The chief operating officer of a mid-sized pump manufacturer comes back from a trade fair with three MES brochures. Each promises the same thing: real-time shop-floor visibility, an OEE figure on a screen beside every machine, no more paper job cards. The board approves the budget, the chosen vendor starts its pre-implementation analysis, and six weeks later, in a project meeting, someone asks the question nobody asked earlier: where is the MES supposed to get the routing for the impeller housing that engineering redesigned last month?

It turns out that the routings in ERP are current for roughly two thirds of the part numbers. The rest live in process engineers’ spreadsheets and in the heads of shift supervisors. The bill of materials in ERP is a hand-typed copy of the engineering structure, without the subassemblies that only exist on the shop floor. Engineering changes reach production by email. A nine-month MES project turns into a master data project, and nobody planned budget or people for it.

This article is for COOs and the people who prepare their decisions about manufacturing software. The thesis is this: the order of implementation in a plant is set by the flow of data, not by feature lists or vendor promises. You start by deciding which system owns each object, then fix the flow of engineering changes from PLM to ERP, then run one line with the full data exchange and a mock recall, and only then scale. Choosing the MES as step one usually means automating a data mess, and paying for it twice.

Why is “which MES should we buy?” asked too early?

Because it assumes the problem is a missing system, when it is usually missing agreement between the systems you already have. A typical plant runs an ERP, an engineering tool (often with document management or a full PLM system), a set of PLCs and HMIs on the machines, a quality system or inspection spreadsheets, and dozens of files filling the gaps. Each holds its own version of the truth about the same product.

The MES is the system that depends on those versions most. To dispatch an order to a work centre it needs the routing with operations and standard times. To record consumption it needs the bill of materials in the revision valid for that lot. To block nonconforming product it needs the inspection plan. If that data is inconsistent, the MES will not fix it. Either it inherits the errors and reports them faithfully, or the implementation team starts correcting data inside the MES, creating yet another separate version of the truth.

That is why the conversation about a specific product should be step three, not step one. Step one is a map of data objects and their owners. Step two is checking whether that data can reach the places that need it at the pace production changes. Only when you know what the MES must receive and what it must send back can you compare proposals in a meaningful way. Vendors given that description will price a comparable scope, rather than each pricing its own vision of your factory.

This approach is a few weeks slower at the start. In exchange it reduces the risk of discovering, six months in, that the most expensive part of the project is neither the licence nor the vendor’s implementation work, but your process engineers’ time on data that nobody put in the schedule.

How do ERP, MES and PLM differ, and where does each one stop?

The simplest way to tell them apart is by the question each answers and the rhythm it works in.

ERP answers “what, how many and at what cost”. It takes customer orders, runs material requirements planning, manages purchasing and inventory, calculates production cost and keeps the books. It works in days, weeks and months. It knows that an order for 500 units is due on Friday, but not that machine 4 has been idle for an hour waiting for a tool.

MES, the manufacturing execution system, answers “what is happening right now”. It dispatches orders to work centres, guides operators through operations and collects good quantities, times, material consumption with lot numbers, scrap, downtime and inspection results. It works in shifts, hours and minutes. It knows exactly which raw material lot went into which finished unit, but it does not calculate margin.

A PLM system answers “what is the product”. It holds engineering documentation, CAD models, the engineering bill of materials, variants, specifications and, above all, the history of changes with their approvals. It works at the pace of design projects and engineering changes, meaning weeks and months, but every decision it records eventually reaches the shop floor.

The ISA-95 model arranges the first two in a hierarchy of levels. As described by the International Society of Automation, level 0 is the physical process, level 1 is sensing and actuation (sensors, smart devices, valves), level 2 is monitoring and supervisory control (PLCs, DCSs), level 3 is manufacturing operations management, meaning MES and supervisory systems, and level 4 is business planning and logistics, meaning ERP. Popular explanations of the model, such as one by the MES vendor Symestic, add a simple time scale: level 4 plans in months, weeks and days, level 3 works in days, shifts, hours and minutes, and levels 0 to 2 react in seconds and milliseconds.

PLM has no level of its own in that hierarchy, because it sits beside it. That matters for sequencing: PLM is the source of the product definition for both upper levels, and its integration with ERP is usually the first link to break.

What does ISA-95 say about the boundary between ERP and the shop floor, and what does it leave out?

