Podcast (practitioners-unplugged): Play in new window
Subscribe: Apple Podcasts | Spotify | RSS
Watch the episode: How Lesaffre Manages Industrial Data Across 82 Plants
Key Takeaways
- Lesaffre runs 82 plants, 34 of them large fermentation sites, with a central digital team of 5 people.
- 22 plants are live on a single standardized data model, heading for 32 by mid-2027.
- Standardization was never an ideal. With 5 people and 82 sites, custom work per plant was arithmetically impossible.
- The standard is treated as living, with a control loop that absorbs each new plant instead of freezing the spec.
- Context is worth as much as the data. Structured data let one expert diagnose a failing installation in half a day instead of days.
- Foundations decide what AI you can run. Clean ERP practice is the gate on anything agentic.
- 1 loaf of bread in 3 worldwide is made with Lesaffre yeast.
Get future Practitioners Unplugged episodes
Standardizing data across manufacturing plants sounds like a governance exercise. At Lesaffre it was closer to a math problem. François Pineau leads Factory 4.0 at the 170-year-old fermentation company, and his central team is 5 people serving 82 plants.
He joins Practitioners Unplugged hosts Dante Vaccaro and Sree Hameed to explain how one data model reached 22 of those plants. Specifically, they get into why it is still not finished, and what breaks when the foundations underneath it are weak.
Why Standardizing Data Across Manufacturing Plants Is Arithmetic, Not Ideology
Lesaffre is a 170-year-old yeast company that now classifies itself as biotech. In total, it runs 82 plants worldwide, 34 of them large fermentation sites. Notably, 1 loaf of bread in 3 globally is made with its yeast.
Yet Pineau’s central digital manufacturing team is 5 people. Consequently, the usual approach was never available to them.
“We are only 5 people on the team, we cannot run 82 custom projects obviously. So it came as something very obvious and unthinkable otherwise, that standardization was not a simple ideal; it really was arithmetic.”
François Pineau, Factory 4.0 Group Leader, Lesaffre
That reframing matters. Specifically, standardization stops being a preference about tidiness and becomes the only arrangement the headcount permits.
Every Plant Is Different, and the Standard Has to Absorb That
In practice, the first obstacle was variation. Every Lesaffre plant makes broadly the same product, yet each differs by location, climate, age, and how it was built.
Therefore the team began by measuring the gap between a harmonic data standard and what was actually running on each shop floor. That meant defining the vocabulary before writing a single template.
Process engineers and domain experts had to define a separator, a filter, a fermenter, a tank. Only then did anyone write a template.
As a result, the model plugs into almost any architecture, whatever the DCS or SCADA underneath. Moreover, mapping is where local knowledge becomes essential, because plant teams know their own equipment and the central team does not.
This is the same semantic gap Jonathan Wise of CESMII described on an earlier episode, approached from inside a single company rather than across an industry.
Context Is Worth as Much as the Data
For example, a new installation at a plant in Poland was failing. In fact, the site reported lost time and lost money on a line that runs continuously.
Consequently, Lesaffre sent an expert from France. Instead of troubleshooting blind, he went straight to the local PI System, built his own dashboard in minutes, and diagnosed the fault in half a day.
Before standardization, that work took days. Furthermore, the team learned something larger from it. The same diagnosis can now happen remotely, without anyone boarding a plane.
“Context is as important as the data itself, because if you don’t understand what you’re reading, you cannot do much about it.”
François Pineau, Factory 4.0 Group Leader, Lesaffre
In other words, the historian stores values and the context layer explains what those values mean. Without the second, the first is just numbers.
The Standard Is Never Finished
Notably, Sree Hameed draws the contrast with a conventional IT project, where you design the to-be process, deploy it, and walk away. Pineau rejects the premise outright.
“If you really are done, it means that nobody’s using your data. We are never done because people are using data.”
François Pineau, Factory 4.0 Group Leader, Lesaffre
Instead, Lesaffre runs the standard as a control loop. Every new plant surfaces new ways of gathering information, and those get absorbed into the model rather than rejected by it.
In fact, the team describes the standard as alive. Additionally, they monitor compliance remotely across sites while feeding the asset framework new components that other plants can reuse.
