Walk into any integrated steel plant IT room in 2026 and the wallboard tells the same story — SAP S/4HANA or ECC running the enterprise, SAP PP driving production planning, SAP PM handling maintenance, a Level 3 MES sitting between ERP and the floor, historians pulling from Level 2, and now a growing stack of AI models writing recommendations back into all of it. The ISA-95 hierarchy that once looked clean on the whiteboard has quietly turned into a spaghetti diagram. Getting AI to add value without destabilising the ERP-MES data fabric is now the architecture problem. Book a demo to walk through a reference architecture built for exactly this.
INTEGRATION LEAD · STEEL · ERP-MES-AI ARCHITECTURE
Steel Plants Now Run SAP, ERP-PP, ERP-PM, MES, and a Growing AI Layer — Here Is How to Wire Them Together Without a Spaghetti Diagram
A clean ERP-MES-AI architecture keeps SAP as system of record for orders and financials, MES as system of execution for the floor, and AI as system of insight — with strict layer boundaries and no direct writes from the model layer into ERP master data.
5
ISA-95 Levels the Steel Stack Has to Respect
40-70%
MES Projects Reported to Underdeliver on Scope
4
Integration Patterns Every Steel Plant Needs
THE INTEGRATION REALITY
Why the ERP-MES Boundary in a Steel Plant Was Already Hard Before AI Arrived
In an integrated steel plant, SAP S/4HANA plans the campaign, SAP PP releases production orders, SAP PM issues work orders, and SAP MII or a third-party MES translates it into shift-level execution. That interface — Level 4 to Level 3 — is one every mill has been quietly fighting for years. Production orders arrive as IDocs, confirmations flow back as B2MML messages, goods movements post overnight, and the numbers never quite reconcile.
Now add AI. Caster breakout models, BOF end-point predictors, and quality inference engines all want to read Level 2 data, correlate it with Level 3 records, and write recommendations back somewhere. The wrong answer is to let each model write into whichever system it can reach — caster model updating SAP PM notifications, quality model rewriting MES routing, all going into ERP master data. That is how a plant ends up with fourteen brittle bidirectional interfaces and a shift lead who trusts none of them.
THE REFERENCE ARCHITECTURE
The ISA-95 Stack for a Steel Plant Running SAP, MES, and AI Side by Side
The five ISA-95 levels are not a suggestion — they are the discipline that keeps ERP planning at days-to-months, MES execution at hours-to-shifts, and AI inference at seconds-to-hours from stepping on each other.
LEVEL 4 · ENTERPRISE
SAP S/4HANA — ERP, PP, PM, FI/CO
System of record for campaigns, production orders, maintenance work orders, financials, and master data. Timescale: days to months. Owns the plan and the ledger.
LEVEL 3 · MANUFACTURING OPERATIONS
MES — Order Dispatch, Genealogy, Quality, OEE
System of execution: dispatches SAP orders to the floor, tracks heat and coil genealogy, records confirmations, posts goods movements to SAP. Timescale: hours to shifts.
LEVEL 3.5 · AI INSIGHT LAYER
On-Prem AI Models & Reasoning Copilot
Reads from historian, MES, and SAP master data. Writes recommendations into a supervised action queue — never setpoints. Timescale: seconds to hours.
LEVEL 2 · SUPERVISORY CONTROL
DCS, SCADA, Historian
Owns real-time process control, alarm management, recipe execution, and the historian that AI models read. Timescale: seconds. No AI writes without engineering-of-change control.
LEVEL 0–1 · PROCESS & SENSING
PLCs, Instruments, Sensors
Physical process and instrumentation — thermocouples, load cells, mold-level sensors, drive currents. Timescale: milliseconds. Read-only for everything above the DCS.
