SAP S/4HANA private AI layer blueprint generator
This generator turns a few facts about your SAP S/4HANA landscape into a structured architecture plan for a private AI layer: data access path, reporting copy, retrieval index, model serving, tool server, identity and audit. Export it to PDF or Word and take it to your architecture and security review.
Fill in
SAP S/4HANA On-Prem Private AI Layer Blueprint Generator
Private AI Layer Architecture Plan: SAP S/4HANA
Prepared for: Acme Precision Ltd System: S/4HANA 2022 on SAP HANA (on-premise) Hosting: On-prem server Named users: 800 Sites: 6 plants Target go-live: 2027-03-31
Context and scope
Acme Precision Ltd runs S/4HANA 2022 and wants an AI assistant that works on its own data without moving that data to an external service. The assistant is read-only first, answers with cited records, and writes back only through supported interfaces after human approval.
Main use cases: Order and delivery status, material availability, GR/IR and journal questions, quality notification search, drafted supplier emails
Data that must never leave the site: Payroll and HR master data, bank accounts, controlled technical data, unreleased pricing
Components
| Component | Design | Notes |
|---|---|---|
| Data access path | Released OData APIs (for example the purchase order and sales order services) for live reads; extraction-enabled CDS views via ODP or SLT for analytics | Prefer released and stable interfaces; avoid raw ACDOCA and MATDOC reads against production |
| Reporting copy | Approved CDS extraction views replicated to a store Acme Precision Ltd controls, aggregates plus recent periods | Line-item depth only where a use case needs it |
| Retrieval index | Attachment Service and GOS attachments, DMS documents, external SOPs; company code, plant and sales organization stored as metadata | Rebuild on authorization or org changes |
| Model serving | Open-weight model on vLLM or Ollama, Two to four data-centre GPUs | Private edition: agree the network path between GPU cluster and hosted system |
| Tool server | Named tools wrapping OData requests with typed parameters and paging | No model-built filters from free text |
| Identity and permissions | PFCG roles plus access control (DCL) on every CDS view; identity propagated or authorization values mirrored | Test with restricted users before go-live |
| Audit | Log of question, tool, request, and returned keys; idempotency keys for approved writes | Immutable storage |
Security controls mapped to your inputs
| Control | How it is met |
|---|---|
| Data stays on site | Hosting: On-prem server. Inference and index run on your premises or in the same sovereign zone as the system. |
| Stable contracts | Only released APIs and released CDS views, so upgrades do not break the tool server. |
| Row-level access | DCL on every CDS view and identity propagation for 800 named users; fail closed on unknown users. |
| Protected data | Excluded from extraction and index: Payroll and HR master data, bank accounts, controlled technical data, unreleased pricing |
| Write-back | Released APIs only, after human approval, with idempotency keys and the exact payload stored. |
Three-phase rollout
| Phase | Scope | Exit criteria |
|---|---|---|
| 1. Read-only | Read-only questions through released APIs and the replica for a pilot group | Answer accuracy on a fixed question set agreed with business owners; zero authorization leaks in testing |
| 2. Drafts | Drafts (supplier emails, exception summaries, requisition drafts) approved by a person | Drafts accepted without edit at an agreed rate; audit log reviewed |
| 3. Approved actions | Approved actions through released APIs with idempotency and audit | Approval workflow tested; rollback and idempotency verified |
Open questions for IT
1. Is the system on-premise or private cloud edition, and what network path is allowed to our GPU cluster? 2. Which released APIs and CDS views cover our use cases, and which gaps need custom CDS? 3. Which clean-core or extensibility rules apply to new CDS views? 4. How are business roles and org-level restrictions maintained? 5. What is the volume of ACDOCA and MATDOC, and what retention do we need in the replica? 6. Which GPU option (Two to four data-centre GPUs) fits the peak concurrency for 800 users? Use the sizing calculator to check.
Approvals
This is a planning starting point, not a design guarantee. Verify interface availability, object names and licence terms against your exact release and contract before building.
Review of your SAP S/4HANA AI blueprint
Optional: send the exported plan and we will comment on the data access path, permissions model and rollout order. The generator works fully without it.
Optional. The tool and every export work without this. No spam.
