Skip to main content

Unlocking AI Value: Build AI Agents on SAP Data in Databricks

Summary

  • SAP data unlocked: Discover how Delta Sharing enables building domain-specific AI agents on SAP data without complex ETL, accelerating AI innovation on the Databricks Data and AI platform.
  • Business intelligence: Learn how generative AI agents reason over contextual SAP data to deliver actionable insights for financial analysis, cash flow forecasting and business decisions.
  • Governed integration: Understand how Mosaic AI and Databricks security controls enable enterprises to build and deploy SAP-powered AI while maintaining data governance and compliance.

Unlocking AI Value: Build AI Agents on SAP Data in Databricks

Watch: Unlocking AI Value: Build AI Agents on SAP Data in Databricks
Talk by Qi Su, Lead Solutions Architect, Databricks
Discover how enterprises are turning SAP data into intelligent AI. By tapping into contextual SAP data through Delta Sharing on Databricks - no messy ETL needed - they’re accelerating AI innovation and business insights. Learn how they:

  • Build domain-specific AI that can reason on private SAP data
  • Deliver data intelligence to power insights for business leaders
  • Govern and secure their new unified data estate

Databricks Named a Leader in the 2025 Gartner® Magic Quadrant™ for Data Science and Machine Learning Platforms: https://www.databricks.com/blog/databricks-named-leader-2025-gartner-magic-quadrant-data-science-and-machine-learning
Build and deploy quality AI agent systems: https://www.databricks.com/product/artificial-intelligence
See all the product announcements from Data + AI Summit: https://www.databricks.com/events/dataaisummit-2025-announcements

Chapters

FAQs

How do I build AI agents on SAP data?

This video shows how to use Delta Sharing to connect SAP data directly to Mosaic AI on the Databricks Data and AI platform, enabling you to build reasoning agents that analyze business data without manual ETL.

What is Delta Sharing?

Delta Sharing is a secure, open protocol for sharing data. In this context, it allows SAP data to be shared safely with AI systems on Databricks for building intelligent applications.

What business problems can these AI agents solve?

The video demonstrates cash flow analysis and financial simulations that help organizations forecast business outcomes, optimize working capital and make data-driven decisions.

How is this secured and governed?

The Databricks Data and AI platform provides Unity Catalog governance, authentication and audit trails, ensuring SAP data security while enabling AI capabilities for business users.

Full transcript

0:00 Introduction to SAP and AI

All right. Hi everybody. Uh welcome to my session. So today we're going to be talking about the value of building Genai agents on SAP application data. Uh so to introduce myself. I'm a solution architect with data bricks uh leading customer engagements globally for our SAP team. So I've been having lots of customer conversations over the last three years about ERP, HANA, BW and data sphere integration and also about uh this year a lot of conversations about business data cloud what that partnership means to our customers our joint customers with SAP and data bricks uh depending on where they are in their S4 and rise journey.
Okay. So okay so just quick reminder that there will be a a feedback survey uh so please fill that out if you're able to after this uh session that'll help us improve our sessions. Uh so this is a news that was released uh last week that we announced that we are in the latest Garner 2025 data science and machine learning magic quadrant that we've moved up a little bit further. we were already a leader and so now we are the leader uh being the kind of the most upright and have leap across leaprocked AC um above the big three clouds uh in terms of uh leadership.
So something that was very very proud of and so and that's also of course part of this uh conversation is about around the data science and machine learning topic today. Um so we all know right that data is uh data and AI especially AI and geni these days is a hot topic for all customers across all industries right customers are talking about that and how do you get value from that across the different industries right every single industry has use cases and also all line of business horizontal line of businesses across all these industries have AI and genai use cases and so there's some um garner uh garner analysis about kind of value and growth in AI software and so on.
So of course all of that to have AI you have to have a data foundation to have AI right? So to be able to train your AI, you need data and SAP applications. And this is of course the most well-known application is SAP ERP and that could be legacy ECC as well as a modern ERP, S4 HANA, right? All of those are gold mines for data, right? Because ERP is arguably kind of the heartbeat of the company, right? So when we say, hey, company XYZ closes the books, right? as in closed accounting books.
Well, that's done in accounting system which is in in ERP and SAP is the dominant player in the ERP market and so and but that's not it right ERP is not the only product SAP sells SAP has a very wide portfolio of all these line of business applications that's now terminology business suite so from the finance and accounting and also to things like supply chain logistics and so on but also to other non-ERP line of business applications like success factors for human capital management agement, Aribba for procurement, Concur for travel and expense, Fieldglass for contingent management, the CX portfolio of salescloud, service cloud, customer cloud and so on for CRM and so on, right?
That's a very broad broad range of applications that are all generating very valuable enterprise data and so those are that the data that we need to combine with non SAP application data with IoT data and so on so forth to be be able to derive insights to for analytics and for AI okay but historically SAP application data especially SAP ERP data has been very difficult to access for our customers right and so that's the challenge for the to build AI is to be able to get access to the data to the enterprise data.

