# AI for Rockwell FactoryTalk ProductionCentre and Plex MES, Without Exposing Plant Data > Source: https://netray.co/erp-ai/rockwell-factorytalk-plex-mes-ai-integration > Publisher: Netray (netray.co), AI and integration engineering for ERP and manufacturing systems. > Section: AI for ERP ## Short answer AI for Rockwell MES puts a private language model over FactoryTalk ProductionCentre's relational data and web services, or over Plex MES through its REST APIs, so supervisors and quality staff can ask plain-English questions about orders, downtime, genealogy and quality. The model is grounded in your records, shows its query, and does not write back without human approval. ## Kind platform ## Erp Family mfg-systems ## Erp Products - Rockwell FactoryTalk ProductionCentre - Plex MES - FactoryTalk Historian ## Industries - Automotive - Aerospace - Food and Beverage - Industrial Machinery ## Region Global ## Persona Plant Operations Manager ## Eyebrow FactoryTalk and Plex MES + AI ## Intro Rockwell Automation's MES story has two different engines. FactoryTalk ProductionCentre, which has its roots in the Datasweep product, is a configurable on-premises execution system built on a relational database and a business object layer. Plex, acquired by Rockwell in 2021, is a cloud-native manufacturing platform where MES, quality and ERP functions share one multi-tenant data model and are reached through APIs. Plants often run one, the other, or both after acquisitions. The two engines present very different integration problems for AI. ProductionCentre is typically on-site, customized by an integrator, and wired to Allen-Bradley PLCs, FactoryTalk Historian and an ERP such as SAP or Oracle. Plex is hosted, so the question becomes which APIs and extracts are available, what the vendor's own analytics already cover, and whether a customer's data policy permits piping production records to a separate AI service. Rockwell and Plex are both building AI features, and for routine analytics they may be enough. The gap tends to appear when a plant wants cross-system answers: why did a line miss its rate, combining PLC downtime from the historian, quality results from the MES, and material availability from the ERP, in one grounded response that a supervisor can read in the shift meeting. This page explains how a private AI layer connects to each engine, what it should and should not do, and which questions to ask any vendor before they touch production or quality data. ## Pain Points ### Two MES engines, one plant floor **Detail:** After acquisitions, a company may run ProductionCentre at one site and Plex at another. Reporting is rebuilt per site, and corporate asks for cross-site comparisons that nobody can produce without a spreadsheet exercise. ### Downtime reasons do not match reality **Detail:** Operators select a coded reason in a hurry, and the real cause sits in PLC fault history or free-text notes. Pareto charts built on codes alone tell an incomplete story. ### Historian data and MES data live apart **Detail:** Time-series from FactoryTalk Historian and transactional MES records join poorly. Answering why a batch failed means correlating tag trends with lot and order data by hand. ### Customized business logic is poorly documented **Detail:** ProductionCentre sites often depend on an integrator's custom business objects and reports. When that person leaves, the knowledge of how a status is derived leaves with them. ### Cloud analytics limits and data policy **Detail:** Plex data is hosted, and some customers cannot send extracts to third-party AI services under customer contracts or export controls, yet still want natural-language access. ## Use Cases ### Shift handover summaries **Description:** Generates a grounded summary of the last shift: orders completed, scrap, top downtime events and open quality holds, with each figure linked to its source record. **Erp Objects:** Production orders, downtime events, scrap transactions, holds, shift calendars **Outcome:** Replaces manual handover compilation; supervisors still review before the meeting. ### Downtime root-cause assistant **Description:** Combines coded downtime with PLC alarm history and operator notes to propose likely causes for a long stop, with evidence from each source. **Erp Objects:** Downtime reasons, FactoryTalk Historian tags and alarms, maintenance work orders **Outcome:** Faster triage of repeat stops; the model proposes, the maintenance lead decides. ### Lot genealogy and containment **Description:** Answers forward and backward trace questions across work orders, lots and shipments, joined to ERP supplier receipts. **Erp