Manufacturing Operations Software: A Guide to ERP, MES and IIoT

When Savvycom’s 11-engineer team built SCG’s Smart Living IoT platform on AWS IoT Core, the number the client could point to was the cloud bill. Re-engineering the AWS infrastructure behind device connectivity and data streaming cut infrastructure costs by 20%. SCG (Siam Cement Group) is Thailand’s largest industrial conglomerate, but that platform monitors solar roofs and air-quality devices, not machines on a production line. It still belongs at the top of a factory guide. Once physical equipment talks to software, the expensive problems sit in how data moves, not in the screens people see.

Manufacturing operations software is the set of systems that plans, runs, controls and records production. In most plants that means an ERP for business planning and a manufacturing execution system (MES) for the shop floor. SCADA handles supervisory control of machines, and a quality management system (QMS) handles inspections and corrective actions. An industrial IoT (IIoT) layer carries machine data between them. This guide answers the 13 questions buyers ask most often when they plan, budget and staff a manufacturing operations software project in 2026.

The appetite is real. In Deloitte’s 2025 Smart Manufacturing survey of 600 executives at large manufacturers, 92% named smart manufacturing as their main competitive driver for the next three years. Respondents also reported net gains of 10% to 20% in production output since implementation. Those figures come from companies with more than $500 million in revenue. A 400-person plant in Binh Duong or Rayong starts from a different baseline, and the sections below are written with that buyer in mind.

1. What systems run a modern factory, and how do ERP, MES, SCADA, QMS and IIoT connect?

ISA-95 (published internationally as IEC 62264) splits a manufacturing business into levels. The model is old, and people argue about whether the pyramid still fits cloud-era plants. It remains the most useful map for one practical question: which system owns which decision, and how fast does that decision need to happen?

Data moves through this stack in three directions, and each direction has its own failure mode:

1. Downward, plans become instructions. ERP releases a production order with its bill of materials and routing. MES breaks it into dispatch lists per line and shift, and recipes or setpoints reach SCADA and the PLCs. If master data in ERP is wrong, every level below executes the wrong plan very efficiently.
2. Upward, events become records. PLC counters and sensor values flow into SCADA and a historian. MES turns them into good counts, scrap, downtime reasons and lot genealogy. It then confirms quantities and material consumption back to ERP so costing closes.
3. Sideways, quality cuts across. QMS pulls inspection results from MES or lab systems, opens nonconformance reports, and can place a lot on hold before it ships.

IIoT does not sit on one level. An IIoT platform reads machine data at Levels 1 and 2, usually through OPC UA or MQTT, and shares it with MES, analytics and AI. Values no longer have to climb the pyramid one system at a time. Engineers call that shared layer a unified namespace. Our view: the pyramid still governs control and safety, but data no longer has to respect it.

2. What is the difference between MES and ERP?

The overlap is where buyers get confused. Most ERP suites include a shop-floor control module, and for a low-volume job shop that may be enough. On high-speed lines or regulated products it runs out of road quickly. It was never designed to take a count every few seconds.

The most common mistake we see is asking ERP to do MES work. Operators type quantities into ERP screens at the end of a shift, the data arrives eight hours late, and its accuracy depends on memory. If you are still planning the business layer, the ERP implementation guide covers phases, cost and rollout order. For the execution side in more depth, including dispatching, genealogy and labor tracking, see the guide to MES functions in manufacturing.

3. How do manufacturers collect real-time shop-floor data?

Real-time data on a shop floor comes from four kinds of source, and most plants need all four:

1. Controllers. Modern PLCs and CNC controllers expose counters, states and alarms. Many do it through a native OPC UA server, or a vendor protocol such as MTConnect for machine tools.
2. Retrofit sensors. Older machines with no usable controller get current clamps, vibration sensors or photoelectric counters wired to an IoT gateway. This is cheaper than people expect, and usually the fastest route to a first OEE number.
3. Operator input. A tablet or HMI at the line captures why a machine stopped, which material lot was loaded and which parts were scrapped. Machines know that they stopped, rarely why.
4. Cameras. Vision systems turn movement and appearance into data where no PLC is watching: parts on a conveyor, pallets between work centers, labels on finished goods.

Two design choices matter more than the protocol list. Buffer at the edge, so a network outage leaves a delay rather than a gap in the OEE report. And send changes, not everything. Streaming every tag at 10 ms floods networks and cloud bills, while most MES use cases need only state changes and counts. Machines reporting to other systems without a person in the loop is machine-to-machine (M2M) communication, the foundation every later layer depends on.

Cameras deserve a separate note, because plants underuse them. In a container yard project for a global logistics provider, Savvycom used YOLOv8, PaddleOCR and DeepSORT to track containers by camera instead of manual records. Swap containers for pallets or work-in-progress carts, and the same architecture tracks material flow between factory work centers.

4. How is OEE measured and improved with software?

• Availability = run time ÷ planned production time. A line planned for 450 minutes that stops for 50 has 88.9% availability.
• Performance = (ideal cycle time × total count) ÷ run time. It exposes slow running and short stops that nobody logs.
• Quality = good count ÷ total count, where good means right first time, without rework.

