In my conversations with European SAP customers about their AI journey, they tend to fall into one of two camps: excited or overwhelmed—sometimes both. Then I ask a simple question: what’s running in production today?
Usually, there’s a pause. And the answer, if there is one, is rarely confident.
On one side, SAP has articulated one of the most aggressive enterprise AI roadmaps in the industry — including dozens of specialised Joule Agents, thousands of composable Joule Skills, embedded intelligence across 35+ SAP solutions, as well as a Generative AI Hub that gives developers access to frontier models from Anthropic, OpenAI, and Google directly inside BTP.
On the other side, the data tells a sobering story. DSAG’s 2026 Investment Survey puts a number on it: In Germany, roughly 3% of SAP customers have SAP Business AI in production. Definitions of ‘production’ vary across organizations, but the directional gap between announced capability and embedded ERP adoption is consistent. Around three‑quarters of AI‑active enterprises are running their AI somewhere outside the ERP—Microsoft 365 Copilot, team licences for ChatGPT, point solutions, and similar tools.
That gap is the conversation we are having with almost every Head of ERP across our European customer base. Not “should we do AI?” — that argument is over. The real question is: why is so little of it landing in production, and what should we do in the next two quarters that we can credibly point to in our annual report?
The capability is increasingly there. Adoption, in most cases, still isn’t.
After supporting customers across manufacturing, CPG, utilities, and life sciences in Europe, this is the playbook I share with practitioners.
Your starting point depends on where you run SAP today
Before getting into the playbook, one thing that needs to be established: your deployment reality determines your options, and I see too many customers benchmarking themselves against a path that simply isn’t available to them.
If you’re on S/4HANA on-premise, the embedded Joule experience you have been reading and watching about is not available to you yet. What you have is SAP’s embedded AI/ML scenarios via ISLM (Intelligent Scenario Lifecycle Management ). This can be relevant for customers with strict data residency requirements. In practice, ISLM tends to work best for organizations with established data science capabilities and a clear operating model. For bringing in GenAI capabilities, you can use custom AI solution on BTP. This BTP as a sidecar then handles the generative and unstructured workloads: email triage, contract extraction, exception summarization.
If you’re on ECC, the sidecar path via BTP is your only realistic option right now. You don’t need to wait for S/4 HANA migration to start — cash application, invoice handling, and AP exception management are all achievable today from BTP without touching your core. You do need to be cautious about building heavily customized logic on that sidecar before a S/4 HANA migration timeline is established.
If you’re on S/4HANA Cloud (public & private), the embedded Joule experience is available and this is the easiest route to deploy AI scenarios in production. Work with the standard agents first and treat BTP extension as a second phase once you’ve exhausted what’s available out of the box.
The reason this matters: a meaningful share of the 3% production gap isn’t hesitation. It’s ECC customers who have convinced themselves they need to wait for S/4 HANA before any of this becomes relevant. They don’t—provided the sidecar is treated as a deliberately scoped, transitional capability, and not something that quietly becomes permanent by default.
A necessary qualification: contracts and licensing
Earlier I framed the 3% production gap as a discipline problem. That’s only partly true. The embedded Joule experience requires a RISE with SAP or GROW with SAP contract. If you’re on classic on-premise S/4HANA or ECC, the most marketed parts of the platform are not available to you, and SAP Note 3632703 makes that explicit. DSAG’s survey also cites licensing and contract complexity as one of the biggest hurdles, with around 70% of respondents calling it out.
The AI conversation cannot be cleanly separated from this reality. Which AI capabilities can you access on your current contract? What would a licensing change cost in full? And is that the right trade‑off against your migration timeline? These are not secondary questions—they shape what’s practical right now.
How we actually do this
- Start with three processes, not a strategy.
The customers I see actually getting things into production didn’t begin with an AI roadmap — they began with a shortlist. Cash application, AP exception handling, dispute resolution, master data validation. Pick the ones where you have a baseline metric today — cycle time, error rate, weekly manual hours – and a process owner who will stand behind the outcome. Three to five processes maximum. Board mandates don’t ship.
2. Before you build anything, exhaust what SAP already provides.
The instinct to customize is strong, especially in organizations that have spent years tailoring their ERP. Resist it. Standard Joule Agents and embedded AI scenarios cover more ground than most customers realize, and SAP is shipping improvements every quarter. Custom agents have a total cost of ownership that doesn’t show up in the initial business case — maintenance, versioning, keeping pace with the underlying model updates. Build only where there is genuinely no standard path.
The standard-first approach is starting to scale beyond SAP itself. Infosys has published 20 Joule Agents on the SAP Business Accelerator Hub, designed to be discovered, governed, and deployed through SAP’s own lifecycle and consumption model rather than as bespoke builds that sit outside it. The number will grow. What matters is the pattern: pre-built, hub-published agents are how customers get to production faster without giving up governance — and that’s deliberately how we’ve built them.