3:45 SAP Business Data Cloud

So this is where the partnership that we announced in February with SAP about business data cloud right so what that takes is that there's number of components to business data cloud some of which are existing components like data sphere business warehouse on rise private cloud edition as well as SACE SAP analyst cloud for BI and also for planning those are existing components what's new is that SAP is OEMing data bricks so that's called what we calling SAP data bricks that's the OEM version of data bricks that's tailored for a IML use cases on SAP centric workloads and data sets right so that's also that's provided as part of business data cloud in a seamless commercial manner and also tightly integrated with business data cloud and data products and on the bottom here you have the data product so that's a brand new concept that SAP has introduced in business data cloud where the idea is that historically think about how customers do their analytics whether it is in especially in non SAP data platforms is that they're extracting raw data from ERP that could be ECC tables, could be S4 CDS views and building out business logic to form let's say sales order business objects, purchase order business objects or combining them to form order to cache as a business object in a uh non typically a nonSAP data platform right and that could be 30 years ago that could be terod data five years ago that could have been Hadoop and these days is data bricks and our competitors right so they're building out these data models to be able to have a data warehouse with the semantics of the SAP application data to be able to do the analytics whether it's descriptive analytics all the way to advanced analytics AI and genai right so first to be able to do that you have to have that semantic foundation of SAP application data so the idea of the data products is that well you don't have to do that yourself because that's a lot of labor to do the ELT extraction load and transform SAP will do that for you and that becomes a managed service.
So that means quicker time to insight by outsourcing that you don't have to deal with complexities and labor to build out the ELT, right? So you go straight to the output of that T of the transformation is that these enriched data products that you can directly consume. They're ready for analytics, ready for AI, right? So that's something to my opinion that's the most valuable part of business data cloud is this concept of SAP curated data products for specific line of business applications including S4 on the cloud as well as success factors arba concur CX and so on so forth but also there are some uh SAP will be building crossline of business applications uh um data products as well that span let's say a employee right employee exists in S4 and an employee exists in success factors or a business partner right which exists in S4 and also in Aribba to be able to build these richer cross line of business data products as well.
So that is something that's made available to customers of business data cloud who are using native capabilities such as data sphere sac but also to us at data bicks as a partner right so data bricks customers can consume these data products as well in the form using SAP data bricks which comes as I said um as integrated within business data cloud commercialized by SAP or right so we have 12,000 plus customers many of you guys are here right there you are existing data bricks customer.
So what you have is native data bricks on the right. You already have existing data bricks and so that will also be able to consume these data products and be able to do a IML analytics and so on. So a couple key things that SAP has adopted that as part of this partnership. One is Delta Lake, right? So Delta Lake is the open cloud storage format that we open sourced, right? That we believe to be strong in. And so SAP has adopted that as the persistency format in their managed cloud storage.
And second is a delta sharing protocol, right? Which is the only open data sharing protocol out there that we open source. So SAP along with many other of our partners out there has adopted that as well as the data sharing mechanism as open data sharing mechanism. And so that allows our customers to be able to access SAP ERP data SAP data products in BDC without the traditional way which is physically copying, replicating and etling SAP application data.
