Guide · SAP Business AI
EU AI Act and SAP Joule: What to Document Before Go-Live
EU AI Act dates after the 2026 Omnibus, what applies to SAP Joule and Business AI now, and the inventory, evidence and sign-off to collect.


If your S/4HANA or RISE program has Joule or other SAP Business AI switched on, the EU AI Act already applies to part of it, and the rest has a deadline about a year away. Before go-live, you need an inventory of every AI feature in use, a risk class for each one, the AI disclosures users see, a personal-data check, and a named human owner for each use case. This guide covers the steps, the dates as of 18 November 2026, and what a test lead should collect for sign-off.
What the EU AI Act dates look like today
The date you may have in your head is out of date. On June 29, 2026, the Council of the European Union gave final approval to the "Digital Omnibus" simplification package, formally pushing back the compliance deadline for stand-alone high-risk AI systems under Annex III. The Omnibus is now Regulation (EU) 2026/1744, and it entered into force on 27 July 2026 and moved obligations for stand-alone high-risk AI systems listed in Annex III from 2 August 2026 to 2 December 2027.
Two things did not move. The Article 50 transparency obligations requiring disclosure of AI interactions remain in force on their original August 2, 2026, schedule, and Article 5 prohibitions have applied since February 2025. Article 50 transparency duties apply from August 2026 regardless of the high-risk timeline. The Omnibus also does not remove duties. It does not remove the obligations. Treat the delay as time to build evidence, not a pause.
Also check the wording of the final text yourself. One firm notes that the new dates only bind once the Omnibus is published in the Official Journal, which it now is, but your legal team should confirm the article references against the published regulation before you cite them in a sign-off pack.
- Feb 2025Prohibitions and AI literacyArticle 5 prohibitions apply. Deployers must show staff are competent in using AI, per UNIORG's summary.
- 2 Aug 2025General-purpose AI obligationsGPAI model duties start. These sit with model providers, not with your program.
- 2 Aug 2026Article 50 transparencyUsers must be told when they interact with AI. This date has passed.
- 2 Dec 2026Watermarking grace endsGrace period for watermarking requirements on systems already deployed, per the Omnibus summary.
- 2 Dec 2027Annex III high-riskEmployment, credit, education and similar use cases must comply.
- 2 Aug 2028Annex I high-riskAI embedded in regulated products, such as machinery and medical devices.
Sources for the dates: Cloud Security Alliance research note, Technology.org, and Usercentrics. Confirm against the Official Journal text before you rely on them.
Step 1: Inventory every AI feature that is switched on
Start with what is active, not what the project plan says. Joule ships with skills, agents and grounding sources that someone may have enabled during a sprint and never documented. Organisations that do not yet know which AI features are in use start with an inventory of activated Joule skills, SuccessFactors AI features, and GenAI Hub connections.
For each item, record the following in your register:
- Feature or agent name, release, and where it runs (S/4HANA Cloud Private Edition, Public Edition, SuccessFactors, or BTP).
- The business process it touches, for example accrual proposals, procurement approvals or candidate ranking.
- The model path, including whether it goes through the Generative AI Hub, and which grounding sources feed it.
- Who enabled it, when, and who owns the outcome.
- Whether custom prompts, custom agents or Joule Studio content were built in-house.
Custom content deserves its own line. If an agent is used for a high-risk use case, high-risk obligations apply. If a deployer uses an agent in a way the provider did not intend, the deployer may become the provider of a high-risk system. A team that repurposes a standard SAP agent for a new HR question has changed its own risk position.
Under RISE, SAP is your processor, but the accountability stays with you. One practitioner summary puts it this way: as deployer, you remain responsible for risk classification, human oversight, and documentation, regardless of SAP's role as processor. The RISE testing ownership question has the same answer for AI: SAP delivers the platform, your program proves the use.
For background on what these features are and how they are licensed, read SAP AI for Practitioners. For the grounding layer you will be asked about, see Grounding SAP AI in BDC Data Products and Knowledge Graph.