Objects:** Lot genealogy, component consumption, shipments, supplier lots **Outcome:** Cuts initial containment scoping from hours toward minutes. ### Quality results and SPC narrative **Description:** Explains why a characteristic is trending toward a limit in plain language, citing recent samples, tooling changes and material lots. **Erp Objects:** Inspection results, control charts, tool changes, material lots **Outcome:** Earlier attention to drift; control limits and responses stay with quality engineers. ### Cross-site OEE comparison **Description:** Normalizes availability, performance and quality across ProductionCentre and Plex sites and explains the largest gaps. **Erp Objects:** Run time, downtime, ideal cycle time, good and total counts **Outcome:** A defensible comparison once definitions are aligned; most effort is agreeing on OEE rules. ### Work order status Q&A for customer service **Description:** Customer service asks whether an order will ship on time and gets an answer that combines MES progress with ERP promise date. **Erp Objects:** Production order progress, sales order promise dates, shipment records **Outcome:** Fewer calls to the floor; answers show their basis and confidence. ### Maintenance and changeover insights **Description:** Looks at changeover durations and maintenance work orders to identify sequences that consistently run long and the resources involved. **Erp Objects:** Changeover events, work orders, resource calendars, SKU sequences **Outcome:** Input to scheduling and SMED efforts; benefit depends on event data quality. ## Architecture **Overview:** A private AI layer reads from the MES through the interface appropriate to each engine: curated read-only views and web services for ProductionCentre, and the Plex REST APIs or governed data extracts for Plex. It adds a Historian connector for time-series, joins ERP keys, and serves open-weight models on infrastructure you control. **Layers:** #### Engine-specific connectors **Detail:** A read-only SQL path and service calls for ProductionCentre; REST API clients with scoped credentials for Plex; an OPC UA or Historian query path for time-series. #### Unified plant model **Detail:** A mapping that normalizes orders, resources, downtime reasons and quality characteristics across engines and sites so queries mean the same thing everywhere. #### Grounded query and retrieval **Detail:** Validated, parameterized queries and document retrieval with sources cited; the underlying query is shown with each answer. #### Private model serving **Detail:** Open-weight models on local GPUs, with no outbound inference calls. For Plex, only the minimum extract needed is brought on-prem. #### Governance **Detail:** Role-based access mirroring MES and ERP roles, full logging, and human approval on any write-back or notification. ## Deployment Options ### On-prem at the plant **Fit:** ProductionCentre sites with segmented OT networks **Detail:** GPU server in the plant or regional data center, reaching the MES database and Historian over approved network paths. ### Private cloud with governed Plex extracts **Fit:** Plex customers whose policy allows data in their own tenant **Detail:** Extracts via Plex APIs land in a private environment you control, where the model runs. ### Air-gapped **Fit:** Defense and export-controlled production **Detail:** Isolated deployment with scheduled one-way data transfer and offline updates. ## Compliance ### Traceability, containment and process-control answers cite source records to support your own audit evidence. **Framework:** IATF 16949 ### Customer-controlled data stays inside your boundary; the layer logs who asked what and what was returned. **Framework:** AS9100 and customer flow-down ### Forward and backward trace queries can support your traceability plan; your records remain the system of record. **Framework:** FSMA 204 and lot traceability (food) ### The AI layer connects from the IT side through defined conduits, with read-only access and no path back into control networks. **Framework:** ISA/IEC 62443 and OT segmentation ### On-prem or air-gapped options let you meet your own controls. Compliance is determined on your environment. **Framework:** ITAR / CMMC ## Integration Notes - For ProductionCentre, prefer read-only views over its relational store and avoid querying its internal tables directly, which are subject to change between releases. - For Plex, use the Plex developer APIs with a scoped integration credential and respect API rate limits; pull incrementally rather than bulk scanning. - Connect FactoryTalk Historian through its supported query interfaces, and