So these data products you see here in that blue box on the right within BDC is that these data products are managed by SAP's data engineering pipelines and persisted in SAP managed cloud storage within BDC right and then and then data bricks whether you're a customer using SAP data bricks or native data bricks you can consume that those data products including all those rich semantics on the fly right without physically copying that data you can directly do analytics directly do MLflow or mosaic AI and genai on top of that right without physically cop making copy of that data.
So that reduces data proliferation and also right you still get the benefit of the SAP having done the data engineering to add the semantics to it. Uh and also you you may notice that the delta sharing arrow is implying birectionality right so that's also very important is that it's not just okay well data brick customers want to get data out that we can also bring insight back into the SAP analytic ecosystem. So um for those of you who watch the uh unleash demo uh when the unleash event when BDC was introduced on February 13th, there was a demo there of the use case of um of account receivable uh invoice uh forecasting of taking invoice data which is a data product that SAP has curated and accessing it.
In the case of the demo it was in SAP data bricks but you can do that with native data bricks as well to do access that on the fly and doing machine learning to do just traditional predictive analytics to forecast the individual invoices. what's the payment data going to be? So now you have a ML inference data set that is managed by data bricks. So what we want to do is also to bring make that available to data sphere without copying it into data sphere's haunted memory right making that available on the fly and then data sphere is able to read that and then data sphere can show that in the sec dashboard and so that's the end of the demo and there's a similar demo at sapphire a couple weeks back where the demonstration was about success factors is that you have headcount data and you know employee central kind of data bring that into data bricks and similarly make some predictive analytics forecasting about attrition and then bring that back into data sphere.
But the possibility doesn't end there at this kind of analytical feedback loop, right? There's also quite a bit of potential for actually actioning back into ERP and into SAP applications, right? you could think about okay once I bring that insight about whether it's cash flow forecasting or invoice date or employee attrition and whatnot into data sphere then next step I could do is I could use services like BTP integration service to take action in the SAP transactional applications maybe I should go create a sales order or if it's a predicted maintenance maybe I want to go and create a work order to have a technician go inspect a piece of equipment that the data bricks ML has forecasted to have have problems potentially will fail right so this is very interesting Interesting that there's actually full life cycle of not just analytical but also to take action.
That's very powerful. Okay. So let's use a banking example. So this is one of our uh motivated by one of our customers who used um genai in data bricks to be able to and they were able to increase the engagement on the mobile app by over 4x. So traditional um AI kind of chat bots is really about retrieving information, right? Right? This is where you ask a question about hey you know what is the uh what is what is the what kind of credit cards are relevant to me and so on and they were able to retrieve information and techniques they use are things like retrieval augmented generation or rag which uses a vector database to fetch the information to kind of limit the information set and then bring that to a AI for summarization and so on or fine-tuning to incorporate domain right two different ways of incorporating domain knowledge fine-tuning versus rag Uh so yes so this example here is that hey um I want to know about is my credit score high enough to open a new credit card and then so here's some information that's relevant to here's where you go to find the right cards and apply versus something that's more a proper agent is much more proactive and is able to provide recommendations and assist you in proactive way.