ISA-95, published internationally as IEC 62264, is a family of standards for integrating enterprise and control systems. According to ISA it now has eight parts. Part 1 (models and terminology, 2025 edition) defines the functions at the interface between control and enterprise. Part 2 (objects and attributes, 2018) describes the data exchanged between level 3 and level 4. Part 3 (activity models of manufacturing operations management, 2013) describes what happens at level 3. Part 4 covers objects exchanged between activities within level 3, Part 5 business-to-manufacturing transactions, Part 6 a messaging service model, Part 7 an alias service model and Part 8 information exchange profiles. ISA itself stresses that the standard deals primarily with the interface between levels 3 and 4.

In practice, three things from ISA-95 are worth taking. The first is a shared vocabulary: production schedule, production performance, product definition, production capability, and the resource types of personnel, equipment and material. It means the process engineer, the IT architect and the MES vendor are talking about the same thing. The second is the split of level 3 activities into four areas: production, maintenance, quality and inventory operations. That is a useful checklist which shows that an MES is more than an OEE dashboard. The third is a ready-made exchange format: B2MML, an XML implementation of ISA-95 published as open schemas by MESA International. MESA ceased operations at the end of June 2026 and ISA took over its programmes and content; the schemas remain publicly available.

What does ISA-95 leave out? It does not tell you which system should own the routing in your plant. It does not decide whether the manufacturing BOM is built in PLM or in ERP. It does not cover PLM integration at all, because that is not its subject. Nor does it say how often to send confirmations or what to do when ERP is down. Those are design decisions you have to make and write down. ISA-95 gives you a language for describing them, not the answers. A company that buys an “ISA-95 compliant” MES gets a compliant vocabulary and data model, not a finished integration with its own ERP.

Who owns the part number, the bill of materials and the routing?

This is the most important question in the whole programme and the place to begin. Every data object that moves between systems should have one owner: the system where it is created and changed. Other systems receive a copy and may enrich it with their own attributes, but they do not change the data the owner is responsible for.

The table below shows a typical split in a discrete manufacturing plant. It is not the only correct answer, but any answer has to fill in every row.

Data objectTypical ownerConsumersMost common failure
Part (item) – engineering attributesPLMERP, MESThe same product created separately in PLM and ERP under different numbers
Part – logistics and cost attributesERPMES, reportingEngineers fill in PLM fields that nobody maintains
Engineering BOM (EBOM)PLMProcess engineeringUsing the EBOM as the build structure
Manufacturing BOM (MBOM)ERP or PLM (design decision)MESNo owner, retyped by hand after every change
Routing and operations with standard timesERP, or PLM with a process planning moduleMESRoutings in spreadsheets, ERP holding values from years ago
Work instructions and inspection planPLM or quality systemMESPrintouts on the floor in an outdated revision
Production orderERPMESOrders created directly in the MES “because it is quicker”
Detailed shift scheduleMES (or a dedicated scheduling tool)ERP (status)Two schedules that drift apart within an hour
Confirmations, consumption, scrapMESERP (cost, inventory)Confirmations once a day, ERP inventory never current
Lot and its genealogyMESERP, qualityLot numbers assigned in two systems
Inspection result and nonconformanceMES or quality systemERP (holds), PLM (changes)A hold in the MES that ERP cannot see, so the product ships

Look at the row for the manufacturing BOM. It is the only one where the table gives no single answer, because this is where companies genuinely differ, and it is the subject of the next section. In the other rows the answer is often obvious, but unless it is written down, every team works from its own assumption.

Filling in this table for your own plant usually takes a few workshops with process engineering, design engineering, planning, quality and IT. It cannot be handed wholesale to the MES vendor, because it is about organisational decisions: who has the right to change a routing and who answers for it. For the broader question of which parts of data management you can outsource and which must stay in-house, see our article on data management services.

Why is the EBOM not the MBOM, and where should the manufacturing structure be built?

The engineering bill of materials (EBOM) describes what the product consists of from a functional and design point of view. The manufacturing bill of materials (MBOM) describes how it is built: in what sequence, with which intermediate assemblies that only exist on the shop floor, with consumables, with scrap allowances, and with a split between what is made in-house and what goes to a subcontractor. One product can have a single EBOM and several MBOMs if it is built in two plants or on two different lines.

The most common mistake is for the EBOM from the engineering system to land in ERP unchanged and become the structure used for requirements planning and consumption. At first nobody notices, because the differences are small. Then process engineers start adding items to the ERP structure that the design engineers never see, and at the next engineering change someone overwrites them with an import from the engineering system. Manual reconciliation after every change begins, and over time it becomes somebody’s full-time job.

So where should the MBOM be built? There are two sensible schools, and both work if applied consistently.

