Manufacturing digital transformation is a sequencing problem before it is a technology problem. The project list is short and well understood: predictive maintenance, visual quality inspection, demand and inventory models, and process and yield optimisation. What separates a programme that moves a plant number from one that produces dashboards is the order those projects are built in, and whether the data each one needs already exists. This guide covers the four project classes, the plant metric each one moves, what has to be in place first, and a realistic order of operations.
The clearest public record of what these projects move is the World Economic Forum’s Global Lighthouse Network, which added 16 sites in June 2026 to reach 238 recognised factories. Each site publishes the metric it moved and the project class that moved it. These are the largest and best-resourced plants in the world, and most deployed 25 to 50 separate solutions to reach those results, so read the figures as evidence of which project moves which metric rather than as a forecast for a single mid-market site.
Why sequencing decides the outcome
Each of the four projects depends on data a plant either already collects or has to start collecting. That dependency, more than model sophistication, sets the calendar.
A visual inspection model needs labelled images of defects, which means a camera position on the line and someone marking good from bad. A demand model needs order history and a clean item master. A predictive maintenance model needs failure records that name the cause rather than the repair. Where the data already exists, a first result lands inside a quarter. Where it does not, building the record is the first phase of the project, and that is measured in months.
The sequencing rule follows from that: start where the data is closest to ready, bank a number your finance team already reports, and use that result to fund the projects with longer data lead times. The digital transformation in manufacturing programmes that compound are the ones that opened with the most instrumented project rather than the most ambitious one.
The projects, and the plant number each one moves
Four project classes cover most of the available value in a discrete or process plant. Each maps to a metric a plant manager already reports monthly.
Predictive maintenance moves unplanned downtime
A predictive model estimates remaining useful life on a specific asset and names the failure mode that is developing. Deloitte’s Industry 4.0 research on predictive technologies for asset maintenance reports maintenance cost reductions of 25% to 30% and equipment downtime reductions of up to 50% at the mature end.
Rotating equipment is the usual starting point, because vibration, motor current and thermal readings show wear developing weeks ahead. The choice between a commercial platform and a model trained on your own machine data is a decision in its own right, and it usually turns on whether machine data is allowed to leave the site. Where it is not, a model built and hosted on your own infrastructure is the route that clears procurement, which is what our private, on-premise AI work covers.
Visual quality inspection moves scrap and rework
A camera and a trained model check every part at line speed and grade it against your own defect classes. The plant metrics are first-pass yield, scrap rate and rework hours, all of which sit inside the cost of quality your controller already tracks.
Two Lighthouse sites show the range. CIMC Reefer Containers in Jiaozhou reduced defects by 47% and conversion costs by 24% across more than 50 digital applications. Rockwell Automation’s Singapore site, running over 1,000 SKUs and more than 20,000 changeovers a year, reduced defects by 35% and raised units per person-hour by 43% using AI-enabled quality control alongside flexible automation.
Inspection is often the quickest route to a visible number, because the project generates its own data. You install the camera, you collect the images, and the labelling is work your quality team can already do. Our computer vision development engagements usually open with one cell and one defect class.
Demand and inventory models move carrying cost and service level
Forecasting and inventory optimisation run on transactional data rather than sensor data, which makes them viable at plants with very little instrumentation. Hitachi Vantara’s Norman site cut inventory by 50% and order-to-ship lead time by 77% after consolidating inventory visibility into a single platform. Unilever’s Haridwar site, absorbing a 50% increase in SKUs and a fourfold rise in demand volatility, cut response times by 72% and lifted service levels to 99% with AI-enabled planning and sourcing. Schneider Electric’s El Paso site raised on-time delivery from 61% to 97% and cleared $43 million in backorders.
The prerequisite here is master data rather than hardware. Duplicate item codes, inconsistent units of measure and stale supplier lead times constrain a forecast far more than the choice of algorithm does.
Process and yield optimisation moves yield and throughput
Process models read setpoints, sensor readings and quality outcomes together, then recommend operating conditions. This class carries the highest ceiling and the longest data requirement, because the model has to have seen the process run across a wide range of conditions.