The first plant went live in February 2024. Currently, 22 are running, with 32 targeted by mid-2027.
Foundations Decide What AI You Can Actually Run
Meanwhile, Lesaffre’s largest AI project improves the fermentation process itself. Thousands of fermentation data sets feed a data lake where models identify which parameters matter and how to influence them.
In practice, the results are measured in energy and water. Given the volumes a fermentation site consumes, a 2 percent improvement is a substantial win.
However, none of it works without clean, identified, contextualized data. Specifically, a model trained at one site deviates unpredictably at another when the underlying data was never comparable.
Importantly, Pineau splits AI into two categories. Enterprise tools are managed globally, while applied AI connects a local language model through to plant systems so maintenance and management get answers from their own data.
“The data that we’re gonna use in 10 years, we are producing now. So we have to be sure that it’s super clean, super usable by whatever is coming in the next 10 years.”
François Pineau, Factory 4.0 Group Leader, Lesaffre
What the Team Looks Like Now
Dante Vaccaro asks what talent profile this work demands. Pineau’s answer inverts the usual job description.
Specifically, he wants people who understand processes, AI, and interoperability without being an expert in any of them. By contrast, deep specialists in one topic are what he is explicitly not hiring.
Indeed, his team of 5 all arrived from different backgrounds. For example, one came from retail and is now the strongest prompt engineer among them.
Overall, the adoption numbers have moved. More than 250 people use the system daily across 22 plants, with 80 use cases identified, 12 in progress, and 3 already scaled to multiple sites.
Frequently Asked Questions
What does standardizing data across manufacturing plants involve?
Standardizing data across manufacturing plants means agreeing a common vocabulary and structure for equipment, processes, and measurements, then mapping each site onto it. Specifically, Lesaffre defined what a separator, filter, fermenter, and tank are before building templates, which lets the model plug into almost any control architecture.
Why standardize instead of building per-plant solutions?
Arithmetic. Lesaffre’s central digital team is 5 people serving 82 plants, so 82 custom projects were never possible. Standardizing data across manufacturing plants was the only arrangement that let a small team serve a large network without multiplying the work site by site.
What is an asset framework in manufacturing data?
An asset framework is the contextual layer sitting on top of raw measurements that explains what each value represents. Notably, Pineau makes a distinction here: his team built their methodological framework inside the platform, rather than letting the platform dictate the structure.
How long does deploying a standardized data model take?
Lesaffre’s first plant went live in February 2024 and 22 are running now, with 32 targeted by mid-2027. Notably, the pace accelerated after roughly 10 sites, once the team had worked through the technical and cybersecurity details that slow early deployments.
What foundations are required before scaling AI in manufacturing?
Clean, identified, contextualized data above all. Pineau notes that a model trained at one site will deviate unpredictably at another if the underlying data was never comparable. Consequently, ERP practice and data governance gate anything agentic, however capable the model is.
How does standardized data speed up troubleshooting?
It removes the discovery phase. For instance, at a Lesaffre plant in Poland, an expert built his own dashboard in minutes and diagnosed a failing installation in half a day, work that previously took days, because the data arrived already labeled and structured.
Is a manufacturing data standard ever finished?
No, and Pineau argues a finished standard is a warning sign rather than an achievement. If nobody asks to extend the model, nobody is using it. Instead, Lesaffre runs a control loop that absorbs each new plant rather than freezing the specification.
What skills do manufacturing digital teams need now?
Breadth over depth. Pineau hires people who understand processes, AI, and interoperability without being an expert in any single one, then relies on the team’s combined backgrounds. For instance, one member came from retail and is now its strongest prompt engineer.
Bottom Line
Standardizing data across manufacturing plants is not really a technology decision. Ultimately, it is a decision about how much variation an organization is willing to carry.
Lesaffre chose to carry less, and a team of 5 now serves 82 sites because of it. Above all, the standard works because nobody has declared it finished.
To submit a request for a new episode topic from Practitioners Unplugged,
visit our contact page.
Explore more about Schneider Electric & AVEVA
Sponsored by Optimum Productions
Author: Wes Garrett
Content and Growth Marketing Producer | From Strategy to Execution, Delivering Impactful Media Solutions and Client Success
