VitrA spent about €1m on AI for one tile line. Scrap fell by half

VitrA Tiles put machine learning on one production line at its Bozüyük plant for about €1m. Scrap fell 56% and the company built the models itself.

VitrA Tiles, the ceramics business of Türkiye's Eczacıbaşı Group, spent nearly €1 million putting machine-learning process control on one production line at its Bozüyük plant, and reports that scrap fell 56%, equipment effectiveness rose 19% and energy use fell 14% in 2023. The models set milling, spray-drying and sludge-preparation parameters from the plant's own historical data, and the company says it built them with its own staff.

It is a useful case right now because most of what a plant manager hears about AI this month is about the next frontier model, and whether the labs will slow down. VitrA's results came from none of that. They came from getting process data out of the machines and letting a model pick settings that used to depend on the experience of whoever was on shift.

What VitrA actually ran

Bozüyük is not a small site. It employs about 1,500 people, can make 31.2 million square metres of tile a year and carries more than 4,200 SKUs, according to Eczacıbaşı and the World Economic Forum, which added the plant to its Global Lighthouse Network in December 2023. Tile making burns a lot of natural gas, mostly in drying and firing, and Turkish energy prices have been volatile. That, VitrA's chief executive Hasan Pehlivan said, is what the programme was for.

The programme is called DigiTile. Its base layer is a manufacturing execution system and an industrial IoT network that collect production data and move it to the cloud. On top of that sit several narrow models, each tied to one process step. DigiMill uses data analysis and predictive models to cut the energy used in grinding raw materials. DigiSpray keeps spray-dryer conditions at optimum values. DigiOK sets how much sludge to add in sludge preparation. DigiQuality supports quality-control machines with AI algorithms, DigiAlarm flags problems early, and DigiBoard is a production control tower.

None of these is a chatbot or a general-purpose model. The company describes the core of it as AI that analyses historical data to find the best working parameters for the line.

What did it cost, and how far did it reach?

Pehlivan put the investment at nearly €1 million for hardware and analytics capability. In October 2024 the company also said the digitalisation investment was running on a single production line, Production 3, and Pehlivan tied the gains to the digitally transformed line.

That scope matters when reading the numbers. The 56% scrap cut is not a plant-wide result for 31 million square metres. It is what one line did once models were choosing its settings. In its 2024 sustainability report the company says wider rollout of the DigiTile projects raised its digitalisation rate from 13% to 26%, with more expected in 2025. The report does not define how that rate is measured, so we read it only as the company's own sign that the rollout is far from finished.

None of the savings needed a bigger model. They needed every machine on the line to record what it was doing.

What changed on the plant floor

The 2024 report gives the savings in physical units, which is how an energy manager would want them. DigiOK optimised sludge additions so the plant no longer had to add sludge a second time, saving 12,260 cubic metres of water and 898,294 cubic metres of natural gas in 2024. DigiSpray saves 10,230 cubic metres of water and 749,555 cubic metres of natural gas a year. Taken together, by our arithmetic, those two models account for about 1.65 million cubic metres of gas a year.

The report also says that equipment, software and AI modelling for these projects were carried out entirely by Eczacıbaşı's own people, and that in 2024 the business added Digital Transformation Manager and Engineer roles. It ran 296 digital transformation training sessions that year, and 24 employees completed a further 84 hours of specialised training. The stated aim is to capture the expertise of experienced operators in models so results depend less on who is on shift.

Does this apply to a 200-person plant?

Partly. The method transfers: pick the one process step that burns the most energy or makes the most scrap, capture its settings and outcomes over time, and let a model recommend settings. The budget does not transfer directly. VitrA spent about €1 million on one line of a 1,500-person plant, and it already had an IT function able to build models in-house.

A smaller plant is likely to find the data layer is most of the work. If the line's settings and its quality results are not already recorded together with timestamps, no model can learn from them, and that is the first thing to fix. VitrA's programme is built the same way, with the MES and IoT network underneath the models.

What it does not mean

It does not mean a tile plant is using frontier AI. Pehlivan has said wider use of generative AI in product design is a goal. It is not one of the reported results, and nothing here depends on it.

It does not mean every figure is independently audited. The sustainability report carries limited assurance from RSM Turkey on selected indicators. The press-release figures for scrap and OEE are the company's own, repeated by the World Economic Forum.

And it does not mean the lighthouse label is what produced the savings. The award followed the results. It did not cause them.

What to do next

For an operations manager, the useful move is a single-step audit rather than an AI strategy. Name the process step where one wrong setting costs the most in gas, water or scrap. Check whether its settings and results for the last year are recorded in one place. If they are, that step is a candidate for the kind of model VitrA runs. If they are not, spend the first budget on recording them, and name one person inside the plant who will own the result. Report the saving in cubic metres and tonnes, as VitrA does, so nobody has to take a percentage on trust.