10:56 GenAI agents for business

So it might be that okay so that same question same prompt different kind of response is that okay well you know uh in addition to some recommendation about the credit card and also other things like looking at your trips and providing recommendation on the trips as well uh so that's kind of the more proactive example okay and so thinking about building these kind of chat bots or these kind of genai or power chat bots there a number of things that you need to do is that first is that the data silos right so okay you know we might have a database about all the credit card offerings and so on, right?
But more than that, what else could we possibly use that will make this interaction much more engaging, right? Things like we might want to look at this customer's mobile app history, uh going to the backend banking system to see what transactions they have made, right? So, there's all kinds of additional information or look at the multi- channelannel, what they're doing on the mobile app, on the website versus in the customer service that they're calling in to the customer service center. you're looking at the interactive voice response IVR system like that multi-channel omni channel information customer 360 is very important obviously right to be able to break that silo so that requires having a data platform that allows you to access all those data and also the you know lookup functions and so on to be able to tap into these different sources to look up information and then the second is that using the foundational models is that do we have access to the right tooling to have you know the state-of-the-art foundational models the AI the best AI right so tomorrow Right?
If the next deep sea comes out, do we have the ability to leverage that or are we kind of stuck with whatever foundational models a particular platform is offering us? And then the pre-built connectors and functions to use basically a toolkit that helps you pre-built things to help you uh build out integrations and so on faster so that you can uh get the data that you need faster and then be able to build the genai uh in a more effective manner.
Okay, so this is the philosophy behind our data intelligence platform is all about democratizing data and AI, right? So this is where you think about general intelligence is really about you know what Facebook and others are doing in the community which is super important is building a general intelligence model using publicly available data sets, right? But that what you need is not I would argue is not sufficient is that you as enterprise you need access to your own customer data but also your proprietary knowledge right so if you're a u equipment company to think about your internal manuals or any company think about your internal wikis and information about processes you think about if you're HR professional thinking about your human resource HR policies right all of these are proprietary data sets proprietary knowledge you want to incorporate them depending on the use case into your AI Right?

14:05 Data intelligence platform

So to be able to do that's the data intelligence is that it's leveraging enterprise data right not just what's whatever data is available to llama to train on the internet and so on and so this is where um we are pretty early on uh we were early leader in that early this is in February of 2024 uh that our co-founder mate who leads a research lab in addition to his job being a co-founder uh working at data bicks he also leads a research lab at came up with this paper about multimodel compound agent systems is that really thought that to have agents to be very effective that you really need a multi- multiple function multiple um steps and also multiple components multiple agents collaborating together that gives you the richness and to be able to do more accurate AI and also to be able to take action and be more effective right so first is that so these building blocks on the left you see that is to be able to reason over your data so this is more about breaking silos so that talked about in the previous slide you right so kind of generalize to say the SAP and the nonP silos right historically that's been there because SAP ERP data especially is very difficult to tap into to be able to tap into that as well as your nonp IoT Salesforce whatever data all of that data to be able to tap into that's that because that's your enterprise data specific to you as a customer right and then to be able to understand and evaluate your models right so the key question to ask is for whatever framework that you're using you know data bricks non data bricks whatever can you answer this question which is what is the accuracy performance of my model right are you able to answer that and then the next question if you are able to is that okay do I know how to improve my model right do I have a way to get to that right so these are we'll dive into a little bit deeper a little bit later as well on that uh and then governance right so what we don't want is to uh give so I just said in breaking the silo is I want all the data right but you also have to think about that very carefully is that depending on who's using the genai so I want geni to have access to all the data but when the AI is being invoked I don't actually necessarily want all that data to be available so think about in the SAP context is that I am a finance or accounting analyst right and I work in the I work in the United States for a multinational company and I want to do some genai about u you know general ledger or AP accounts payable accounts receivable whatever ever that may be.
Okay. Well, you do the enterprise should not give me all of finance, all of general ledger, right? Because I'm not privy to see the general ledger finance data for Europe because I don't that's not my function, right? My function is US finance or accounting analyst, right? So in that case is that the gen should have all the access data but the gen should also respect the role of the user and not give them more data than they should be privy to.
Right? Right. So this is very important is that you don't want to be building a data access framework. Right? That's that's that's work is very important but hopefully you don't have to build that yourself in every single geni system that it should come out of the box to respecting the access control policy of the user who's invoking the AI right so this is where our philosoph philosophical design of AI and data is all tied into unity catalog and unity catalog is a mechanism that provides this uh very granular access control right so that when you have AI and it's invoked by me the US finance analyst automatically the AI is going to respect that and I'm only going to get AI responses based on the data that I can see which is in this case US finance data and not Europe finance data right so that's super important especially in this kind of enterprise application data space whether it's you're talking about finance data or you're talking about HR data and so on okay so uh in our mosaic portfolio this is just small sample of some of the key capabilities and there why there's this kind of bar in the middle is that on the left is more of a classical ML on the right is more about this agentic and things like that and genai capabilities so what we offer historically have been things like MLflow which is the industryleading open source framework for full ML life cycle right everything from data preparation to model training experimentation to model deployment productionalizing a model deploying that in an endpoint point of whether it's a real-time inference or batch inference and then to model refinement monitoring to monitor for data drift as well as model drift and then to geni right which is there we have a different slide here uh that's also there's a lot more capabilities there this is just a small sample of that and we'll have some significant announcements coming over the next two days so look for that in the keynotes so first of course just as with classical ML we need to prepare the data right and then some ji specific capab capabilities.

