Why Do So Few SAP Customers Run Business AI in Production?
A 2026 DSAG survey, as reported by Börse Express, put productive SAP Business AI use at 3 percent. The gap is rarely the model. It is scope, access, and governance.
Reportedly only about 3 percent of surveyed SAP customers use SAP Business AI productively, according to a May 2026 Börse Express article citing the DSAG Investment Survey 2026. The gap is usually not model quality. It is scope, access to the right SAP data and transactions, and governance that nobody has signed off on.
A caveat first. I have not seen the DSAG survey itself, and the Börse Express piece does not describe sample size or method. The same article notes that parts of it were AI-assisted and then edited, and its headline frames the figure around Joule although the survey statement is about SAP Business AI more broadly. Treat the number as a directional signal, not a benchmark.
What else did the survey reportedly find?
Per the same article, 77 percent of AI-active companies rely on non-SAP solutions such as Microsoft Copilot. That fits what I see in the field: companies are experimenting with AI, but much of it lives next to SAP rather than inside a business process that touches SAP.
The article also states that Joule requires RISE with SAP or GROW with SAP contracts and excludes on-premise installations. SAP's own positioning evolves quickly, so verify current entitlements with SAP before planning around any of this. ERP Today's coverage of SAP TechEd 2026 (October 27) describes Joule Work and Joule Studio general availability as rolling out in phases, with some capabilities scheduled for Q4 2026 and Q1 2027. Those two sources do not agree on timing, which is itself a hint of how hard it is for a buyer to know what is actually available today.
Why do SAP AI projects stall before production?
Three patterns come up repeatedly in mid-market SAP shops.
Scope is a platform, not a process
A team licenses or builds a general assistant and then hunts for use cases. Answering questions about SAP data is useful, but it rarely moves a P&L line. The projects that reach production start from one measurable workflow, such as sales order entry from emailed purchase orders, and work backward.
Read access is easy, write access is the hard part
Most pilots stop at reading. Production value usually needs the agent to create or change documents: a sales order, a delivery, a change to a master record. That means a dedicated user, authorization scoped to the transaction, a simulate-before-create step, and an audit trail a controller will accept. Without that, security and audit teams rightly refuse to go live.
The environment does not match the product
Many mid-market companies still run ECC or a heavily customized on-premise system. SAP's ECC 6.0 mainstream maintenance is widely reported to end at the close of 2027, and the Börse Express article says more than 10,000 customers face a migration decision. Teams in that position often cannot wait for a cloud-first AI product to reach their landscape, and they should not.
What should a mid-market SAP team do instead?
Pick one transactional workflow with clear volume and a clear error cost. Define the acceptance criteria before building: what share of documents must post without human touch, what accuracy is required, what happens on an exception. Run the agent against a non-production client first, with every proposed posting logged. Only then widen scope.
Ownership is a legitimate choice here, not an attack on any vendor. SAP's tooling and a self-hosted agent layer can coexist. SAP is an AWS partner, and many customers will run both, using Joule where it fits and their own agents where the process or the landscape demands it.
For context on our side, SayfeAI's aggregate production figures across customers are 120,000+ orders processed, about 95 percent touchless, and 99.2 percent accuracy, with a typical six-week path from pilot to production. These are SayfeAI platform aggregates, not the result of any single named customer, and they apply to the shipping products only: easyOrder, Regulatory Operations, Product Catalog, and Sales Hunter. SayfeAI (sayfe.ai) is a self-hosted agentic AI platform for mid-market SAP. It is a separate company from Sayfe.ai (sayfeai.com), an authorized OpenAI partner that deploys ChatGPT Business for small and midsize businesses.
Frequently asked questions
Is the 3 percent figure reliable?
Treat it cautiously. It comes from a secondary news article citing the DSAG Investment Survey 2026, with no stated sample size or method, and it concerns SAP Business AI generally. It is a useful signal that production adoption lags interest, not a precise measurement.
Does low adoption mean SAP AI does not work?
No. It more likely reflects early-stage rollout, licensing and landscape constraints, and the work needed to govern write access. SAP continues to ship agent tooling, and ERP Today reports further capabilities scheduled through Q1 2027. Low adoption of any new platform in its first years is normal.
What is the fastest way to get one AI use case into production on SAP?
Choose a single high-volume transactional process, define acceptance criteria up front, give the agent its own scoped SAP user, require a simulation step before any posting, and log every action. Prove it in a controlled client, then expand. A narrow scope is what makes governance approvable.
Sources: Börse Express (May 12, 2026) reporting on the DSAG Investment Survey 2026; ERP Today, SAP TechEd 2026 coverage (October 27, 2026).