The 85% OEE figure often quoted as the benchmark is commonly attributed to Seiichi Nakajima’s work on Total Productive Maintenance (TPM) in Japan. It is built from roughly 90% availability, 95% performance and 99.9% quality. Treat it as a long-term reference, not a target for month one. Use the calculator below with your own line’s numbers.

Software moves OEE in a predictable order. Automatic capture comes first, and the first honest number is usually lower than the paper one, because short stops are rarely written down. Reason codes come second, because a stop without a reason cannot be assigned to maintenance, materials or changeover. Visibility comes third. A shop floor production dashboard shows the current shift’s losses. A supervisor can then fix a jam at 10:00 instead of reading about it in a weekly report. Treat the first month of automated data as a baseline, not a performance review, or operators will learn to game the reason codes.

5. Where does AI deliver measurable value in manufacturing?

Adoption is still early. In the same Deloitte survey, 29% of large manufacturers reported using AI or machine learning at facility or network level. The use cases worth funding first share one trait: the input data exists today.

1. Visual quality inspection. Cameras with object detection models flag surface defects, missing components or wrong labels at line speed. You need a library of defect images. Run the first deployment in assist mode, where the model flags and a person confirms, until false-reject rates are known.
2. Predictive maintenance. Vibration, temperature and motor current predict bearing and spindle failures well, if you have failure history to learn from. Most plants do not. The realistic first step is condition monitoring with anomaly alerts, which pays back sooner and builds the history that prediction needs.
3. Drawings, BOMs and engineering documents. Computer vision and OCR can read engineering drawings and extract material tables, a slow manual task in every plant that builds to order. Savvycom delivered this for a large Japanese construction corporation in five months, using a YOLOv8 and PaddleOCR pipeline. Against the manual process, operational costs fell 20%, process automation improved 65% and productivity rose 35%. The client builds infrastructure rather than products, yet manufacturers face the same drawing-to-BOM problem. See the AI blueprint digitization case study.
4. Scheduling. Optimization models that sequence orders to cut changeovers help most in high-mix plants. They depend on accurate routings and cycle times, which is another reason MES data comes first.

The bet we would make in 2026 is unglamorous. Put AI on a clean data foundation, scope it to one defect type or asset class, and measure a baseline before the model goes live.

6. How much does a manufacturing software project cost?

These are engineering-only planning ranges, not quotes. For reference, Savvycom’s Vietnam-based engineering runs $25 to $45 per hour in 2026, with senior AI and data roles reaching $55. Team mix, seniority and engagement length decide where a project lands within each range.

What the table leaves out is often larger than what it includes:

• Hardware. Sensors, edge gateways, industrial PCs, cameras and lighting, plus network upgrades on the shop floor.
• Licenses. A commercial MES, historian or SCADA package, and ERP user licenses for new integrations.
• OT work by your automation integrator. PLC program changes, panel wiring and on-site commissioning.
• Cloud running costs. Telemetry ingestion and storage grow with every machine you connect. That is why the 20% AWS saving on the SCG platform mattered as much as the build itself.
• Validation. Regulated plants (medical devices, pharma, food) need documented test evidence, which can add a third or more to QA effort.

The biggest single cost driver is the number of distinct machine types and protocols, not the number of features. Twenty identical injection molding machines are one integration. Twenty machines from eight vendors and three decades are eight.

7. Should you build a custom MES or buy an off-the-shelf one?

Our honest view: many mid-size plants buy a full MES suite and use a fraction of it. Meanwhile the integration work that decides success is quoted separately. Before signing, list the five shop-floor decisions you want to improve and check which modules they need. If the list stops at data collection, OEE, downtime reasons and genealogy, a focused custom build or a light product will do. If it includes recipe management, electronic batch records and multi-site scheduling, a mature commercial MES is the safer path. Keep custom work to integrations and dashboards.

8. How do you integrate legacy machines and PLCs with new software?

Two problems catch teams off guard. Old Windows-based HMIs often cannot be patched, which turns a data project into a security project. And PLC programs on legacy lines are sometimes undocumented, the original integrator long gone. Budget a few days per machine type for discovery; it is cheaper than guessing.

9. What are the cybersecurity risks of connected factories?

Manufacturers know the exposure. In Deloitte’s survey, 55% of executives strongly agreed that unauthorized access to operational technology is a high concern. The difference from IT security is priority. In a plant, availability and safety come before confidentiality, so you cannot patch, scan or reboot the way an IT team would. The NIST SP 800-82 Rev. 3 guide to OT security is the most practical free reference on those trade-offs for SCADA, PLC and distributed control systems.

The controls that matter most when you add software to a plant:

1. Segment into zones and conduits. The ISA/IEC 62443 series groups assets into zones with similar security needs and controls the conduits between them. At minimum, put an industrial DMZ between Level 3 and Level 4, so nothing in the office network talks directly to a PLC.
2. Make cloud connections outbound only. Edge gateways should publish data out through the firewall, for example over MQTT with TLS, rather than opening inbound ports to the plant.
3. Control vendor remote access. Route machine-builder and integrator access through a jump host with multi-factor authentication and session recording. Switch it off when not in use.
4. Keep an asset inventory and PLC backups. You cannot protect devices you have not listed, and a known-good copy of each PLC program shortens recovery after an incident.
5. Use compensating controls where patching is impossible. Application allow-listing and tight network rules protect HMIs that must stay on old operating systems.