The first: MBOM in the PLM system. Process engineers work in the same environment as design engineers, transform the EBOM into an MBOM while keeping the links, and see straight away which parts of the manufacturing structure an engineering change touches. ERP receives a finished MBOM and routing. This requires a mature PLM deployment with a process planning module and people who will actually use it. It gives the best change control.

The second: MBOM in ERP. PLM sends the approved EBOM to ERP and process engineers transform it there. This is simpler when PLM is used mainly for document management and process engineering has worked in ERP for years. Its weakness is poorer visibility of change impact: when the design changes, ERP has to be able to show the process engineer which MBOMs need review.

There is no good third way in which the MBOM is built in the MES. The MES should receive the structure and routing, not be the place they are defined, because otherwise ERP plans and costs against a different structure from the one actually used on the line. This is one of the reasons sequencing matters: until you know where the MBOM is built, you cannot properly design either the PLM-ERP integration or the ERP-MES integration.

How does an engineering change travel from PLM through ERP to the shop floor?

The engineering change is the stress test of the whole architecture. A single master data flow can be built in a week. A change flow that stays consistent across three systems takes months, because it touches decisions that are not made in code.

The typical path looks like this. An engineering change request is raised in PLM, goes through impact assessment and approval, and the approved change order carries an effectivity: the point from which it applies. Effectivity can be a date, a serial number, a lot number, or the use-up of existing stock of the old part revision. That information has to reach ERP together with the new revision of the structure and routing, and ERP has to use it in planning: orders opened before the effective point stay on the old revision, new ones go on the new revision, and remaining stock of old parts has to be consumed, reworked or scrapped.

On the shop floor the MES has to know which revision of the structure, routing, work instructions and inspection plan applies to a given order. If the order was released on the old revision but the work instruction in the MES has already changed, the operator gets contradictory information. This is one of the most common sources of nonconformances that are hard to explain afterwards.

Three decisions should be made before any integration work. First, which kinds of effectivity the company supports and which system decides them. Second, whether a production order freezes the data revision at release (recommended) or always takes the current one. Third, who confirms, and how, that the change has reached the floor: is there a feedback loop from the MES to PLM or to quality? Without that loop nobody knows when the first unit of the new revision actually came off the line.

This part of the project has the greatest effect on the cost of nonconformance and customer complaints, and in vendor proposals it is usually covered by one sentence about “ERP integration”. Write it out as test scenarios before you choose a vendor, not after.

What messages need to flow between ERP and MES?

Once ownership is settled, ERP-MES integration comes down to a handful of streams. Write them down with frequency, direction and failure behaviour, because that document is what the MES vendor and the ERP team will build the real integration from. How to carry these messages is covered in ESB vs API vs event-driven architecture.

Down, from ERP to MES: master data (parts, structures, routings, resources), production orders with their data revision, material reservations, order changes and cancellations. Master data changes rarely and can flow in batches or as events. Orders should reach the MES within minutes of release in ERP.

Up, from MES to ERP: operation confirmations and good quantities, material consumption with lot numbers, scrap with reason codes, goods receipt of finished product, order status changes, quality holds. Here frequency directly affects how much you can trust ERP: if confirmations arrive once a day, the planner spends the whole day working with inventory that does not exist.

A few technical rules save a great deal of trouble. Every message should carry a unique identifier, so that a resend after a failure does not post consumption twice. The MES must keep working when ERP is unavailable and send the backlog in the correct order once the connection returns; the line cannot stop because the ERP server is being patched. Integration errors need a business owner and a work queue, not just a log entry that someone in IT reads once a week. And measure latency and the number of rejected messages from day one, because that is the simplest health indicator for the whole connection.

The integration technology itself, whether direct APIs, a middleware layer or an event stream, is secondary to those rules. For an overview of the patterns and their trade-offs, see our article on software integration patterns, and for the basic definition, the glossary entry on systems integration. In manufacturing the key point is that whatever you choose meets the resilience requirements above, however fashionable its name.

Where should traceability start, and how deep does it really need to go?

Traceability is often the MES benefit that is easiest to justify financially, because without it every recall is wider and more expensive than it needs to be. Before designing data capture on the floor, though, answer two questions: what is the legal minimum, and what depth makes business sense?

The legal minimum depends on the industry. In the EU food and feed sector the baseline is Article 18 of Regulation (EC) No 178/2002, the General Food Law. Operators must be able to identify anyone who supplied them with a food, a feed or any substance intended to be incorporated into one, and the other businesses to which they supplied their products, and they must make that information available to the competent authorities on demand. That is the “one step back, one step forward” principle; the UK kept the same text in its retained version of the regulation. Other sectors, such as automotive, aerospace and medical devices, have their own, usually stricter requirements from regulation and from customers, and you should check those for your sector before the project starts.