3. Don’t let data quality be the reason you wait.
Several finance‑oriented AI scenarios—such as invoice extraction or bank reconciliation—can be technically deployed with the level of data quality most organizations have today. The straight-through automation rate you’ll actually achieve is mostly a function of scope discipline. Narrow the first release, accept that exceptions will still need humans, and resist the temptation to claim a percentage. Reserve the more complex automation — the stuff that needs clean master data and consistent coding — for after your core data programme has made progress. There is real value available before your S/4 HANA migration is complete. Don’t strand it.
One reference point: a large services firm with multi-country operations, running on S/4HANA, deployed seven AI agents against its internal AR collections function. Reported outcomes are faster cash collection and a 40–50% reduction in subprocess effort. The detail worth pulling out isn’t the agent count — it’s that the scope was narrow at go-live and the firm ran it before scaling the pattern.
4. For unstructured content, start with the Generative AI Hub on BTP.
Supplier emails, contract clauses, customer dispute narratives — the Generative AI Hub on BTP is the right place to start. Its adoption still lags its potential. It’s the path of least resistance — model choice and governance in the same place your SAP data already lives, without standing up parallel infrastructure.
5. If you’re mid‑migration, use AI as an accelerator—not a reward.
Joule for Developers belongs in your delivery plan now, not after go‑live. Code generation and automated testing are helping teams move faster through the brownfield work that normally drags. Framed honestly, AI as a migration accelerator is a better near‑term story than AI as a post‑migration promise.
Finally, measure from day one. Set a baseline before you go live, agree the improvement target with your sponsor, and review it every four weeks. This sounds obvious. Almost nobody does it rigorously. Projects that don’t track outcomes lose executive attention by quarter two, and then they quietly die. The measurement cadence is what keeps the momentum going into the next wave of use cases.
Economics of AI
SAP has communicated a shift toward consumption‑based pricing for AI services starting in mid‑2026. While list pricing and minimum commitments vary by contract and deployment model, in current commercial models, base entitlements included in RISE contracts often cover limited production usage. Teams planning to scale AI beyond isolated scenarios are well advised to model consumption early—particularly peak volumes and end‑of‑period spikes—before committing to broader rollout. Developer access has been communicated as being available without additional charge through September 2026, which is genuinely useful for prototyping. Use that window to instrument consumption, not just outcomes.
Honest summary
Most enterprises will end up running AI in two places: inside SAP for transactional workflows where the data and the system of record already live, and outside it for general knowledge work that happens in email, documents, and chat. The interesting question is governance across both, not which one wins.
The share of AI‑active enterprises running AI outside the SAP stack isn’t necessarily a failure. A lot of it is customers making a reasoned choice. Microsoft Copilot ships through M365 contracts most enterprises already have. For meeting summaries, document drafting, code review, general knowledge work — these tools are simply where the work already happens. There are two more structural reasons to keep some AI deliberately outside SAP: model choice (you’re not bound to whichever frontier models SAP has currently signed up for), and EU AI Act compliance, which is often easier to govern on platforms designed to be model-agnostic from the outset.
The customers extracting real value from SAP Business AI right now aren’t the ones with the boldest strategy slides. They picked three unglamorous processes, used standard agents, instrumented the outcome, and came back next quarter with three more.
Enough has shipped in 2026 for almost any European SAP customer to get two or three production AI use cases live this fiscal year. The constraint isn’t the technology. Contracts, risk appetite, and operational trust are what’s actually holding this back — and those move slower than roadmaps. It’s the discipline to choose boring wins first.
What I’ll be watching at Sapphire this year
A few things on my list, in case yours is similar:
The SAP AI Agent Hub roadmap, particularly governance, observability, and cost attribution across SAP and non‑SAP agents. This is where the multi‑agent story either becomes enterprise‑ready or remains a demo. When we talk about the AI Agent Hub here, it’s worth being explicit that SAP LeanIX is emerging as the system‑of‑record layer for discovering and governing agents across SAP and non‑SAP platforms—rather than as an execution or runtime environment.
Embedded intelligence in industry solutions—retail order reliability, supply chain orchestration, and the Joule + Signavio integration that finally puts process mining and AI in the same conversation, without needing a process analyst as an intermediary.
ABAP AI announcements. The VS Code direction and side‑by‑side architecture will reshape how brownfield S/4 HANA programmes are delivered over the next three years. As SAP’s ABAP AI and Joule for Developers capabilities mature, the functional gap with specialist tools for custom code analysis, refactoring, and Clean Core alignment is narrowing—shifting the question from “can SAP do this?” to “is native enough, good enough?”
Looking forward to the conversations in Madrid.
