AI for On-Premise SAP: What Runs, What Is Gated, What Is Missing

I am at the DSAG-Jahreskongress in Cologne this week (6 to 8 October 2026, Koelnmesse, motto "Claim Your Ground. Lead Your Business."). DSAG, the German-speaking SAP user group, said it plainly in its press release on day one: key innovations must also be available to S/4HANA on-premise customers and must not be tied exclusively to commercial cloud contracts.

I wanted to know what that means in practice. So I compared SAP's own documentation and announcements with the SAP PRESS handbook Künstliche Intelligenz in SAP-Systemen (Sebastian Sav and Michael Schwarzenberger, 2026), page by page where it mattered.

The short version: it is not one gap, it is three buckets. Some AI runs on-premise today, some is available only if you commit to the cloud, and some is missing or undocumented. SAP itself describes the condition for the second bucket in four different ways.

The claim from Cologne

Three lines from the DSAG press release of 6 October frame the topic:

The numbers behind it come from the DSAG Investment Report 2026 (198 respondents, survey run by DSAG itself). 56 percent use S/4HANA on-premise, 17 percent Private Cloud and 5 percent Public Cloud. For 2026, 42 percent plan high or medium investment in on-premise, against 22 percent for Private Cloud and 6 percent for Public Cloud.

On AI, 43 percent of respondents have implemented use cases and 51 percent have none. Of those with use cases, 77 percent run non-SAP solutions in production, while only 3 percent run SAP-based ones. DSAG asks to read these figures in the context of the survey base and the complexity of the use cases, and it names complex licence models and mixed landscapes as reasons for the hesitation.

How SAP delivers AI

The handbook sorts SAP's AI into two kinds. Embedded AI is standard functionality built into SAP applications that you can use without a project of your own. Custom AI is what you build yourself on your own data and processes (pages 60 and 61). Underneath both sits SAP Business Technology Platform (SAP BTP) with SAP AI Core, the generative AI hub and Joule, SAP's assistant.

Two details matter for on-premise. First, the embedded scenarios the book describes live in the cloud editions: it names S/4HANA Cloud Public Edition and Private Edition as the systems that ship predefined AI scenarios embedded in the application (page 48). Second, the base layer of Joule is part of the cloud subscription. SAP's commercial model deck calls it "Included at no additional cost for SAP cloud customers", while premium agents and features are billed in AI Units (pages 68 to 70).

Where SAP drew the line, and how it moved

SAP gives two reasons. Klein at Sapphire, as reported by diginomica: "When we are thinking about ERP on-prem, we thought about eight year long upgrade cycles. In the cloud, we moved to quarterly, then to monthly. Now you can add agents and skills every hour." And the SAP associate writes that AI agents "require standardized systems to operate reliably".

Behind both sits an architectural fact that the handbook states plainly: the S/4HANA stack has no runtime for foundation models, so embedded scenarios can only do classic machine learning, and everything generative runs side-by-side on SAP BTP (pages 389 to 392). SAP's on-premise documentation shows the same logic: its generative AI scenarios are built on SAP AI Core and SAP BTP. That is why on-premise is not simply switched off, and why the interesting question is who sets up the connection.

What runs on-premise today, with no commitment

This is the part I found most useful. SAP's own S/4HANA 2025 documentation for on-premise describes it:

One difference is easy to miss: on-premise, you wire it yourself. SAP's documentation tells you to buy SAP AI Core, create the service instance (plan "Extended") and the service key, and connect ISLM on your own. For Private Edition, SAP automates the same setup through a service request, and the page for that request says: "The features described here are only available in SAP S/4HANA Cloud Private Edition."

What the documentation does not say is which of the SAP-delivered generative AI scenarios an on-premise customer is entitled to switch on. It describes the mechanics, not the entitlement. That is a question for your account executive.

What is gated

Joule is where the commitment starts. SAP's Joule integration guide (28 September 2026) states: "Joule is available for SAP S/4HANA Cloud Private Edition only within Rise with SAP." and "You can enable Joule for SAP S/4HANA Cloud Private Edition only in an SAP managed data center setup." In the release notes for S/4HANA 2025 feature package stack 02, which cover on-premise and Private Edition together, the Joule item is tagged for Private Edition.

The May 2026 offer changes this for customers who commit. SAP Learning says customers committed to the RISE with SAP journey with complex ERP landscapes "can access Joule Assistants and Agents on their on-premises ERP systems, including SAP ECC and SAP S/4HANA on-premises, during their transformation." The same page adds: "Customers must meet eligibility requirements which includes minimum cloud commitments."

Joule for Developers, the ABAP AI capabilities in the ABAP Development Tools, follows the same pattern. An SAP Community post of 12 June 2026 lists SAP BTP ABAP Environment, Public Edition and Private Edition (releases 2021 and higher) and then adds: "For SAP S/4HANA On-Premise 2025, 2023, 2022, and 2021, availability is subject to the special RISE with SAP contract conditions announced at SAP Sapphire." Note that this is a community post by SAP, not a formal product document.

The condition itself reads differently depending on where you look:

DSAG's own reading on 12 May was a RISE contract plus converting at least 50 percent of the volume into cloud subscriptions, and it added: "Unklar ist zum jetzigen Zeitpunkt jedoch welche KI-Szenarien bzw. KI-Agenten freigegeben werden." (It is currently unclear which AI scenarios and agents will be released.) Almost five months later I found no SAP list of which assistants and agents run on ECC or on S/4HANA on-premise, and no confirmed general availability date. The Register wrote that general availability was "planned for May 2026"; I could not find SAP confirming that.

Capability by deployment option

"Yes" means documented and available. "If committed" is the May 2026 offer described above. For its cloud portfolio SAP states "more than 300 embedded AI scenarios across product lines" (TechEd, 5 November 2025) and "over 30 specialized agents and more than 2,500 Joule Skills" (14 April 2026). I found no official count or table by deployment option.

One more catch: the API Policy

Anyone who plans to let an AI system call S/4HANA interfaces (application programming interfaces, APIs) should read SAP's API Policy, version 4.2026a. Clause 2.2.2 says SAP prohibits API use for "interaction or integration with (semi-) autonomous or generative AI systems that plan, select, or execute sequences of API calls", except "through and within the limits of SAP-endorsed architectures, data services, or service-specific pathways". Clause 1.2 mentions private cloud and on-premise deployments, so the text reaches beyond the cloud editions. How it binds your licence is a question for your contract.

DSAG chairman Jens Hungershausen, quoted by The Register on 30 April: "It's not clear which kind of APIs are allowed for use and which ones are not allowed to be used." The DSAG position of 6 October asks for no gray areas on data access via APIs, usage and access rights, governance and costs.

What I would ask SAP for

What I would do in the next 90 days

The handbook ends with a playbook (pages 505 to 509) that fits an on-premise estate well:

How I checked

My take

SAP has a point that agents need standardized systems, and I understand why the cloud is its focus. But a customer who followed the roadmap to S/4HANA on-premise should not need a migration contract to test an assistant. Until the conditions are clear and in writing, I would start side-by-side: your own SAP BTP, released APIs, one process with a human sign-off. That is available now, and SAP's documentation already describes how.

Sources