The legal minimum says nothing about what happens inside the plant. If you only know who supplied the raw material and who bought the product, a problem with a single delivery forces you to recall everything produced in that period. Internal lot genealogy, linking each finished lot to the raw material and intermediate lots that went into it and to the production conditions, lets you narrow a recall to specific orders. That is a business decision, not a legal one: the deeper the genealogy, the smaller the recall, but the more data you have to capture, and every extra operator action costs time and introduces errors.

The practical sequence is this. First, define the traceable unit, meaning what gets a number: raw material lot, container, order, individual unit. Then define the capture points, meaning the operations where the system links inputs to outputs; usually the operations where materials are combined or split are enough. Then decide how capture happens: a barcode scan, an automatic read from the machine, or a manual pick from a list. Finally, plan a mock recall: pick a raw material lot and measure how long the system takes to list every finished unit that contains it and every customer who received one. That drill is the best acceptance test for the whole integration, because it runs through PLM (product revision), ERP (receipts and shipments) and MES (genealogy).

If data is to be collected automatically from machines, an industrial internet of things layer comes into play; the glossary entry on the Internet of Things (IoT) covers the basics. Connecting the plant network to the office network safely is a separate, large subject that we do not cover here, but it has to be planned alongside the project, not after it.

Why start with one line rather than the whole plant?

Because one line with the full data exchange tells you more about the project than the whole plant with half the functions. A pilot in which the MES runs on every work centre but only records downtime tests none of the hard parts: engineering changes, lot-tracked consumption, confirmations into ERP, quality holds. A pilot on one line with all of those flows tests every one of them, and its scale lets you fix mistakes before they multiply.

Choose the pilot line on a few criteria. It should be representative, with the operations, material combinations and inspections typical of the plant, rather than the simplest line you have. It should have a stable production lead with time for the project. It should make products that genuinely change, so that at least one engineering change passes through it during the pilot. And its master data should be fixable within a few weeks.

Write the exit criteria before the pilot starts. For example: all orders for the line released from ERP to the MES without manual intervention; confirmations and consumption reaching ERP within an agreed delay; at least one engineering change carried through all three systems with matching revisions on the floor; a mock recall completed within an agreed time; the share of rejected messages below an agreed threshold. If the criteria are not met, scaling stops. It does not quietly continue “in parallel”.

This logic is close to the incremental replacement of legacy systems, where the new solution takes over areas one at a time instead of switching everything on a single day. If the plant runs an old in-house shop-floor system that has to be replaced, read about the strangler fig pattern: the same principle lets you move lines across one by one while keeping both systems consistent through the transition.

What about starting with OEE and machine data instead?

This is the strongest counter-argument to the thesis of this article and it deserves a fair hearing. Advocates of “machine data first” say: connect the machines, collect run states and downtime, calculate OEE, and within weeks you will see where the time goes. The result is fast and visible, it does not require clean master data, and it gives the board proof that digitisation pays. Structures and routings can wait.

There is a lot of truth in that. Machine monitoring really can be up and running quickly, and it often reveals losses management did not know about: micro-stops, long changeovers, waiting for material. For a plant with no shop-floor data at all it is a sensible first step that builds confidence in the programme.

The problem starts when that step is called an MES implementation and ends there. Machine monitoring tells you how long a machine ran, not what it made, from what, in which revision and for which customer. It does not shrink recalls, it does not correct ERP inventory and it does not close the engineering change loop. And if the monitoring system quietly starts taking over MES functions, for instance operators typing order numbers into it from a paper card, the company is building yet another data island that will later have to be integrated or switched off.

So the right answer is: machine monitoring, yes, but as a parallel and bounded stream, not a substitute for getting the data right. Design it so that from day one it uses the same resource identifiers as ERP and the future MES, so that downtime has a reason code list agreed with maintenance, and so that it can later be linked to orders. For the wider picture of Industry 4.0 technology, from sensors to digital twins, see Industry 4.0 in practice; this article is about what has to happen first so that those technologies have something to stand on.

Buy, extend or build: where does custom manufacturing software make sense?

The core of MES, PLM and ERP consists of mature product categories developed over decades by specialist vendors. Building your own ERP or a full PLM system from scratch rarely makes sense for a mid-sized manufacturer; we lay out the arguments for ERP in detail in Custom ERP: build, extend or buy. MES is similar, although here you more often find plants running in-house shop-floor systems that were built before the market matured.

