The ASUG Tech Connect AI review - Walter Sun reveals why SAP announced its RPT-1 foundation model, and how it differs from an LLM

Now that SAP TechEd Berlin and ASUG Tech Connect '25 are history, let's ask the tough question: did SAP differentiate on SAP AI? As I've written, differentiating on AI is hard:

Most enterprise vendors - SAP included - are pursuing similar strategies for better AI context and output (even though they might claim AI uniqueness). 

This is not a bad thing: responsible/accurate/customer-specific data inputs make LLMs much more compelling, and doing that privately and securely has proven approaches now.

Granted, AI differentiation is also about results. You can have similar approaches to LLM context, and have a different result for customers. But that requires deep use case comparisons between vendors - including pricing, licensing, and data access factors. The maturity/adoption curve is not ready for that yet. 

Differentiating on AI - how can SAP pull it off? 

But I do see two areas of SAP AI that are potentially differentiating. One is end-to-end worklows: I'll get to that at the end. The other is SAP's foundation model. As I wrote in ASUG Tech Connect meets SAP TechEd - SAP makes its agentic AI case, but are customers ready?

But SAP's foundation model for tabular/structured data is a different pursuit - one that most vendors have not taken on. If this impacts the use of tabular data in SAP's agentic scenarios, this could shift from a different AI offering to a true AI differentiator, if you get my drift.

Access to innovation - regardless of ERP release - also matters. For SAP customers, this has not always been clear:

Does SAP provide access to innovation to customers now, or must they move to RISE and go through cloud migrations first? 

Part of this is about customer messaging, not technical limitations or RISE agreements. After all, you can certainly access AI services via BTP, even while on ECC. And: Joule is typically provided as a service via BTP.  Yes, SAP has heard this access-to-innovation criticism before. So has the messaging shifted? 

In Louisville, ASUG did a partial simulcast of the SAP TechEd Berlin keynote with Muhammad Alam, member of the Executive Board of SAP SE and leader of SAP Product & Engineering. Alam hit directly on this issue, prompting me to write: 

Yes, the "flywheel" catchprase isn't going anywhere. But I am noticing a change: a bit less on accessing innovation only through RISE, and more on BTP as the AI engine, and maybe even "teammate" - a tech platform which extends to ECC users as well.

Evidence of this 'innovate wherever you stand' mentality was present at ASUG Tech Connect as well. I caught several strong presentations from customers not running on RISE (yet), but pushing ahead with BTP/AI and Business Data Cloud (more on this shortly). But are BTP customers aware of the possibilities with Gen AI, which ships with BTP? 

SAP's Walter Sun boils down the SAP TechEd AI news

There's no one better to talk to about this than Dr. Walter Sun, SVP and Global Head of AI at SAP, who was also on the ground at ASUG Tech Connect this week, appearing in keynotes and customer AI sessions. 

I asked Sun: what are the key AI news items that SAP professionals should be watching? Sun says that the TechEd AI news boils down to "empowering developers to drive the new business AI revolution." SAP released a slew of news items on these points. During our talk, Sun noted the SAP BDC-Snowflake partnership. He also pointed to the following: 

• ABAP AI in the Generative AI Hub: "We're releasing the ABAP AI model in Gen AI hub this quarter, and that's actually useful, because developers previously could access it through Eclipse IDE, but now they can actually access it directly through the Generative AI Hub."
• SAP AI global skills initiative: "We have pledged to equip 12 million people worldwide with AI-ready skills by end of 2030."
• Joule Studio agent building: "We have agent building capabilities inside of Joule Studio, so developers can now create and deploy their own custom agents on top of the out of the box agents that we have already."
• RPT-1, SAP's foundation model release - along with the RPT-1 playground.

SAP announces its RPT-1 foundation model - but why? The short version

SAP has now formally launched its foundation model, SAP RPT-1, or SAP Relational Pretrained Transformer, which is billed as a new foundation model for structured business data (On LinkedIn, Dr. Philipp Herzig, Chief Technology Officer at SAP SE, explained why SAP is geeked up about this foundation model news.). For starters, here's a few RPT-1 bullet points: 

Transformer-based - It's not an LLM, but it's about on the same kind of transformer-based architecture that powers modern LLMs.

Tabular data - RPT-1 excels at predicting on tabular data, which is not a strength of LLMs.

Aggregated-anonymized enterprise/ERP data - RPT-1 draws on aggregated ERP tabular data.

Predictive use cases - RPT-1 is strong on predictive use cases across enterprise domains.

Architectural benefits - In theory, a broad foundational, tab-friendly model will save customers from building/training a bunch of predictive models from scratch.

