Oct 15 2026
/
Manufacturing AI Projects: Fix Your Data Foundations First
Quick answer: Most manufacturing AI projects stall because the data behind them is incomplete, inconsistent, or trapped in separate systems, not because the AI itself fails. The fix is to start from one business problem, identify the data it needs, and clean and connect that data first. AI then has a reliable foundation, and each new use case builds on the last.
Manufacturing AI is moving from conference slides to real factory floors. Demand forecasting, predictive maintenance, quality prediction, and AI assistants inside ERP are all practical today. Yet many manufacturers find their first AI pilot stuck after an exciting demo. In fact, Gartner predicts that through 2026, organizations will abandon 60% of AI projects unsupported by AI-ready data.
The pattern is familiar to mid-market manufacturers in Malaysia and Singapore. For example, a model works on a clean sample dataset, then struggles with the factory’s real data: item codes that differ between systems, machine readings with gaps, and quality records still on paper.
Why do manufacturing AI pilots stall?
They stall when the data isn’t ready for the decision the AI is meant to support. Common reasons include:
- Data lives in silos: ERP, spreadsheets, machine controllers, and paper logs each hold part of the picture, and nothing connects them.
- Master data is inconsistent: the same part has different codes, units, or descriptions in different places.
- Machine data is missing: many machines aren’t connected, so downtime and output come from manual records.
- Context is lost: a temperature reading means little unless you know which order, product, and shift it belongs to.
- Nobody owns the data: teams fix errors in reports but never at the source, so they keep coming back.
In other words, none of these are AI problems. They are data problems that AI makes visible.
What does AI-ready data look like?
AI-ready data passes four simple tests:
- Accurate: master data, such as items, bills of materials (BOMs), routings, and customers, is correct and maintained.
- Connected: ERP, shop-floor, and machine systems share common identifiers, so records link together automatically.
- Captured automatically: machine states, output, and quality results are recorded as they happen, not typed in later.
- In context: every data point carries the who, what, and when, linking machine data to orders, products, and shifts.
However, you don’t need perfect data across the whole business before you start. You need data that passes these tests for the specific problem you want to solve.
Which manufacturing data matters most for AI?
It depends on the use case, but five types of data come up again and again:
| Data type | Where it lives | Example AI use |
|---|---|---|
| Master data | ERP | Every use case; errors here spread everywhere |
| Transactions | ERP | Demand forecasting, inventory optimization |
| Machine data | Controllers, sensors, IoT platforms | Predictive maintenance, energy optimization |
| Quality records | Quality systems, MES, paper | Defect prediction, root-cause analysis |
| Maintenance history | Maintenance systems, logbooks | Failure prediction, spare parts planning |
Connecting machines through IoT and recording production in a manufacturing execution system (MES) fills the two biggest gaps most factories have: machine data and context.
How do you build manufacturing AI data foundations without a big program?
Work backward from one use case instead of trying to fix all data at once:
- Choose one business problem with a clear outcome, such as reducing unplanned downtime on a bottleneck line or improving forecast accuracy for a key product family.
- List the data that decision needs, and where each piece lives today.
- Clean the master data in scope, not the whole database.
- Connect the systems involved, so data flows automatically with shared identifiers.
- Assign data owners who fix errors at the source.
- Run the AI pilot on the prepared data, and measure the business outcome.
As a result, each use case leaves behind cleaner, better-connected data, so the next project starts further ahead. For more on where AI fits in operations, read integrating AI for smart manufacturing.
Where does generative AI fit in manufacturing?
Generative AI assistants built into ERP can answer questions, summarize exceptions, and draft documents in plain language. However, they are only as reliable as the data they draw on.
An assistant asked “which orders are at risk this week?” gives a useful answer only if order dates, inventory, and capacity data are accurate and current. In short, the same foundations that support predictive models also make generative AI trustworthy. See how these trends are shaping the industry in our look at 2026 manufacturing trends.
Start with one problem, not a data program
In the end, manufacturing AI succeeds or stalls on its data. Pick one business problem with a clear outcome, prepare only the data that decision needs, and make sure it is accurate, connected, captured automatically, and in context. Each use case then leaves cleaner data behind, so the next project starts further ahead. A practical first step is a short data assessment on the problem you most want AI to solve.
Frequently asked questions
Do we need a data warehouse before starting with AI?
Not necessarily. Many first AI use cases can run on data from ERP and connected machines, prepared for that specific problem. A data warehouse or data platform becomes valuable as the number of use cases grows and more teams need the same trusted data.
How long does it take to get data AI-ready?
It depends on the use case and the state of your systems. Scoping to one problem keeps the effort manageable, because you clean and connect only the data that decision needs. A data assessment at the start shows how big the gap is.
Can we start AI if our machines aren’t connected?
Yes, depending on the use case. Forecasting and inventory optimization rely mainly on ERP data. Use cases such as predictive maintenance need machine data, which you can add through IoT sensors and gateways on existing equipment without replacing it.
Who should own manufacturing data?
The business teams that create it. Production owns production records, quality owns inspection data, and planning owns demand data. IT provides the platforms and integration, but data accuracy improves fastest when the people closest to the work own it.
As an Infor Gold Channel Partner, Phitomas helps manufacturers in Malaysia and Singapore build AI-ready data foundations across Infor CloudSuite Industrial, IoT, and MES. Talk to our team about your first AI use case →