19:37 Mosaic AI capabilities

For example, the vector database, right? To be able to do things like ragg, you need the vector database as a component and then the building agents to be able have access to state-of-the-art commercial as well as open-source foundational models and also classical ML models in some cases as well and having the framework to be able to um provide get additional data to provide lookup functions to go look at let's say go to a CRM database to look up customer records, right? the example we're talking about with that banking chatbot to be able to go to the interactive voice IVR system to look up you know what what what's the call history what this person did with customer service when they called in and so on right so these are functions and tools so you need a framework to be able to tap into data from different sources and then once you have developed your agent right kind of deploying the agent right framework for that the ML apps and also understanding the lineage of what the agent is doing and then the evaluation is super important is that so I I posed the question to you is which is that do you know what's the accuracy of your uh of this agent or of this model right so to do that how do you need to do evaluation so the evaluation has multiple parts there's a gold standard data set how do I curate a gold standard data set do I need to go hire a bunch of army of people to label datas so that there's a gold standard label and then I can evaluate against and once I have that label data set how do I uh evaluate it right so do I and do I just have humans go prompt questions and then evaluate the answer do I use Right.
One industry trends is using LMS as human-like judges to be able to do this kind of evaluation. And then tracing is also important is that thinking about what you're doing that right because we're talk we just talked about multi- aenting right so there's a lot of different components right I did this lookup function to go to the IVR system to look up what they're the customer did in the um in the customer service call center right and then I had this different LM steps where I'm calling a llama model and I'm calling a deepseat model or whatever a is the measuring the accuracy right which requires this um gold standard data set of the ground truth and then be able to evaluate but also right so that's the accuracy part of performance and there's also latency part of performance so I want to be able to in this complex agent orchestration that I'm doing I want to understand be able to look into individual steps and understand how the performance has changed whether it's accuracy or latency performance right I will want to know okay you know my perform my whole holistic agent got slower but is it because of particular step oh because the lookup to this IVR database some somehow slowed down right do I have the visibility to be able to dive down to even see that level of detail.
Then I can go look okay why is that because that database is overloaded or whatever or is it because the llama model that I'm using that some whoever is hosting it for me for serving has slowed down for whatever reason right is it because of my quotas are running lower or whatever so um and then on the governance side there are a number of capabilities that we're providing for guardrail so you know very basic things like profanity and so on but also more interesting twists to the guardrail that you might not think about is like corporate branding standards right so did you might the way you guys you as a brand talk about certain things, right?
You might want to impose that, right? Which is very very different from like things like profanity and those kind of things that you don't want in the in the Gen AI response, right? And then of course the uh things like usage tracking to be able to you're you're um centralizing the management of the access to let's say Azure Open AI and to be able to manage tokens set rate limits for specific groups of users and so on and the managing centralizing management of the credentials and so on.
So there are a lot of things because genai is complex especially multi- aent setups are complex to be able to manage this end to end with a lot of these well thoughtout features. Okay. And so this is a one customer example where we're able to leverage the multi- aent and uh these things about using these multiple agents working together we're able to significantly increase the accuracy of the um of the re uh of the offers and also kind of increase the user engagement as well.
Okay. So in this slide right we just want to quickly talk about um enhancing the decision making is that so this domain specific AI right what that means is that as we said incorporating you as a enterprise your SAP and nonSAP data right being able to tap into all of that breaking the data silos which really enhances the AI I would argue that makes AI more accurate right versus just having AI on SAP only part of the house and versus having AI on the nonSAP part only of the house really combined finding both sides and finding all of that data really makes your AI more accurate.
Have all of that AI enterprise data connected to any avail any gen AI model that have available have that available to in your at your fingertips the commercial ones as well as open source ones right without worrying about how do I host the latest version of you know llama 4. 0 or whatever that comes out, right? And doing that in a cost-effective way and also having all the guard rails about safety and reliability and kind of being able to monitor and diagnose the quality and also the latency performance aspect of that.
So this is where really business data cloud uh and the partnership comes through with us by providing seamless access to SAP application data and also the semantics of that very important right. So we don't have to do a lot of extra work of data engineering to curate these SAP application data combining with the power of data bricks that we just talked about over the last few slides about our native capabilities for building custom AI what is classical a IML to genai okay so let's switch to do two demos so the first one is a um notebook book.