DATA OBJECTS ACROSS THE BOUNDARY
The ERP-MES Data Objects an AI Layer Has to Read and Never Overwrite
Every architecture conversation ends at the same question: which data objects cross which interface, and who owns each one. The table below is the master-data contract most steel plants land on.
| Data Object |
Owner |
Direction |
AI Layer Access |
| Production Order |
SAP PP |
ERP → MES |
Read only |
| Heat / Coil Genealogy |
MES |
MES → ERP |
Read only |
| Goods Movement (261/101) |
MES posts to SAP MM |
MES → ERP |
Read only |
| Quality Result |
MES / LIMS |
MES → ERP QM |
Read + advisory |
| Maintenance Notification |
SAP PM |
MES → ERP |
Recommend only |
| Grade Recipe / BOM |
SAP PP |
ERP → MES |
Read only |
| Setpoint Change |
DCS / Operator |
Human only |
Recommend only |
The Integration Contract Keeps AI From Destabilising the SAP Data Plane
A clear owner and direction on every data object — and a strict advisory boundary for AI — is the difference between a reference architecture and a spaghetti diagram.
FOUR INTEGRATION PATTERNS
The Four Patterns Every Steel Plant Uses to Move Data Across the ERP-MES-AI Boundary
ISA-95 defines the layers, but not the transport. Every steel plant integration combines four patterns, each suited to a different data object and time horizon. Getting the pattern right per object is the difference between a bidirectional data fabric and a fragile daisy chain.
P1
Order Dispatch — IDoc or OData
Production orders flow from SAP PP to MES as IDocs (LOIPRO, LOIROU) or via OData services on S/4HANA. Batched or event-driven, this is the most standardised part of the stack.
P2
Event Stream — Kafka or Service Bus
Heat completions, coil ready events, and quality outcomes publish to a broker. Idempotent handlers on both sides, outbox pattern on the transactional side, and replay support keep the system honest under load.
P3
Historian Federation — Read-Only
The AI layer reads tag data from the plant historian through a scoped, read-only federation. No AI process opens a direct connection to a Level 2 DCS or PLC — the historian is the contract.
P4
AI Advisory Queue — Human-in-the-Loop
AI recommendations write to a supervised queue, not to SAP or the DCS. A qualified operator or planner promotes each recommendation into a work order, notification, or setpoint change with full audit trail.
ANTI-PATTERNS TO AVOID
The Five Integration Failures That Kill Steel ERP-MES-AI Projects Before Go-Live
Every stalled ERP-MES-AI project in steel maps to the same handful of architectural mistakes. None are software bugs — they are shortcuts taken at design time that a good integration lead catches at kickoff. The five below appear most often in post-mortems.
AP1
Bidirectional Master Data Writes From AI
An AI model updates SAP material master or PM equipment records directly. Master data drift begins immediately, ERP and MES fall out of sync, and no one can trace which system holds the correct value.
AP2
Overnight Batch Reconciliation as the Only Sync
Confirmations, goods movements, and scrap postings only reconcile in a 02:00 batch. By 09:00 the shift lead has three sets of numbers that all disagree, and the batch is running longer than its window.
AP3
Point-to-Point Interfaces With No Broker
Every new AI model gets its own bespoke pipe into SAP or MES. After six models, the plant has fourteen brittle interfaces and no team willing to touch the ones the previous integrator built.
AP4
AI Writing Setpoints Directly to the DCS
An optimizer changes burner setpoints or casting speed without an operator in the loop. Management-of-change is bypassed, the safety case is invalidated, and the next audit is a very bad conversation.
AP5
Historian Bypass for Level 2 Data
The AI stack pulls directly from PLC tags or DCS controllers, bypassing the historian and its access control. Every tag change breaks the model, and no one knows which model reads which sensor.
ROLLOUT SEQUENCE
How a Steel Integration Lead Sequences an ERP-MES-AI Rollout Without Freezing Production
The right rollout order is not "install everything at once." It is a staged sequence — get the ERP-MES contract right first, add the event backbone second, introduce the AI layer only once the ISA-95 boundary is stable. Plants that reverse this order almost always rebuild the middle.
