DocUX4T-AB-015Ver1.0ClassPUBLICIssued2026-06-28Reviewed2026-06-28Read~9 min Living reference
UX4TECH· Architecture Brief · Series III · No. 015
Architecture Brief / for SAP customer architects, CIOs & CISOs
Grounding SAP Joulethe Knowledge Graph, the gaps, and the readiness path.
Joule is the headline. The Knowledge Graph is what
makes it trustworthy instead of generic — it grounds the model in SAP's
semantics so answers reflect your business, not the open internet.
But a map of SAP's world is not a map of yours. The last mile —
your data, your correctness bar, your audit trail — is still on you.
Document typeArchitecture Brief / L1
AudienceCIO · CISO · Platform Lead
DomainEnterprise AI Architecture
SourcesSAP Learn · AI Foundation · 2026
grounded still on you
§1 The briefing Read · 60s
What you need to know in sixty seconds.
Grounding1.1
Grounding is not one feature — it's a stack.
Five layers sit under every grounded Joule answer, from the Knowledge Graph to your own documents. Each carries its own readiness and metering implication.
Context1.2
The Knowledge Graph grounds SAP's world.
It makes entities and relationships explicit across roughly 452,000 tables and 7.3M fields — automatically. That part you get for free.
Last mile1.3
Grounding your world is the project.
Document Grounding and Company Memory bring your data in — and that ingestion, curation and limits are where the real work lives.
Trust1.4
Grounding ≠ verification.
It reduces hallucination; it doesn't certify the number. Anything feeding a control, a close, or a filing needs a verification layer on top.
Cost1.5
Grounding is metered.
Every grounded call draws AI Units; agentic tasks draw materially more. Plan consumption governance before you scale.
Purpose1.6
This brief is the working map.
The stack, two reference architectures, three line-of-business use cases, the gaps, and a readiness plan — with the door kept open with SAP.
§2 Under the answer Five layers
The grounding stack.
Joule does not "just call an LLM." Each grounded answer draws on a stack — and each layer carries its own readiness and metering implications.
01
SAP Knowledge Graph
Makes entity meanings and relationships explicit to reduce hallucination; built on a graph spanning roughly 452,000 ABAP tables and 7.3M fields (the published breakdown also cites ~80,000 CDS views). The source of SAP-process context.
02
Domain Models + SAP-RPT-1 GA planned Q3 2026
SAP-process reasoning plus a foundation model purpose-built for tabular ERP data. In Early-Adopter today.
03
Generative AI Hub
The model-access layer. Every grounded call meters against AI Units.
04
Document Grounding / Custom Knowledge Grounding
Grounds your documents via SAP AI Core — ingestion handles thousands of documents (per-pipeline and per-file limits apply).
05
Company Memory
Extends the Knowledge Graph with your policies, SOPs and tribal process knowledge.
// ARCHITECT'S NOTE — Layers 01–02 ground SAP's world automatically. Layers 04–05 — grounding your world — are where the real project work lives. Most "Joule isn't accurate for us" complaints trace to 04/05 being under-built, not to the model.
§3 Two ways to draw it Fig. 3.1 — 3.2
Reference architectures.
Two reference flows. RA-1 shows what happens at answer time — native grounding inside the tenant perimeter. RA-2 shows what has to be true beforehand: the AI-native data layer is load-bearing; if it is weak, every downstream answer is weak.
RA-1 — Native grounding flow. A user question enters Joule, which hands off to a grounding orchestration node that fans out in parallel to three retrieval paths — GraphRAG over the Knowledge Graph, vector RAG over HANA Cloud, and Document Grounding over your own documents. Results are fused into a single grounded, cited answer. Everything from Joule onward lives inside the dashed SAP tenant / BTP identity perimeter — retrieval never leaves the trust boundary.
RA-2 — Grounding-readiness flow. Before Joule can ground anything, raw sources are extracted, then cleansed and shaped, then landed in the AI-native data layer — drawn deliberately larger and tagged ★ load-bearing. The grounding pipeline (stages 04–05) indexes and serves from that layer up to Joule. The shape makes the thesis literal: if the data layer is weak, every downstream answer is weak. This is the gap most readiness assessments miss.
§4 Across the business UC-1 — UC-3
Three use cases.
Chosen across lines of business to outlast any single release. Each follows the same shape: grounding supplies the context and the speed; the certified number, the cross-system signal, the audit trail and the curated corpus are the readiness layer on top. The stack-maps below make the shape literal — grounding at the base, the gap floating on top.
UC-1Finance
Period-end close & financial reporting
The Financial Closing Assistant orchestrates Joule Agents — a GL period-end agent surfaces parked documents and open GL items by company code; a journal-entry agent trains on your policy and posting history, so controllers review exceptions, not every line. (Suite GA planned across 2026.)
The gap"Audit-ready" inside the system is not a certified consolidated number — cross-entity reconciliation and an audit trail an external auditor independently accepts are a verification layer on top.
UC-2Supply chain
Production planning & order release
The Production Planning & Operations Agent checks material, capacity and scheduling availability, recommends workarounds, and can automate the release of production orders. (GA planned Q2 2026.)
The gapThe signals that break a plan are often cross-system — supplier status, non-SAP logistics, disruption — and regulated manufacturing needs a traceable "why released or held" record.
