# How to Add AI to Siemens Opcenter Execution (MES) On-Prem > Source: https://netray.co/erp-help/private-ai-layer-siemens-opcenter-execution-on-prem-architecture > Publisher: Netray (netray.co), AI and integration engineering for ERP and manufacturing systems. > Section: ERP Help ## Short answer Opcenter Execution is a family with different data models, so start by identifying your variant (Foundation, Discrete or Process) and its supported service and reporting interfaces. Read through those, or through a reporting copy of the database, behind a read-only tool server. Serve an open-weight model on local GPUs and send any write-back through the MES's own transactions with approval. ## Kind advanced ## System mfg-systems ## Erp Siemens Opcenter Execution (MES) ## Module Opcenter Execution Foundation / Discrete / Process ## Query how to add AI to Siemens Opcenter Execution on-prem ## Alternate Queries - Opcenter Execution AI integration without cloud - Camstar MES AI assistant on-prem - Opcenter Execution Foundation API read-only - private LLM for Siemens MES data ## Applies To Opcenter Execution Foundation (Camstar lineage), Opcenter Execution Discrete and Opcenter Execution Process, on-prem. Interfaces and database engines differ by variant and version; confirm against Siemens documentation for your release. ## Steps ### Reference build, in order **Items:** - Identify exactly which Opcenter Execution variant and version you run, and which database engine hosts it. - List the supported read interfaces for that version: service APIs, the integration product your plant uses, and reporting views or a reporting database maintained by Siemens or your integrator. - Create a read-only identity and, where the data model is heavy, stand up a replicated reporting copy for analytical questions. - Define the business objects the assistant needs (lots or containers, work orders, operations, resources, specs, nonconformances, history) and map each to one read operation. - Index MES-adjacent documents (work instructions, SOPs, deviation records) with source permissions. - Serve an open-weight model with vLLM or Ollama on local GPUs and connect it only to the tool server. - Mirror MES roles, factories and resource groups into allow lists and log every call. - Roll out read-only, then drafted outputs, then approved MES transactions through the supported interface. ## Sections ### Know your variant before you design anything **Body:** Opcenter Execution Foundation comes from the Camstar platform and is a model-driven system: factories, products, specs, workflows, resources and containers (lots) are configured as objects, and every shop-floor transaction is recorded against them, producing history for traceability. Opcenter Execution Discrete and Process are different products with their own data models, and Siemens documents integration points for each. Because of that, a generic list of table names would be wrong for some readers. The reliable approach is mechanism-first: use the supported service APIs for transactional and object reads, use the Opcenter reporting and intelligence offerings or a reporting copy for analytical reads, and have your integrator document the views you will rely on. For ERP connectivity, Siemens offers an integration layer (Opcenter Connect MOM) that plants commonly use to exchange orders and confirmations with ERP; if you have it, it also tells you how orders and lots map across systems. **Code:** # Pseudo-request to a tool server operation (shape only; the underlying call uses your version's documented service API) POST https://ai-tools.plant.local/tools/get_lot_history { "lot_id": "L-2026-004512", "max_events": 200, "include": ["operations", "resources", "nonconformances"] } # The tool server maps this to a read-only call against the supported API or reporting copy, # enforces the caller's factory and resource-group allow list, and returns a compact summary. ### Reference architecture **Body:** Run the AI stack beside your MES reporting tier. A read-only reporting copy (or a read replica where your database engine and Siemens support it) keeps analytical questions off the transactional database that operators depend on. The tool server calls supported APIs for object lookups and the reporting copy for aggregates. The retrieval index holds work instructions, specs descriptions, deviation and CAPA documents, and an export of MES configuration names (products, operations, resources) so users can ask in plain terms. Keep the lot and WIP data itself out of the index and read it live through tools. In an air-gapped plant, the whole stack runs unchanged on your own servers, with models and containers loaded from approved media and logs kept in the isolated zone. ### Permissions and traceability obligations **Body:** MES roles typically scope what users can do per factory, operation or resource group. The assistant should carry the user's identity into the tool server and apply the same scoping to every lookup and every retrieved document. If the MES cannot be called as the user, fall back to a service identity with an allow list derived from the user's roles. Regulated plants (aerospace, defense, medical devices) have record-integrity and traceability duties. The assistant should be read-only against the electronic record, and any information it produces should be clearly labelled as a draft that is not part of the controlled record until a person commits it. Keep an audit trail of questions, sources and tool calls so you can show who looked at which lot data and when. ### Write-back, sizing and phases **Body:** Write only through MES transactions exposed by the supported interface, with a person approving the exact payload: for example opening a nonconformance, adding a comment to a lot, or placing a hold. Never insert or update MES tables directly, which would bypass business rules and history. As a range, one 24 to 48 GB GPU serves tens of occasional users; hundreds of active users usually need two to four GPUs. The document index for tens of thousands of files is a few GB to low tens of GB. Phase 1: read-only lot, WIP and nonconformance questions for one factory. Phase 2: drafted deviation reports and shift summaries. Phase 3: approved holds, comments and requests via the MES's own transactions. ### Limits and when Siemens tools are enough **Body:** If your users need standard operational reporting, Opcenter's own reporting and intelligence tools may cover it, and the assistant is optional. The AI layer helps when questions cross systems (MES, ERP, quality documents), or when users cannot be expected to know the data model. Expect integration effort to be heavier than for SQL-first SCADA systems because the data model is rich and version-specific. Do not over-promise a single answer across MES and ERP until each read operation is tested. Siemens also offers its own AI-related features; evaluate them first if your data policy allows. ## Pitfalls - Designing against table names from a blog post that match a different Opcenter variant or version. - Querying the transactional MES database for analytical questions during production. - Writing to MES tables directly instead of using supported transactions, which breaks history and business rules. - Presenting AI-drafted text as part of the controlled electronic record. - Ignoring factory and resource-group scoping in retrieved documents. - Indexing live WIP data, which goes stale within minutes. Read it through tools instead. ## Ai Angle Netray builds the tool server and retrieval layer against the interfaces you approve, and joins MES data to ERP records through ERPray so a held lot can be tied to its work order, supplier lot and open purchase order. Sources and tool calls are shown to the user. ## Faq ### Is Opcenter Execution Foundation the same as Camstar? **Answer:** Foundation descends from the Camstar platform, but product names and versions have changed over the years. Always confirm interfaces and database details for your specific release in Siemens documentation. ### Can we read the database directly? **Answer:** Only through views or a reporting copy that Siemens or your integrator has documented for your version, and only read-only. Prefer the supported APIs for object lookups, and never write to the tables. ### How do we keep records compliant? **Answer:** Keep the assistant read-only against the controlled record, label its output as draft, and commit changes only through MES transactions that a person approves, so history and electronic signatures stay intact. ### Can it work air-gapped? **Answer:** Yes. Models, embeddings, the index, the tool server and the audit log all run on your own servers, with software brought in on approved media. ## Search terms AI for Siemens Opcenter, Opcenter Execution AI integration, Camstar MES AI, Opcenter Execution Foundation API, local LLM MES, air-gapped MES AI, Siemens MES on-prem AI, Opcenter Connect MOM ERP, MCP server MES, private AI traceability --- Talk to an engineer: https://netray.co/contact?intent=erp_ai | Machine catalog: https://netray.co/agent/catalog.json