Concept card

SAP Business AI — data privacy and data usage

As of 2026-09-25

Edited by Cédric-Olivier Mâry, sole editor and publication director — 27 years in the SAP ecosystem. How these pages are produced, and where automation is used.

What is SAP Business AI?

SAP's public commitment is specific, not generic: customer data is not shared with third-party model vendors to train or improve their LLMs, and third-party foundation models process a prompt without retaining it beyond the inference call. What SAP does not promise — full data residency inside the EU for every model, or that its own masking detection catches every personal-data pattern — is exactly what a consultant has to check per engagement, per model, per region.

The specific commitment, stated precisely

SAP's public position on customer data and AI is narrower and more useful than the marketing phrase "we don't train on your data" suggests, because SAP states the actual mechanics: SAP does not share customer data with other vendors for the purpose of training or improving their large language models or image models, and any data stored in SAP AI Launchpad is not used for quality improvement of SAP products, for example to train LLMs. When a customer uses an SAP AI capability that calls a third-party-hosted foundation model — GPT, Claude, Gemini and others reachable through the generative AI hub — the prompt is sent to that model provider for processing only; it is not stored or otherwise retained by the provider beyond the inference call. That is a contractual and architectural claim, not a vague reassurance, and it is the one line worth quoting verbatim to a client's data protection officer, because it answers the question DPOs actually ask: does our data end up inside someone else's training set. The documented answer is no — but the answer covers training and retention by the model vendor specifically, not every privacy question a deployment raises.

Why it matters

SAP's documented commitment is specific — no sharing of customer data with third-party model vendors for training, and no retention by the model provider beyond the inference call — which is quotable verbatim to a client DPO instead of paraphrased from marketing language.
Masking entity detection is scoped: person names reliably in English, addresses and locations for the United States only — a European deployment needs custom entities tested on real data, not the default profile out of the box.
"Not retained by the model provider" is not the same claim as "stays in the EU": the model endpoint's hosting region still drives the Schrems II analysis, independently of SAP's no-training commitment.

Key points

SAP does not share customer data with third-party vendors to train or improve their LLMs or image models; data used by third-party-hosted foundation models is processed, not retained, beyond the inference call.
Data stored in SAP AI Launchpad is not used for SAP's own product quality improvement or LLM training.
The privacy architecture chains BDC (data stays in the managed landscape) → AI Core (assembles the prompt inside the authorization scope) → generative AI hub (one governed channel to the chosen model).
Orchestration masking: anonymization is irreversible; pseudonymization is restored in the answer; entity detection is uneven — English person names, US-only addresses/locations by default.
Grounding retrieves only what the caller's permissions scope, not an unscoped corpus.
"Not retained by the provider" is a retention commitment, not a residency guarantee — the model's hosting region still governs the Schrems II analysis (C137).
SAP's Product Development Schedule can allow SAP to use customer data for general R&D unless the customer's own contract restricts it — check the order form, not the trust-center page.
Three-question per-engagement checklist: which model and which region, does masking match the languages/formats in use, does the contract restrict general R&D use.

Terms used on this page

No-training commitment: SAP's stated policy that customer data is not shared with third-party model vendors to train or improve their LLMs or image models.
Anonymization: Masking method replacing personal data irreversibly; the model and the answer only ever see placeholders.
Pseudonymization: Masking method replacing personal data with restorable placeholders; the answer and tool-call arguments can show the original values again.
Product Development Schedule: The SAP contractual document setting terms for how SAP may use customer data for general product research and development.
Data residency: Where data is physically processed or stored; a separate question from whether a provider retains or trains on it.
Grounding scope: The retrieval boundary in the orchestration service, set by the caller's own data permissions rather than an unscoped corpus.

Sources

Full card available to members. What the full card adds: the full decision framework · the common pitfalls and their fix · the cheat sheet · the code blocks · the facts worth quoting.

Open in the app →