Saudi Aramco’s Hawiyah gas and NGL complex increased production volumes by 26%, improved product quality by 43% and raised overall equipment effectiveness by 44% across more than 50 use cases including digital twin optimisation. DCM Shriram’s Gujarat caustic soda site, where power dominates operating cost, cut power costs by 32% and material costs by 15% with AI-enabled process control.
Plants running a historian with several years of tagged process data start from a strong position here. Plants without one build the historian first, and that is a worthwhile project on its own terms.
The four projects side by side
| Project | Plant metric it moves | Data it needs | Prerequisite | Typical sequence position |
|---|---|---|---|---|
| Visual quality inspection | First-pass yield, scrap rate, rework hours | Labelled defect images from the line | A camera position and a quality team able to label | First in most plants, because it generates its own data |
| Demand and inventory models | Inventory carrying cost, service level, lead time | Order history, item master, supplier lead times | Clean master data in the ERP | Early, and viable with no plant instrumentation at all |
| Predictive maintenance | Unplanned downtime, maintenance cost | Sensor history plus work orders naming the failure cause | Condition monitoring and disciplined downtime coding | Second, or first where failure history is already clean |
| Process and yield optimisation | Yield, throughput, OEE, energy per unit | Tagged process data across a wide operating range | A historian with multi-year coverage at usable resolution | Later, once the historian is complete and trusted |
What has to exist first: MES, historian and clean asset data
Three systems carry the data these projects consume, and their current state determines what is buildable this year.
A manufacturing execution system records what was made, on which line, in what order and against which work order. Production counts, downtime reasons and quality dispositions live there. Without one, downtime and scrap have to be reconstructed from paper or from an operator’s recollection, which caps how precise any model built on them can be.
A process historian stores time-series tags from PLCs and instruments at high frequency. It is the substrate for process optimisation and for much of predictive maintenance. The question to ask is how long it retains data and at what resolution, since a historian that rolls up to five-minute averages after 90 days will not support failure prediction.
Asset and item master data is the least glamorous of the three and the most common constraint. An asset register that names each machine consistently, records its criticality and links to its work order history is what turns maintenance records into training labels.
This is where Industry 4.0 and smart factory vocabulary becomes practical. Smart manufacturing at a level a plant can act on means the MES, the historian and the asset register agree with each other, and that manufacturing analytics runs on that agreement rather than on a spreadsheet export. Most smart factory solutions and manufacturing analytics software assume the integration already exists, so the plants that reach value quickly are the ones that closed those gaps first. Industrial data analytics is largely this work.
A realistic order of operations
For a mid-market plant starting from a working MES and a partial historian, an order that holds up looks like this.
Quarter one: one inspection cell. Pick the defect class with the highest scrap cost, install the camera, label a few thousand images and run the model in advisory mode beside the operator. You get a number, and the plant gets used to a model being occasionally wrong without production stopping.
Quarter two: maintenance data discipline alongside a demand model. Begin coding failure causes on work orders properly, which is a process change rather than a technology project, while a forecasting model runs on order history you already hold. Two workstreams, no shared dependency.
Quarters three and four: predictive maintenance on the assets with the best history. The failure coding now has six months behind it, and the earlier results have funded the work.
Year two: process and yield optimisation. With a historian that has been maintained and trusted for a full year, this becomes the project with the largest remaining upside.
Where machine and production data has to stay on site, which is common in defence work and in contract manufacturing under customer NDAs, all four can run on a private, on-premise deployment inside your own network. The sequence stays the same; the hosting changes.
Across all four, the engineering effort sits in the data pipelines, the retraining schedule and the handover to the people who will operate the models. That is where our manufacturing consulting and machine learning consulting engagements focus, and the wider set of plant use cases sits on our manufacturing AI solutions page.
Deciding which project to start with?
Bring your MES, your historian retention settings and your top scrap and downtime costs. We will come back with the project that moves a plant number first and a realistic date for it. See manufacturing AI solutions or book an AI strategy session.