Step 2: Classify each use case by risk
The Act uses tiers. Most S/4HANA assistance, such as summarizing a purchase order or drafting a text, is not high-risk. The high-risk tier matters where AI affects people's jobs, pay or access to credit. Annex III covers recruitment, credit scoring, law enforcement, education, and border control tools, so a SuccessFactors feature that ranks candidates or recommends pay moves needs a full classification review.
Use a simple test for each row in your inventory:
- Does the output decide, or strongly shape, a decision about a person? If yes, check Annex III.
- Does it touch a regulated product or safety function? If yes, check Annex I.
- Does a person see the output directly as AI-generated? If yes, Article 50 disclosure applies.
- Is it a finance or supply chain assist with a human approval step? Record the approval control as your evidence.
Record the result and the reason. A written "not high-risk, because" line with the decision owner's name is worth more at audit than a clean-looking spreadsheet.
Step 3: Check the disclosures users actually see
Article 50 is the obligation that already applies, so test it first. To promote ethical use of AI, SAP requires that all its user interfaces display a short AI notice when the user is using an AI feature or its result has been created by AI. On Joule, the same idea shows up as an AI notice on the chat screen. SAP's own documentation says users are informed through the AI Notice on the Joule screen that they should not enter any sensitive personal data into the chat.
Your test is whether the notice appears in each channel your users use: the Fiori launchpad, the Joule panel, mobile, and any custom app or embedded widget you built. Custom surfaces are where the notice most often goes missing. Log the screenshot, the release and the test date.
Step 4: Check the personal data path
Privacy and the AI Act overlap. Your GDPR review for the Joule use case should answer three questions: what personal data reaches the model, what masking happens first, and what the grounding source exposes.
On masking, the answer depends on the product. For SuccessFactors agents, SAP documents that before any user input is processed by an AI model, the SAP Orchestration Service automatically detects personally identifiable information (PII) within the input and masks or pseudonymizes the detected PII before sending data to the underlying AI model. That is a documented control you can test. Do not assume the same behavior applies to every Joule skill or to your custom prompts. Check the product-specific data protection documentation for each one.
On the chat itself, SAP advises users not to enter sensitive data. Users are advised not to submit any sensitive personal data (that is, special categories of personal data specified in Article 9 of the GDPR) to Joule. Your test should try to break that rule. Enter a made-up health note or a disciplinary reference in a test prompt, then check what is stored, what is logged and what the user sees.
Grounding is the bigger risk. SAP's own document grounding guide warns that the document grounding capability has no mechanism for determining the type of data that it processes. It also says any ingested documents are available to all Joule business users in the given Joule tenant. Most significantly, the response from Joule doesn't take context information into consideration, for example, user location and user permission. The guide's wording is "for now", so retest it at each release. A test case that logs in as a restricted user and asks about a document they should not see is the most useful check you can add to your pack.
Step 5: Set human oversight and keep the evidence
Oversight has to be provable. A policy saying "a human reviews AI output" is not evidence. You need a record that a named person reviewed a specific output, accepted or rejected it, and could stop the feature. For agents that post or change data, the control belongs in authorizations. The Joule agent authorization testing guide covers how to prove an agent cannot do more than its approved role allows.
For the evidence pack, keep these items per use case:
- Inventory row, with release and enablement date.
- Risk classification memo, signed by the business owner and the compliance lead.
- Screenshots of AI notices in each channel.
- Masking and personal-data test results, including the restricted-user grounding test.
- Evaluation set results and the rerun after each model or Joule update. The guide to testing SAP AI features after a model or Joule update lists what to rerun.
- Named human reviewer, escalation path and the stop procedure.
SAP's own governance is a useful reference. SAP says it derives its AI Governance framework based on best practice standards, guidelines, regulations, and frameworks such as the EU AI Act, NIST AI RMF, NIST CSF, ISO 27002, ISO 42001. That tells you what SAP covers. It does not tell you what your company has decided, so your register has to add that.