UC-3HR · SuccessFactors
Employee service & people intelligence
An HR Service agent answers policy questions; a People Intelligence agent surfaces retention, comp and skills trends — grounded on your HR policies via Custom Knowledge Grounding plus SuccessFactors data. (HR agents planned H1 2026.)
The gapThe your-data last mile bites hardest here — policy-corpus currency, PII boundaries, and a real accuracy bar for recommendations that affect actual people.
UC-1 — Finance period-end close. A Financial Closing Assistant delegates to three grounded agents (GL Period-End Close, Journal Entry, Accrual) that produce a fast, explainable "Gold / close" output at the base. Floating on top: the GAP — the certified consolidated number and the external-auditor trail. Grounding gets you to draft-grade quickly; the certified, audit-defensible number is the layer it does not hand you.
UC-2 — Supply chain. A Production Planning & Operations Agent runs three grounded checks (material, capacity, scheduling) that produce a proposed order release at the base. The GAP layer on top — the trusted cross-system signal and the decision record — is what grounding does not give you for free. Same shape as UC-1: fast grounded base, readiness gap floating above.
UC-3 — HR / SuccessFactors. An HR Service Agent and a People Intelligence Agent ground on a custom knowledge corpus (HR policies via Document Grounding + SuccessFactors data) to return cited, self-service answers at the base. The GAP on top spans three concerns — corpus currency, PII boundary, and a people-impact accuracy bar. The wider gap band makes the point: people-data answers must clear all three before they ship.
§5 A readiness checklist, not a verdict G.01 — G.04
What's still missing in production.
G.01
Grounding reduces hallucination; it doesn't certify correctness.
SAP's strongest accuracy claims are scenario-specific. Anything feeding a control, a filing, or a board number needs a verification layer on top.
G.02
Part of the stack is still roadmap.
Domain Models are Early-Adopter, GA planned Q3 2026. Designs that assume full native grounding today should phase to the GA dates.
G.03
Grounding is metered.
Grounded answers draw AI Units; agentic multi-step tasks draw materially more. The cost meter grows with adoption — plan consumption governance early.
G.04
The your-data last mile is manual.
Native SAP grounding is automatic; grounding your estate (layers 04–05) means ingestion pipelines, curation and limits. This is the work — and the differentiator.
§6 Where UX4Tech helps
We keep you moving with SAP — not around it.
These gaps aren't reasons to wait on SAP. They're a readiness plan. As an SAP partner, we work two ways — both of which make your SAP investment land harder.
SVC.01
Grounding-readiness assessment
Map which Joule use cases need certified numbers vs. contextual answers, and what your data estate needs before grounding scales.
SVC.02
Verification & $-quantification layer
Complementary architecture alongside Joule — a certified-number and audit-trail layer across finance, supply chain and HR.
SVC.03
Roadmap-ready design
Domain Models, Company Memory and agent grounding designed into your estate now, so you're live on day one of GA.
We don't replace Joule or the Knowledge Graph. We make sure your data, your numbers and your audit trail are ready for them.
Inventory which Joule use cases need certified numbers vs. contextual answers — they carry different readiness bars.
Stand up the AI-native data layer before scaling Document Grounding, not after.
Put AI-Unit consumption governance in place before agentic Joule rolls out.
Phase any design depending on Domain Models or Company Memory to the GA window.
Define the verification layer for anything feeding a control, close, or filing.
§8 SAP-published sources R.01 — R.05
Sources backing this brief.
R.01· Cited §1, §2
SAP Knowledge Graph — AI Foundation overview
Entities, relationships, and the grounding substrate for Joule.
R.02· Cited §2
SAP Domain Models — Early Adopter Care
SAP-process reasoning; GA planned Q3 2026.
R.03· Cited §2, §5
Document Grounding — SAP AI Core service guide
Pipeline / Vector API, repositories, chunking and embedding.
R.04· Cited §4
SAP Business AI — Q2 2026 release highlights
Financial Closing, Production Planning, SuccessFactors agents.
R.05· Cited §1, §5
SAP AI Units — consumption & metering
How grounded and agentic calls draw against AI Units.
§9 The brief evolves Last reviewed 2026-06-28
Tracking SAP as it ships.
When SAP ships material changes to the grounding stack — at Sapphire, TechEd, or in release highlights — we log the delta here with citations. The body is not rewritten unless a foundational assumption breaks; updates accrete so you see what changed, when, and against which source.
Baseline
Foundational assumptions established: grounding ≠ verification; the your-data last mile is the project. The body is rewritten only if these break.
Baseline2026-06-28
§4 UC-2
Production Planning & Operations Agent GA restated Q1 → Q2 2026 (SAP Hannover Messe, Apr 2026). Designs sequencing to Q1 should re-baseline.
SupersedesApr 2026
§4 UC-3
SuccessFactors HR Joule agents = planned H1 2026. Only the Performance & Goals agent is GA today; HR Service and People Intelligence are forward-dated.
New2026-06
Grounding answers one question well: does the AI have the right context? It does not answer the next: is the output correct, auditable, dollar-quantified — and running on data that's ready?
The Knowledge Graph is a genuine advance. The next layer — trust on your own data — is the work. SAP gives you the engine; readiness is what makes it drive.