When people talk about AI in manufacturing, they think automatic scheduling, demand forecasting, reused offcuts. Whether it works is decided earlier, somewhere much less glamorous: the quality of your ERP data. A production plan is exactly as good as the material record it reads.
Why so many AI projects stall
Gartner predicts that through 2026, organizations will abandon 60% of AI projects that are not supported by AI-ready data. In the same survey, 63% of organizations said they lacked the right data practices, or did not know whether they had them.
In a plant, AI-ready data has a very concrete face: your material and tool codes, your units, your descriptions. Everything estimators, programmers and buyers have typed in for years, each in their own way.
What we find in almost every ERP
None of this blocks everyday work, because people make up for it with their memory. AI cannot guess.
HDPE-BLACK-0500- Material
- HDPE
- Colour
- Black
- Thickness
- 0.500 in
- Unit
- Sheet
- The same material under three names. “HDPE black 1/2 in”, “HDPE BLK .5” and “HD black 0.5”: to the estimator it is the same sheet. To a system, it is three products and three prices.
- Mixed units. Inches and millimetres, sheets and pounds, sometimes in the same product family.
- Everything written into the name. Thickness, colour and grade sit in the description instead of in fields you can sort.
- Tools without parameters. The code exists, but the diameter and compatible material live in the programmer’s head.
- Dead records. Products nobody has ordered in years that clutter every search.
3D Coupe: before scheduling, we cleaned up
3D Coupe already quotes 4 times faster with AI. Estimating a part went from 58 to 15 minutes, and the team gets back close to 7 months of work a year. The result led to a second, broader mandate.
“The result was good enough that we’re going ahead with Avenor on a much broader mandate: production planning, offcuts, ERP integration.”
Planning production with AI means reading the ERP. So before writing a single line of scheduling, we extracted the 2,902 records of the reference data and went through the 496 material and tool records one by one. Result: 410 documented defects. The team received two naming guides, one for materials and one for tools, with the list of corrections record by record.
Extract everything
Measure
Write the rule
Fix it in one go
Protect what comes next
The renaming happens in a single day, after a full backup, with a single message to the team. Only then does scheduling connect to the ERP. The full story is in the phase 2 case study.
The five-step method
- 1. Extract everything. The full reference data, not a sample. Problems hide in the families nobody looks at.
- 2. Measure before fixing. How many records have a defect, of what type, in which families. That starting number sets priorities and becomes the baseline.
- 3. Write the rule with the people who use the codes. The convention is decided with the programmer, the buyer and the estimator. A rule the floor does not understand gets bypassed within three months.
- 4. Fix everything at once. Verified backup, corrections prepared ahead, renaming in one day. You avoid the period where old and new codes coexist.
- 5. Protect what comes next. Required fields, picklists instead of free text, and a check on every new record. Otherwise the cleanup has to be redone in two years.
AI already helps during the cleanup
You do not need perfect data to start using it. AI is very good at the thankless part: spotting likely duplicates, suggesting a clean name from a free-text description, flagging a unit that makes no sense for the product family. What would have taken months by hand takes a few days.
The choices stay with your team: which convention to adopt, which code to keep when two contradict each other. That is how it should be, since they are the ones who will live with it.
What clean data gives you
- Quotes at true cost. The material price comes from the ERP, not from a separate Excel list.
- A production plan that holds. The plan knows which sheets are in stock, at what thickness, and which tools are ready.
- Offcuts that get used. To reuse a leftover sheet, you need to know exactly what it is and where it is.
- Forecasts become possible. Predicting material needs requires a history where codes do not change names along the way. More in our article on predictive models.
Skipping this step means risking a system that produces a wrong plan with great confidence. People quickly stop trusting it, and they are right.
Where to start
Ask for an export of your materials and open it with the person who does your quotes. Look for a material you buy often, and count how many names it has. That small exercise already tells you whether your first AI project is a scheduling model or a naming cleanup. It is often where a good diagnostic begins.
Frequently asked questions
Do we need to change ERP before doing AI?
Almost never. The problem rarely comes from the software, more from what was put into it. A new ERP with the same naming reproduces the same problems.
Can AI clean our data by itself?
It does most of the spotting and suggests corrections. Choosing conventions and validating stays human, otherwise nobody will trust the result.
Who should lead the cleanup internally?
Someone who uses the codes every day, with management’s backing. An IT person alone does not have the shop-floor context.
Do we have to fix everything before starting an AI project?
No. You fix first the families the project will read. For production planning, that means materials and tools.
This topic falls under our AI advisory and transformation service: we analyze your operations, put numbers on the possible gains and hand you a prioritized plan.
Is your ERP ready for AI?
In 30 minutes, we look at your data with you and tell you honestly where to start.
What this article argues, we put into production for a client. The case walks through what was built, how, and what changed.