Stage 1
Fix the ERP-MES Contract
Master-data ownership, IDoc scope, and confirmation pattern are agreed and documented before any AI conversation. The boring stage that removes 80% of downstream risk.
Stage 2
Stand Up the Event Backbone
A broker (Kafka or equivalent) is deployed between MES and downstream consumers. Idempotent handlers and outbox pattern are proven in production for at least a quarter before anything else plugs in.
Stage 3
Federate the Historian
Read-only historian access for the AI layer is scoped, tag-permissioned, and audited. Level 2 direct access is disallowed by policy and network segmentation.
Stage 4
Deploy the AI Advisory Queue
AI recommendations land in a supervised queue with full audit. Operators and planners promote each item into an actual SAP or DCS action — nothing writes without a human signature.
FREQUENTLY ASKED QUESTIONS
Integration Lead Questions on ERP-MES-AI Architecture in Steel
Does the AI layer replace SAP MII or our existing Level 3 MES in an integrated steel plant?
No. The AI layer sits alongside your MES, not in place of it. SAP MII or your third-party MES continues to own Level 3 execution — order dispatch, genealogy, confirmations, goods movements to SAP. The AI layer reads from those systems and the historian and writes only into a supervised advisory queue. If your current MES is the source of truth for a data object today, it stays the source of truth after the AI rollout.
Book an architecture session to map the boundary against your MES and SAP configuration.
How do we prevent AI recommendations from creating master-data drift in SAP over time?
By architectural rule, no AI process writes directly to SAP master data — material master, equipment master, BOM, routing, or PM tag structures. AI outputs flow into the advisory queue, where a qualified planner reviews the recommendation and, if accepted, applies the change through the normal SAP change process. This preserves the audit trail, keeps master data governance intact, and prevents the slow drift that kills bidirectional integrations.
Talk to integration support to review the write boundaries your steel plant should enforce.
Can we deploy the AI layer before we finish upgrading from SAP ECC to S/4HANA?
Yes, and most integrated steel plants in 2026 are in exactly this position. The AI layer is deliberately decoupled from the ERP version through the event backbone and advisory queue. Whether your ERP is ECC 6.0 with SAP MII, S/4HANA with embedded MES, or a mixed landscape mid-migration, the AI layer reads what is there today and adapts to the new topology when the ERP program lands. The one requirement is a stable Level 3 confirmation path.
Book a session to review your ERP migration plan and where the AI layer plugs in.
How much of the SAP integration effort is reusable across multiple steel plants in a group rollout?
A significant portion — the IDoc mappings, confirmation event schema, master-data contract, and advisory queue pattern all become group standards after the first plant. Plant-specific work then concentrates on grade recipes, historian tag mapping, and MES process variants unique to that mill. Groups that enforce the first plant as the reference architecture report second-plant deployments running in roughly half the time of the pilot.
Contact integration support to discuss a group rollout roadmap for your steel operations.
Does the AI layer need to be on-prem, or can we run it in the SAP-hosted cloud alongside S/4HANA?
For steel plants, the AI inference layer generally has to be on-prem for two reasons: latency requirements on caster and mill control loops, and data sovereignty over process recipes and heat data that most steel groups will not send to public cloud. What can sit in cloud is the analytics and reporting layer that consumes AI outputs after the fact. The typical 2026 pattern is on-prem inference tightly wired to MES and historian, with cloud analytics reading the event backbone one hop removed.
Book a session to review the on-prem versus cloud split for your plant.
READY TO REVIEW YOUR ARCHITECTURE
Give Your ERP-MES-AI Stack the Same Discipline You Enforce on SAP Master Data
A clean ISA-95 boundary, a documented data-object contract, an event backbone, and an advisory queue for every AI output — that is the reference architecture. Book a session to map it onto your steel operations and SAP landscape.