So where do manufacturing software development services earn their keep? In four places.

The integration layer. This is the code that connects PLM, ERP and MES, transforms data, enforces message order, handles errors and keeps things running when one of the systems is unavailable. No single vendor will build all of it, because each answers for its own system. It is the most common and most justified place for custom software.

Plant-specific processes. Unusual flow control, special kitting rules, unique quality tests, recipe calculations. Where the process is a source of competitive advantage and a packaged system would need deep modification of its core, it is better to build a separate module with a clean interface to the MES and ERP.

Operator interfaces. Work centre screens, handheld terminals, apps for the production warehouse. Standard system screens are often overloaded and unsuited to gloves, noise or working on your feet. A well-designed interface reduces capture errors, and the whole traceability chain depends on the quality of that capture.

Cross-system reporting. Analysis that combines engineering, planning and shop-floor data that no single system sees in full: cost of quality per product revision, the effect of engineering changes on throughput, actual order lead time.

What does it cost? It depends on scope, but it can be estimated arithmetically. ARDURA’s custom software development page puts development costs at about €165–650 per developer day depending on seniority, and a typical three-person team (two developers and a QA engineer) at about €14,500–28,700 per month at mid to senior level. An integration layer for one pilot line is a matter of months, not years, but its cost depends heavily on the state of the master data and on whether the systems have documented, stable interfaces. If you would rather strengthen your own IT team than outsource the project, you can add external specialists to it; we describe how that model works in industry in our article on body leasing in the manufacturing industry.

How should the first twelve months be sequenced?

The plan below illustrates the order, not a standard duration. Real timings depend on the number of lines, the state of the data and the availability of people from production, who are usually the bottleneck in a programme like this.

StageIndicative durationOutputGate to the next stage
1. Data and ownership map4–8 weeksCompleted ownership table, decision on where the MBOM lives, list of data gapsBoard approves the split of responsibility; every object has a named owner
2. Data clean-up for the pilot line6–12 weeksCurrent parts, MBOMs and routings for the line’s products; downtime and scrap reason codesLine data verified by the process engineer and the shift lead
3. PLM-to-ERP integrationin parallel with stage 2Flow of approved changes with effectivity into ERPOne test engineering change carried through with no retyping
4. MES selection against the flow description6–10 weeks, in parallelRFP with message list, test scenarios and mock recallVendors price the same scope; scenarios demonstrated on your data
5. MES pilot on one line3–6 monthsFull exchange of orders, confirmations, consumption and holds; lot genealogyPilot criteria met; mock recall within the agreed time
6. Scale-up decisionafter the pilotPlan for further lines, fixes to data and integration, budgetBoard approves the next stage on results, not on a plan

Note that MES selection appears at stage four, alongside the data work, not at the start. That way the RFP contains a concrete message list, engineering change scenarios and a mock recall description, and vendor demonstrations run on your data rather than on the sample factory from the sales deck. It changes the conversation with vendors: instead of comparing feature lists, you compare how each one handles your flows.

A brief note on budget: the line most often underestimated in these programmes is neither the licence nor the vendor’s work, but the time of people from production and process engineering. Plan it explicitly, with names and a percentage of each person’s time, because without it stages 1 and 2 will stretch indefinitely and the project will reach the pilot with data that nobody has checked.

What are the key takeaways for a COO?

Digitising production does not start with choosing a system. It starts with establishing where the data that system will need comes from. MES, PLM and ERP have different roles and work at different speeds, and most implementation problems arise at the joins between them, not inside them.

Remember six things. First, before you talk to MES vendors, fill in a data ownership table for the whole plant and have the board approve it. Second, decide where the manufacturing BOM is built, in PLM or in ERP, but never in the MES. Third, write out the engineering change flow with effectivity and revision freezing at order release, because it decides quality on the floor. Fourth, document the ERP-MES messages with frequency, idempotency and failure behaviour before choosing an integration technology. Fifth, establish the legal traceability minimum, choose the depth of genealogy deliberately and use a mock recall as the acceptance test. Sixth, start with one representative line carrying the full set of flows, and scale only once written criteria are met.

Machine monitoring and OEE can run in parallel and deliver quick results, provided they use the same identifiers as the rest of the systems from the start. The best places for custom software are the integration layer, operator interfaces and the processes that make your plant different from others.

If you are preparing a data map, an MES RFP or an integration layer between PLM, ERP and the shop floor, see how we work in custom software development, or book a call. We start every project with a discovery phase that takes one to two weeks and ends with a defined scope and architecture.