Available as a tool call? Yes, RPT-1 could eventually be available as a "tool" for Joule (and potentially other) LLM agents to query on.

The "playground" - RPT-1 (pronounced 'rapid one' as far as i can tell) includes an SAP RPT-1 "playground"  which allows users to log in and check it out for themselves. As per the playground site: "SAP-RPT-1 model can instantly predict outcomes from any tabular data - just upload and get results." (Sun encourages users to test one of SAP's playground scenarios first, before trying your own). 

Inside the RPT-1 foundation model - with Walter Sun

I asked Sun: why RPT-1, and why now? Sun explained: 

We built this relational pre-trained transformer to kind of complement Large Language Models. LLMs can do a great job predicting natural language, whereas this model predicts transactional, numerical transactions. 

For example? 

If you're a supply chain person, you have a bunch of ledgers that say, 'Okay, I have a bunch of shipments coming in from the South China Sea, some coming from Europe. How do you make predictions? Predicting the future is valuable in many, many regards. But one thing in retail, for instance, is the holiday season coming upon us here in North America. How do we accurately predict, how much inventory to hold, so that I don't have excess that can perish?   

But we've had machine-learning-based predictive models for pretty long time now... So what is the advance here? One big key? Reducing a customer's AI heavy lifting. In years past, Sun told me, SAP spent significant time helping customers train custom models. Sun believes RPT-1 will change that: 

The advancement is that it horizontalizes the system... It basically solves a lot of narrow AI problems with one model. No one's uploading sentiment analyzers anymore. No one's building out summarizers anymore, because you just use one LLM, and so what we hope to do - and you'll start seeing this in our tabular AI features at SAP - is personalized recommendations, data attribute recommendations, sales order auto-complete.

We have a bunch of these narrow AI features that our team has built in AI services over the years, and we're going to slowly swap them out with with this model. The goal isn't meant for people to know that it's happening. The best goal is you don't even realize it. 

Sun contrasts this with out-of-the-box LLMs: both can use a customer's unique data as context, but RPT-1 will have an enterprise-oriented result. His analogy: 

Let's say I want to finish a sentence about an umbrella: ' It's raining. I'm going to go to the closet and get 'blank'' - and then the LLM will tell you. [With RPT-1],  I'm actually going to figure out how many I need, how to stock, how to staff my team next week at a grocery store, based on my transactions the past four or five weeks. Given this information, what you know in the system, what's the predicted incoming traffic?

Sun says this is SAP applying its enterprise data strength, into modern AI models: 

We don't want to re-invent the wheel. Others have spent billions of dollars on training those LLM models, and we obviously want to leverage those using our agents, and others things we leverage. But this is something we feel like is our skill set, our knowledge from 50-plus years of experience in business applications, plus the availability of anonymized and aggregated data sets that we can use to inform and help our customers in a tabular sense, versus on top of existing language models.

Which brings us to SAP's second potential AI differentiator: the end-to-end process advantage. In his keynotes, Alam has been adamant about why "best-of-suite" via SAP is better than "best-of-breed." Personally, I prefer agnostic environments, full of customer choice. But I've also seen how difficult it is to get agents from different vendors to talk to each other, even using emerging standards. I'm hearing from ERP partners about the advantages of building agents on unified data platforms, across business processes. If that's the case, then only a few vendors, at most, can speak to the end-to-end process depth of an SAP. 

For Sun, this is reinforced by RPT-1: 

Because we have products across all different lines of businesses, we are able to actually understand how things are connected, and how the process flows. Imagine if we were a company that built only CRM or something, right? Some companies out there just do that... If they want to build a tabular model, they would have data for CRM. 

But as soon as you go into HCM, or you go into something else, they can't do it. So the idea is that, because we have products across all business lines, we can actually create a very horizontal model that extends left to right. Obviously, this is just the first step... but hopefully, we can solve a bunch of narrow problems across many different lines of businesses.

My take - has SAP turned the corner on access to AI innovation? 

Holger Mueller of Constellation Research, my occasional video sparring partner par excellence, was on the ground in SAP TechEd Berlin. He asked me for the top three ASUG Tech Connect takeaways: 

That fits one of my big questions for SAP: why not make more customers aware of the Gen AI Hub via BTP - and that they can start building on SAP AI now, even on ECC? Sun is definitely on board with this - we both noticed BTP customers that haven't yet experimented with the Gen AI Hub. Why not? They should - if only for sandboxing, and identifying viable use cases. SAP should hammer this, and team with user groups like ASUG on more Gen AI Hub education sessions.

While SAP (and its partners!) have made progress on automating migrations, it remains an area to bear down on. RISE migrations and cloud management still require hard work - managing the granular roles of hyperscaler, SAP, and customer. I don't get why that can't be made easier sooner than later, with all the AI resources at SAP's disposal. 

