditional automation together. If you miss any of the sessions or would like to rewatch them, we will share a recording for all the session so you can watch them on demand. Thank you for being a part of StoneBranch UserVerse online, and we look forward to seeing you again. Thank you. Thank you. Bye bye.