25:39 Cash flow analysis demo

So um one of the uh so in in February when business data cloud was launched and the initial set of what's called working capital data products were released and it include it's it's all it's a subset of S4 finance. So it includes things like accounts payable accounts receivable uh working capital general ledger cash balance day sales outstanding and so on. And so one of those things is of course account um is the cash flow data cash flow data right so these are SAP curated data products so we're going to look at the cash flow data product so this is what that tabular data product looks like in data bricks so uh actually let me past it right here so you see the cash flow ID company code transaction date transaction amount currency and so on right so that's just a sample of what uh what that tabular data set looks like from a SAP curated cash flow data product.
Uh let's aggregate it because there's a lot of columns that we don't necessarily a particular uh cash flow analyst may or may not need. So let's aggregate it to the company code, transaction date, currency and the currency amount and the number of transactions. Right? So that gives us to only 311 uh rows. I think this is whole the whole data set was about 10,000 rows. So it's aggregated 311 rows. Then let's take that tabular structure data set and just turn it into a big string right and because we have to do LMS right we have to prompt it with text so that's the prompt uh well that's the text part of the prompt so then there's the context so we need to provide some system context and that is hey you are a cash flow analyst and here's the tabular data each row you know what what each kind of explaining helping the LLM this is what the row means and then for a particular company code and transaction date and then hey here's the action part can you give me a summary of the data any interesting insight that may be actionable and then write it in the form of a email that I can send to the cash flow manager for review right and then so specifically here we're calling the anthropic cloud model right and you could be calling your model your favorite model whether it's Azure openai know one of those like 04 model or something um or or any other you know vertex models and so on.
Any commercial or open source model uh that you like uh and then here's the response. So we're not going to look through every single bit of it but very quickly right so think about right this is where we talk about the power of AI is to make people more productive. So this is something that I think about a your cash flow analyst normally would do. Slice slice the data and do some analysis might take you know minutes or hours to do that the AI has already be able to do very quickly about total transaction volumes.
These kind of summary statistics and then these are the high value transactions. Um and then currency exposure because we're we are a multinational company here and then seasonality patterns. And then here's the interesting um recommendations for interesting insights right is that okay so you have some large outflow so maybe you want to look at that is that anomalous or is that you know normal course of business and this offsetting transactions high volume days currency risk maybe you want to hedge current currency because you're doing a lot of work a lot of business in Japanese yen and then specific recommendations about deep diving into transaction patterns for a particular sub company code uh what they're doing maybe they're you can might want to do some optimizations there.
The yen uh currency exposure management, hedging and then verification of these unusually large transactions, right? So again, right, the point about this is very um helps with productivity, right? So that this is very something very quick that AI just did, right? Without having the cash flow analyst spend a bunch of time. Now, you know, we we this is, you know, hopefully accurate, but you know, something might want to somebody they actually might want to actually have a cash flow analyst review this information before sending that email to the cash flow manager.
But this is something that's very quick, right? That was just done within a few seconds. The AI is able to come with these interesting insight that would probably take a analyst quite a bit of time to manually come up with these kind of insights, right? So that's the power very very simple example that to show showcase kind of the power of Genai, right? So then there's um so this is a very straightforward just bringing a tabular data set you're essentially telling the AI hey show me something interesting do some summaries and so on right and then the next one is the kind of a different art of the possible is that we're doing a uh this is a demo recorded video is a 4-m minutee long video uh to show um a chatbot right and this is the multi- aent so this is a mockup so don't it's not like actually a functioning system but it's a mockup to show art of the possible give some background we built this fin bot to corporate finance simulation and what if scenarios using enterprise finance data.

