One source of personal frustration with the way enterprise AI has unfolded is how the focus on large language models has virtually shut out discussions about other types of AI — even though other forms of AI have great potential. For instance, what we are seeing with AI technologies like automated reasoning in cybersecurity is an impossibility with an LLM.

That said, it is also easy to understand how tech companies have prioritized LLMs to prevent losing market relevance. At this point, if vendors didn’t have something that at least looked like ChatGPT, they would be branded as out of touch with the market. And the embrace of LLMs is not a bad thing in itself. Some of the bots and agents being embedded in software platforms are doing genuinely good stuff. But in a very short span of time, generative AI features have become table stakes. So, what is a vendor to do once it achieves looking like everyone else?

I believe we are about to enter a really interesting new phase of AI. Now that we have reached something of a common denominator, software companies have the freedom to really lean into their own value propositions and customer feedback — with tailored AI-based solutions to match.

SAP recently showed us a very good example of this, and I think it may be SAP’s best innovation since the introduction of the Hana database management system in 2010. For the past year, SAP’s enterprise AI conversation has been dominated by Joule, SAP’s natural-language copilot. Joule is now the “face” of the SAP software environment — the conversational interface that helps users navigate apps, summarize e-mails, and so on. But while Joule handles the talk, a new model called SAP RPT-1 is emerging to handle the tables.

RPT-1 Versus Joule

To understand the difference, think of Joule as an administrative assistant that is tuned to SAP applications and features. It understands your words, translates your intent into actions, and makes the software feel “human.”

SAP RPT-1 (Relational Pretrained Transformer), in this analogy, is the data scientist. Unlike LLMs that are trained on internet text, RPT-1 is a tabular foundation model. It is pretrained specifically on the language of business: invoices, sales orders, and HR records.

How RPT-1 Ties into SAP’s Value Proposition

The real power of RPT-1 lies in how it reinforces SAP’s core value proposition by managing the vast estate of structured data and processes stored in SAP over decades. This goes beyond querying the data to get historical trends. Like a good data scientist, it can normalize missing data as well as make predictions and forecasts into the future. These outcomes historically required expensive data science teams to build, train, and maintain custom “narrow” models for every single use case, like churn prediction or lead scoring. RPT-1 changes the math by using the techniques of in-context learning (ICL).

Because RPT-1 already understands how business tables work, you don’t need to train it. You simply provide a few rows of historical data as context, and it can immediately predict the outcome for a new row. SAP says that this eliminates months of development time; ideally, it allows customers to realize value from their data instantly, and without bloating their systems with custom code.

This ends up being a great complement for Joule, as LLMs tend to have a higher inference cost and lower accuracy when it comes to working with numbers and dates. A good example of RPT-1 and Joule working together would be in a situation where RPT-1 observes an increase in sales pipeline that leads to a predicted increase in sales next quarter. An agent using RPT-1 could trigger other Joule-powered agents to begin an automated outreach to suppliers to forecast more components and inform sales operations, for instance by changing tack on discounting or other promotions to maximize margins for high-demand products. In many ways, RPT-1–powered agents could become a key trigger for the types of business events that LLMs are well-suited to deliver.

Role-Based Models Are Gaining Ground

While RPT-1 is a nice innovation for the SAP customer, it also reinforces a bigger trend in the AI space: specialization. We have already seen many providers (including SAP and ServiceNow) create agents specifically for role- or task-based activities. If you think about the many details that go along with specific jobs — say, an HR professional or a marketing researcher — it’s not hard to imagine how specialized agents tailored to each role are more likely to gain traction in the enterprise than general-purpose agents because specialized agents have more contextual grounding. What makes RPT-1 interesting is that it is a specialized model that could also be applied to specialized agents.

I also find it intriguing that there is an open-source variant of the RPT-1 model available now. So, developers could theoretically point it at any tabular dataset, not just datasets sitting within SAP systems.

Long story short, while Joule makes SAP easier for people to use, RPT-1 makes it easier to derive forward-looking insights. By turning every table in S/4HANA or SuccessFactors into a potential crystal ball, SAP is moving from being a system of record to a system of intelligence. The future of the enterprise isn’t just a bot you can talk to — it’s a foundation that already knows what your data is trying to tell you.