Though I'm encouraged by what we're hearing on Cloud ALM adoption and transformation observability, I want to hear more success stories on modeling from SAP Signavio right into Cloud ALM (I also think SAP needs a lighter/free version of Signavio for jumpstarting use cases, even in pre-sales, but that's either brilliant or dreamy; you decide).

I still believe the public cloud option needs a push, both from customers, to openly evaluate public cloud ERP, and, perhaps via user groups like ASUG, to challenge SAP where the public cloud edition isn't where customers need it to be, for their industry. During a live ASUG Talks "ask us anything" podcast, ASUG's Geoff Scott, Josh Greenbaum and I got into that, via a sharp audience question. I'll table that until the podcast is out.

Criticisms noted, I did see a shift in the so-called "access to innovation" message - and what customers are doing about it. On the last day of the show, I caught three terrific customer use cases - two from customers on older releases, and one that made a move to S/4HANA cloud on their own terms. All three still considered SAP one of their key transformation partners. And that, after all, is SAP's best customer success goal. Even customers on RISE, running the latest edition of S/4HANA, are not necessarily perceiving SAP as their core AI and transformation partner. That should be SAP's ultimate goal. 

In my view, that starts by enabling customers to innovate on SAP now, regardless of release. So when I see a presentation like "Transforming Supply Chain & Order Management with SAP BTP and AI" with Rust-Oleum, I see progress. During that session, Rust-Oleum's Rutul Patel and Amol Dubal talked about building BTP and AI apps with SAP, including two strong AI use cases in production, while running on ECC. Did the ECC data structure make the projects a bit more challenging? Yes, but they still got a solid result. 

Think that helps build SAP/upgrade momentum internally - much moreso than feeling a pressure to move to RISE before they are ready? You bet. Another standout: "A 'Spirited' Modernization: Brown-Forman’s Five-Year Journey to a Unified Data Platform." In this session, Brown-Forman's Danny Miller and George McCracken shared their data platform story, and planned move off of long-time data warehousing with SAP Business Warehouse, with SAP Business Data Cloud on deck. Although self-service analytics was a big motivator for this project, after their session, Miller and McCracken confirmed this is also a big "AI readiness" move for the organization. Ergo, innovation on SAP can take many forms. 

Some will push back on the notion of end-to-end business processes being an agentic AI advantage. Take Sun's example of a CX vendor. Most of the effective/accurate enterprise agents right now are highly specialized - or, a more general domain agent can 'call' the specialized agent, but it's the specialized agents earning the trust and doing the heavy lifting, e.g. flagging problems in contracts or screening potential suppliers. So, a CX vendor could definitely have specialized agents performing granular sales and service tasks, and deliver value for customers without access to ERP data. SAP might counter, however, that visibility into inventory/supply chain data is a big part of what CX agents might need.

The other problem: many customers have so many CX applications that even inside CX, building agents means a chaotic collection of apps in play. This is what Alam is railing against when he talks about why best-of-breed is in trouble. Yes, when it comes to AI search and decision support, there is a solution to multiple data platforms: standardize the data in an environment like BDC, and then build agentic workflows on that standardized data, properly annotated for AI agents, with all the requisite metadata. 

True, BDC is not the only way you could do that - other vendors could also try doing that with SAP data, and will. (SAP argues that this will provide inferior results, due to the "business context" BDC provides... That will be for customers to decide).

But the same doesn't hold for automating enterprise workflows. If we want to build agents that can invoke and execute enterprise workflows (albeit with human approval steps), an AI-ready data platform isn't enough. That type of data platform supports decision making, not necessarily executing transactions. This is where SAP's potential AI advantage really comes into play, embedding agents into industrial process flows. SAP will need to make good on this soon, because I don't expect those advantages to hold forever. Can SAP's foundation model impact all of this, well beyond making predictive models easier to build? I'll be watching if/when SAP integrates the foundation model with Joule and agent building. 

Agent-to-agent communications across vendors will be a much more difficult problem that marketing teams are assuring us (losing context and compound error rates are flummoxing), but progress is inevitable. At some point, your AI will just need to be better and more intuitive than everyone else's - because data and process access won't be an impediment (e.g. composable process libraries across vendors that agents could access). SAP has a big chance to get entrenched in these emerging workflows first - and earn user trust.

It starts with customer adoption, and we saw mixed signals about that at ASUG Tech Connect. But if SAP's BTP-centric AI messaging really takes hold, next year we should see a lot more customer hands raised, ready to talk about building AI on SAP. We'll track it here...