How to use the generator
Fill the ten fields on the left: company, version, database, hosting model, named users, sites, use cases, data that must never leave the site, GPU availability and target go-live. The plan updates as you type. Be specific in the use cases and protected-data boxes, because those drive the tool list and the exclusions the plan describes. Export the result and mark the open questions that your team cannot answer yet; those become the first workshop agenda.
- Pick the real version, since interfaces differ by release
- List protected data explicitly, one item per line
- Choose Air-gapped if no outbound internet is allowed
- Keep named users realistic, not the licence maximum
What the plan is built on
Every plan follows the same private-AI pattern: read-only access first, a reporting copy or governed interface for heavy questions, a retrieval index over documents with ownership metadata, an open-weight model served locally with vLLM or Ollama, a tool server exposing named read actions, an identity layer that mirrors the system's own permissions, and an append-only audit log. The table rows are specific to SAP S/4HANA: they name its real data access mechanisms rather than generic placeholders.
How to read and challenge the result
Treat the components table as a hypothesis. Ask your SAP S/4HANA administrators whether each named interface is enabled and licensed, whether the permission model can really be mirrored, and who owns each refresh job. Then walk the rollout table: each phase has exit criteria, and the next phase should not start until the criteria are met. The open-questions block is deliberately uncomfortable; unanswered items are where projects stall.
- Confirm licence terms for every data access path
- Test permissions with a restricted user, not an administrator
- Decide the audit retention period before go-live
- Plan for the next system upgrade from day one
Common mistakes
The most common mistake is giving the model direct database access because it is quick in a demo. It skips the application's own authorization and business logic, and it is hard to defend in a security review. The second is skipping the document index metadata, so a retrieval hit shows content the user cannot open in the source system. The third is starting with write-back. Start read-only, measure accuracy, then add drafts, then approved actions.
Frequently Asked Questions
Does the generated plan require any cloud service?
No. The plan assumes the model, index, tool server and audit log run on your own servers or in your private cloud, and the air-gapped option removes outbound internet entirely. You can still choose to use a cloud service for specific non-sensitive tasks, but the plan does not depend on one.
Is the plan accurate for my exact version?
It is a well-grounded starting point, not a guarantee. Interface availability, support package levels and licence terms vary. Verify each data access path with your administrators and the vendor documentation for your release before committing to a design.
Can I use this for a security review?
Yes, as input. The security controls table maps controls to your inputs, which gives reviewers a concrete starting point. They will add their own requirements, such as logging standards, key management and network segmentation.
What does the GPU choice change?
It changes which model sizes and how many concurrent users you can serve. Use the sizing and value calculator to estimate peak concurrency and GPUs from your own question volumes, then update the plan with the result.
If you want a second pair of eyes on the plan, Netray builds private AI layers on SAP S/4HANA and similar on-prem systems and can review your draft with your IT team.
Related Tools
SAP S/4HANA On-Prem AI Assistant Sizing and Value Calculator
Estimate peak concurrent requests, GPUs, tokens per day, index size, hours saved and payback for a private AI assistant on SAP S/4HANA.
On-Prem AIOn-Prem AI Deployment Checklist
A 30-point pre-deployment checklist covering use cases, hardware, security, model operations, and rollout for self-hosted enterprise LLMs.
On-Prem AIOn-Prem AI Security Hardening Checklist
A practical control checklist for securing self-hosted language models, covering model provenance, network isolation, data governance, host hardening, and audit readiness.
On-Prem AIOn-Prem LLM Total Cost of Ownership Calculator
Model the full multi-year cost of running LLMs on your own hardware, including GPU capex, power, cooling, support contracts, and operations staffing.
More SAP tools
See allKeep exploring
Free tools
- SAP Business One Private AI Layer Blueprint GeneratorFill in your version, database, hosting and use cases to get an architecture plan for adding a private, on-prem AI assistant to SAP Business One, with components, security controls and a three-phase rollout.
- SAP ECC Private AI Layer Blueprint GeneratorFill in your version, database, hosting and use cases to get an architecture plan for adding a private, on-prem AI assistant to SAP ECC 6.0, with components, security controls and a three-phase rollout.
- SAP PI/PO Private AI Layer Blueprint GeneratorFill in your version, database, hosting and use cases to get an architecture plan for adding a private, on-prem AI assistant to SAP PI/PO, with components, security controls and a three-phase rollout.