decide up front which tags are in scope; tag naming is rarely consistent across lines. - Align downtime reason codes and OEE definitions across sites before attempting cross-site comparisons. - Join MES to ERP on production order, part number and lot, and confirm that backflush timing matches what MES reports. - Allen-Bradley PLC context (fault codes, recipes) is best consumed through the Historian or OPC UA, not directly from controllers. ## Product Fit ### Custom build **Why:** Engine-specific connectors and a unified plant model across ProductionCentre and Plex are bespoke work. ### ERPray **Why:** Answers the ERP half of shift and order questions, such as promise dates, material availability and costs. ### DataRay **Why:** Queries quality documents and file-share records alongside MES tables, on-prem. ## Engagement ### Assessment **Duration:** 2-3 weeks **Deliverables:** - Map of MES engines, sites, Historian and ERP - Top supervisor and quality questions - Data policy and access review ### Connectors and plant model **Duration:** 3-6 weeks **Deliverables:** - Read-only connectors per engine - Normalized downtime and OEE definitions - Evaluation set of real questions ### Pilot on one area **Duration:** 4-6 weeks **Deliverables:** - Shift summary and downtime assistant - Accuracy report - Logging and approval workflow ### Rollout **Duration:** 4-8 weeks **Deliverables:** - Additional sites and use cases - Operational runbook - Training for supervisors and quality - Support plan ## Buyer Questions - Which interface do you use for ProductionCentre and which for Plex, and who owns credentials? - Does production or quality data from Plex leave our tenant, and where is it stored? - How do you reconcile downtime codes and OEE definitions across sites? - What access do you need to the Historian, and is it read-only? - How are answers verified against source records? - What does Rockwell or Plex already provide natively that overlaps? - How do you prevent a path from the AI layer into control networks? ## Faq ### Is AI for Plex different from AI for FactoryTalk ProductionCentre? **Answer:** Yes, in the connection method. ProductionCentre is usually on-prem with a relational store, so read-only views are typical. Plex is a hosted platform, so the AI layer uses its REST APIs or governed extracts. The semantic layer and the governance model are shared, but the data paths and limits differ. ### Can I use AI on Plex data without leaving the Plex tenant? **Answer:** That depends on what Plex offers for your subscription and on your data policy. If the vendor's built-in analytics meets your need, use it. If you need cross-system answers with your own models, you pull governed extracts through the APIs into an environment you control. ### Does the assistant control equipment or change schedules? **Answer:** No. It reads data and drafts explanations or suggestions. Any change to orders, schedules or notifications is approved by a person and executed in the MES or ERP. There is no write path into PLCs or control networks. ### How does it use FactoryTalk Historian? **Answer:** Through a read-only query path for selected tags, usually alarms, states and key process variables around a downtime or quality event. We agree the tag list up front, since broad ingestion of every tag is costly and rarely useful. ### What about sites running both engines? **Answer:** A normalization layer maps orders, resources and reasons into a common plant model. Most of the effort is aligning definitions such as what counts as planned downtime, not the technology. Cross-site comparisons are only as good as those definitions. ### Which models are used? **Answer:** Open-weight models in the Llama, Qwen, Mistral, Gemma or gpt-oss class, served with vLLM or Ollama on your GPUs. The choice depends on language needs, hardware, and an evaluation on your own questions, not on benchmark claims. ### What does a pilot cost in effort? **Answer:** Expect a focused pilot over roughly 8 to 14 weeks with a small joint team: one MES administrator, one quality or operations lead, and IT security. The main variables are access approvals, the number of customizations, and how consistent downtime coding is. ## Search terms rockwell factorytalk productioncentre ai, plex mes ai, factorytalk mes ai integration, plex mes api ai, datasweep mes ai, private ai rockwell mes, factorytalk historian ai, mes downtime ai assistant, plex smart manufacturing mes ai, rockwell plex ai on-prem, mes shift summary ai, cross-site oee ai --- Talk to an engineer: https://netray.co/contact?intent=erp_ai | Machine catalog: https://netray.co/agent/catalog.json