What to test before sign-off
- AI notice visible in every channel and every role that can reach the feature.
- Restricted-user test: a user without permission cannot get an answer that uses data they cannot see in S/4HANA.
- Masking test: personal identifiers are masked before model calls, for features where SAP documents this.
- Prohibited-use test: the feature does not produce outputs for uses your policy bans.
- Regression rerun after each release, model change or grounding source change.
- Escalation test: the reviewer can stop or reject an output, and the event is logged.
For people who want structured practice, the free SAP Academy course AI Governance for SAP Programs: EU AI Act, SAP Responsible AI and Agent Controls walks through this material. To go deeper on evaluation sets and non-deterministic results, take Testing Joule Agents and AI Features in SAP.
Frequently asked questions
Does the EU AI Act apply to SAP Joule in S/4HANA?
Yes, if your company deploys the AI in the EU or uses it on people in the EU. The Act's classes apply to the use case, not to the vendor. Joule's general assistance is usually not high-risk, but Article 50 disclosure applies to any system where users interact with AI.
Is the high-risk deadline still 2 August 2026?
No. Annex III high-risk obligations moved to 2 December 2027, and Annex I obligations to 2 August 2028, under Regulation (EU) 2026/1744. Article 50 transparency still applied from 2 August 2026, so it is already in force.
Who signs off the AI use case in an SAP program?
The business owner signs the use case and its risk class, the compliance or data protection lead signs the personal-data review, and the test lead signs the evidence. The program director signs the go-live decision. Keep the names in the register, not only in email.
What is the first thing to do this week?
Run the inventory. Ask each functional lead which Joule skills, SuccessFactors AI features and GenAI Hub connections are active in production and test. You cannot classify or test what you have not listed.
Sources
- EU AI Act High-Risk Deadline Pushed to December 2027, Cloud Security Alliance, 2026
- EU AI Act Deadlines After the Digital Omnibus, Technology.org, October 2026
- EU AI Act Deal: Digital Omnibus Now in Force, Usercentrics, 2026
- EU AI Act Omnibus Agreement: Postponed High-Risk Deadlines, Gibson Dunn, 2026
- What Article 50 of the EU AI Act Means for Enterprise AI, SAPinsider, 2026
- Responsible AI, SAP, accessed November 2026
- Responsible AI in Practice: Assessing the Tools We Use Daily Against SAP's AI Ethics Policy, SAP Community, November 2026
- SAP & the EU AI Act: Risks & Obligations, UNIORG, 2026
- EU AI Act from August 2026: What SAP Business AI Users Must Document, Finoptory, 2026
- What is Joule, SAP Help Portal, updated 20 October 2026
- Set Up Document Grounding, SAP Help Portal
- SAP SuccessFactors AI agents: PII protection, SAP Help Portal, 2026
- Principles to Practice: Securing SAP Business AI, SAP Community
- From Prompt to Permission: How Joule Securely Accesses SAP Cloud ERP Private, SAP Community
Free S/4HANA testing templates
Master test strategy, migration test plan, cutover runbook, go/no-go checklist and more.
The RISE Testing Playbook
One practical move a week for testing your ECC to S/4HANA transformation.
More on SAP Business AI
- Cloud ALM AI Requirement Generation for Fit-to-Standard
- Grounding SAP AI in BDC Data Products and Knowledge Graph
- Joule for Developers Not in ADT? S/4HANA Setup Steps
- Test SAP AI Features After a Model or Joule Update
- SAP AI Units: How to Estimate Business AI Cost
- Joule in SAP Ariba: Setup, Entitlement and Test Checks
- SAP AI Core Resource Group & Orchestration 404 Fix
- Accounting Accruals Agent: How to Test AI Accrual Proposals
How this guide is made: researched with AI-assisted search across SAP documentation and practitioner sources, with every fact linked to its source. Found an error or have field experience to add? Tell us.