ISA/IEC 62443 also assigns responsibilities to distinct roles: asset owners, product suppliers, system integrators and service providers. Any software partner should tell you which role it plays on your project, and which requirements it therefore owns.

10. How long does a smart-factory rollout take, and in what order?

The order matters more than the durations. Teams that open with an AI pilot on unconnected machines spend the pilot building the pipeline they skipped. They then find no baseline to prove the model’s value. Start where the data is weakest, not where the demo looks best.

11. What compliance and quality standards must manufacturing software support?

Traceability is the requirement that shapes architecture most. Backward traceability answers which material lots, machines and operators went into a finished lot. Forward traceability answers where every unit from a suspect material lot ended up, which is what a recall depends on. Both need MES to record consumption at the moment it happens, not reconstructed at the end of the shift.

Quality data also has to be usable, not just stored. Applying statistical process control on live line data catches drift before it turns into scrap. A well-designed quality control system then connects those signals to holds, nonconformance reports and corrective actions.

12. How do APAC manufacturers in Vietnam, Thailand and Korea approach digital factory projects?

This is Savvycom’s reading from teams in Hanoi, Bangkok, Seoul and Tokyo, not survey data, so treat it as a field view.

• Vietnam. Foreign-invested electronics and component plants usually inherit their MES from the parent group, so local work centers on integration, supplier portals and reporting. Domestic manufacturers supplying those groups face customer audits that demand lot traceability, which often triggers their first MES-style project. Expect mixed machine generations on the same floor. Check how operator records are handled under Vietnam’s Law on Personal Data Protection, in force since January 2026.
• Thailand. Automotive, petrochemical and building-materials groups dominate. The large conglomerates run digital programs that reach beyond the factory into energy management and connected products, as SCG’s Smart Living platform shows. Thai PDPA governs personal data in those platforms.
• South Korea. Automation levels are high, and a long-running government-backed program has helped many SMEs install basic smart-factory systems. Many projects are now second-generation: replacing first MES installations, adding AI inspection and connecting plants abroad. Korean-language interfaces and PIPA compliance are standard requirements.

One practical point on time zones. Vietnam shares Thailand’s time zone and sits two hours behind Korea and Japan. Line issues that need a developer during the shift are easier to handle with that overlap.

13. How do you choose a manufacturing software development partner?

1. References you can call. Ask for a project where the partner connected physical devices to software at scale, and what went wrong.
2. Protocol and integration depth. Ask how the team would handle a 15-year-old PLC with no OPC UA server. The answer reveals whether they have done it.
3. Data model ownership. Insist that the data model, connectors and source code belong to you.
4. Security role. Confirm in the contract which IEC 62443 responsibilities sit with the partner, and which stay with you or your integrator.
5. Delivery economics. Compare blended team cost and time-zone overlap, not just hourly rate.
6. A clear statement of limits. A partner who claims to do everything from PLC wiring to AI is either very large or overpromising.

Frequently asked questions

Is SCADA the same as MES?

No. SCADA supervises and controls equipment in real time, showing operators machine states, alarms and setpoints. MES sits one level higher and manages production orders, lot genealogy, labor and performance across a shift. SCADA tells you a filler stopped; MES tells you which order, lot and operator were affected, and what the stop cost.

Can a small factory run MES without an ERP?

Yes, for a while. Small plants often start MES-style data collection and OEE tracking while planning still lives in spreadsheets or basic accounting software. The limit arrives when orders, materials and costs must reconcile automatically. At that point an ERP becomes the system of record, and MES should integrate with it rather than duplicate inventory logic.

Should factory data use OPC UA or MQTT?

Most plants use both. OPC UA is strong at machine level, with rich information models and built-in security, and many modern PLCs expose it natively. MQTT is a lightweight publish-subscribe protocol that moves data efficiently from edge gateways to MES, historians and cloud platforms. A common pattern reads OPC UA at the edge and publishes MQTT upward.

What is ISA-95 used for?

ISA-95, also published as IEC 62264, is the international standard for integrating enterprise systems with control systems. Teams use its level model to decide which system owns which function. Its object models define messages between ERP and MES, such as production schedules, material consumption and performance reports, so integration debates share one vocabulary.

What OEE score should a manufacturer aim for?

Aim first for an honest baseline, then steady improvement. The 85% figure is a widely cited benchmark. Plants measuring automatically for the first time often score far lower, because short stops and slow cycles were never recorded. Track OEE per line against its own history, and compare plants only when their definitions match.

Where Savvycom fits on a factory project, and where it does not

Savvycom is a software engineering partner for the layers above the controller. That covers shop-floor data collection, MES and ERP integration, IIoT platforms on AWS and Azure, and computer vision. That work sits within its manufacturing software development services, backed by projects such as the SCG IoT platform and the blueprint digitization pipeline above.