30:22 Financial simulation demo

Specifically, we're analyzing SAP S4 HANA finance data, including accounts payable, accounts receivable, cash flow, cash balance, and more. If you're interested in analyzing these valuable data sets, they are data products that are part of the working capital insights app made available by SAP business data cloud to data bricks via zero copy delta sharing. So let's ask our first question. Given uh economic contraction in the GDP as well as high inflation over the next six months and stimulate impact on our cash flow by business unit and also generates some mitigation strategies.
We see here that uh the AI has finished uh cons uh reasoning uh using the available data and uh we see here the um impacts on the different business units on the cash flow and also the driving factors and we also see the risks from uh the cost side as well as the revenue side and both sides being squeezed these two are working capital pressure and so we see that with the AI came up with a mitigation strategy to with the first bullet and the third one to address on the con side and then with the second one with the accounts receivable collection to address on the revenue side.
And so how the AI went about this, we see on the right in the intelligence panel where we see the first step with a Monte Carlo simulation by first obtaining the economic data and then combining that with the finance data to perform the Monte Carlo simulation and then using AI to generate recommendations based on the uh simulation output to address the identified risk factors and then with this output we can also use some thumbs up thumbs down and provide feedback so that the LM could be further are refined through fine-tuning.
Okay. Next, um let's uh before we take any action, let's dig deeper to identify the root cause of the cash shortfall. So, we're asking the AI to break down the projected liquidity gaps by month and also by business unit. And we see here is that there's the analysis by uh by month by business unit and also by business unit the gap as well as the driving factors similar to what we saw on the previous question. And here the um AI is showing intelligence panel is showing the steps of disagregating the Monte Carlo simulation results and then using AI to summarize the key drive key contributing factors.
So in the third question um let's uh simulate the first strategy offered by the AI in the first question. So what would be the cash flow impact if we uh applied a across the board opex redu uh reduction across all business units. So there we see is that um we see the output here of the u changes based on that what if scenario of reducing opex spend and we see this improvement in uh liquidity position which is good news.
But how did the AI get to that? And what we see here on the right is that um uh in this case, Finnbot is leveraging SACE planning to allocate the cost reduction. And how this is achieved is via SAP data bricks delta sharing its data sets with via zero copy delta sharing to SAP data sphere which in turn is accessed by SEC planning via a live connection which is the seamless planning capability that was introduced uh in the Q4 of last year and then the second step SEC SAP data bricks is analyzing and summarizing the outputs of the SACE planning generated plan that's how we got to arrived at the um output uh and uh revise the liquidity position.

