What an AI slowdown looks like at Unilever's Indaiatuba plant

Wall Street is sorting AI winners from losers. The world's largest detergent powder plant put 20% more capacity and €3m of savings in writing without buying a model.

Unilever's laundry powder plant in Indaiatuba, Brazil, added 20% of capacity in 2024 and saved almost €3 million in the same year, according to the company's own account published in April 2025. It did that with a digital twin, a handful of machine-learning models running on the plant's own process data, and a training programme for every person on site. None of it depends on a frontier model, a chip supplier or a startup's valuation, which is worth remembering this week.

The Information reports that Wall Street has begun sorting AI's winners from its losers as spending on models and data centres slows. That is a story about technology companies. If you run a plant, the useful question is what an AI slowdown changes on your floor. Indaiatuba is the clearest documented answer we have found: very little, because the systems that pay for themselves in a factory were never the ones the slowdown is about.

What Unilever actually deployed

The plant is the largest laundry detergent powder factory in the world. Its problem was not a lack of ambition but a process that was expensive to change. Making powder is chemically fussy, so every new formulation used to need physical trials before it could run at scale. In its January 2023 account, Unilever describes a digital twin that uses machine learning to predict the optimal process parameters for a new formulation, which removed the need for those trials and let the site launch products such as its first anti-residue detergent faster.

The second system went after energy. The old powder process was the plant's biggest source of greenhouse gas emissions and accounted for 80% of its energy use. Unilever switched the process to biomass and put a machine-learning system on the thermal efficiency of the operation. The company says scope 1 emissions fell by 96% and energy consumption and cost by more than half.

Three smaller models sit around those two. One predicts the right moment to carry out equipment maintenance, and Unilever says maintenance costs have almost halved since 2018. One targets right-first-time sealing of packs to cut waste. One replaces a daily human allocation exercise for outbound orders that involved 600 decisions and 13 critical variables with a model that picks the route and allocation from live data.

Not one of these is a chatbot. Every one of them is a model trained on the plant's own numbers, pointed at a single expensive process.

What did it cost, and how long did it take?

Unilever does not publish a capital figure for Indaiatuba, and we will not invent one. What it does publish is the shape of the effort, and the shape is the lesson. The site's improvement programme is built on the World Class Manufacturing framework the company adopted in 2013, which was then reworked into the Unilever Manufacturing System, a digitally enabled version now live in 124 factories across 2,100 lines. Across that network the company reports an average 3% gain in OEE, 5% in labour productivity and 8% in cost savings. Indaiatuba is the outlier at the top: the highest OEE in the group for two consecutive years.

The people side is where the money went. Unilever says the entire Indaiatuba workforce was upskilled through a digital training programme, along with more than 70 employees from seven other factories in the region, and that the site keeps a network of more than 35 partners, including start-ups, universities and suppliers, for fast prototyping. Across the whole manufacturing system, more than 23,000 factory colleagues have been trained. Renato Miatello, Chief Product Supply Officer for Home Care, calls that training essential to the roll-out.

So the honest timeline is not a quarter. It is a decade of operational discipline, with the machine-learning layer added on top of processes that were already measured, and the first headline energy results arriving by early 2023.

Does the AI slowdown reach this?

No, and it is important to see why. The spending that Wall Street is now questioning is on large general-purpose models and the infrastructure to train and serve them. The Indaiatuba systems are narrow models on proprietary process data. Their cost is dominated by sensors, data plumbing and people, not by compute. If model prices rise or a vendor disappears, a digital twin of a spray tower keeps running.

The larger risk on a factory floor runs the other way: the generative projects that were sold on the boom. Even BMW, which has used camera-based AI inspection in series production for years, describes its generative system for tailoring quality checks at Plant Regensburg, a line producing about 1,400 vehicles a day, as a pilot project in its April 2025 release. When the best-resourced plants in the world call something a pilot, a mid-sized manufacturer should read the label.

Does this apply to a 200-person plant?

Not at Unilever's scale, and not by copying the list. A single-site manufacturer does not have 600 daily allocation decisions to automate or a global manufacturing system to plug into. What transfers is the selection rule. Indaiatuba chose the one process that ate 80% of its energy and the one step that blocked every product launch, and put a model on each. Bosch describes the same pattern across its own plants in a December 2023 release: an energy-scheduling model at its Changsha site cutting electricity use by 18%, a solder-joint inspection model at Ansbach checking 5,000 to 8,000 joints per board, a noise-analysis model at Mexicali tested on around 300,000 tools. One process, its own data, one model.

For a smaller plant the equivalent is usually a furnace, a dryer, a moulding line or the pack seal that generates the most rework. The data for a first model is typically already sitting in the PLC historian and the maintenance log. The expensive part, as at Indaiatuba, is the discipline to measure the process properly first and the training that lets the line team trust and correct what the model says.

Verdict for an operator

Ignore the winners-and-losers coverage unless you own the shares. If your plant has a process with a written energy bill, scrap rate or downtime number attached, that number is where a model earns its keep, slowdown or no slowdown. If a vendor's pitch depends on a general-purpose model and cannot name the single process it will improve, treat it the way BMW treats its own generative work, as a pilot, and budget it that way.

A correction policy applies to every figure above. If Unilever, BMW or Bosch amends any of the numbers cited, we will correct this piece at the top, dated.