34:39 Q&A and closing

Let's explore how we can use SAP data bricks to build a genai system on SAP enterprise application data as demonstrated via chatbot. To give some background, we built a Okay. And so does yeah. So in closing uh so we have a we already had a session this morning with one of our product management colleagues talking about uh using SAP data and data brick SQL uh also through the business data cloud partnership. Uh we have a session with uh SAP colleagues um this afternoon about uh the business data cloud partnership as well as additional session tomorrow morning uh with SAP colleagues.
Similar topic as this session about fueling AI with SAP data products. Um and then in the afternoon um around lunchtime we have u colleagues from our product management team including the delta sharing uh PM Darina talking about the business data cloud connector for delta sharing to native data bricks um and then there's a session in an afternoon um with uh with the product leaders uh from us as well as uh from SAP and also from Salesforce talking about this paradigm of zero copy delta sharing since data bricks has a similar partnership with Salesforce as well you have this federated data lake concept of SAP of data bricks compute being able to access kind of enriched data from Salesforce from SAP and be able to do analytics and AI and so on.
Okay, so with that we have a couple minutes for uh questions. So, let's open it up to the room. So, a lot of these um implementations like some of the problems always become with like the accuracy and so people get frustrated because the LLM generated results aren't good enough and you know it's like hey why why waste my time. Wondering if you have seen that how you kind of work through that. Um uh and any other com comments there?
Yeah, so I mean that's where the LM evaluation framework comes in is that to be able to have a ground truth data set that is labeled, right? So that's the accuracy and use the whether it's a human judge or LM judges to evaluate based on that. Right? Because to be able to accurately kind of objectively measure accuracy, you really need to have a ground truth of you know this is the correct answer to this problem so that you can detect things like u hallucinations and so on.
Right? So now you know that there's course there's a lot more nuance to that, right? So definitely you know have just definitely reach out to your uh database account team if you're a customer and kind of have that discussion with the account with the solution architect and also we can bring in the right subject matter experts uh to talk about that. Thanks. Yeah. Back here. So there's actually more of nuance to your question, right?
So your question was the con example you use a warehouse management. We can also use finance as well, right? So kind of like the one I mentioned about finance analyst is that so first is that the access control of let's say an S4 or in your case EWM is not actually transparently transferred over to data bricks right so today what you have to do in the early phases of our partnership you would actually have to define that from scratch in data bricks that's something that we're working jointly with SAP BDC engineering to be able to enhance the delta sharing protocol to be able to pass through access control policy so that we can seamlessly uh understand that and then apply that right but then today you do have to apply that right so let's say you do that manually and applying the access control policy but we do also have consulting partners who can help you with automatic replication of ERP and STP application including BW uh access control policy in data bricks well so that's a separate conversation but once you have defined that in data bricks then yes whether it's genie which is our text to SQL tool or other things like mosaic geni agents and so on we'll be able to respect you the data bricks user whatever ever your role is and the security policy has been defined in terms of what plans or what plants or you know what what what company code information you can see other questions.
Yeah. So dual agents today has a um fairly new the Juul has a release of jewel support for data sphere right. All right. So I see that's probably where we can potentially have an integration point. This is something that's still in the works about uh you potentially using data sphere since we have the birectional delta sharing with data sphere. So data sphere has access to data bricks curated data sets right. So then Juul can access that.
It looks just like a data sphere data set to Juul to the dual agent right. So that could be potentially and then this is also in discussions about further advancements and other uh dual deeper dual category SAP data bricks which is considered an SAP product. Uh so you could do that it's a little bit more complicated is that you can do data bricks federation to data sphere which is through the HANA right which is through HANA right using the open SQL schema you can also do the same thing with BW but then BW this is where you need to have BW you need to have HANA enterprise edition license for BW and then expose BW objects as calculation views then we can data bricks can directly federate to that Yeah.
Well, the federation the existing federation is zero copy and then what's new in BDC is that BDC with BW data product generator which requires BW on rise that's materializes it in the uh SAP managed data store right that's also a live connection through delta sharing uh and also in the future data models will be able to be materialized as a custom data product as well and that's also delta sharing right so there's you have the federation which already exists versus the delta sharing based uh dynamic access uh so it looks like time is up I'm going to be in the back of the room for a couple minutes before I run over to the to my next meeting.
So, yeah. So, stay around if you have a few questions.

Learn more about the Databricks Data and AI platform.

The information provided herein is for general informational purposes only and may not reflect the most current product capabilities or configurations.