How pulp and paper can successfully implement AI

Posted: August 21, 2026

Some pulp and paper producers were early AI pioneers, using artificial neural networks to optimize pulp machines, digesters and bleaching as early as the 1990s. But—along with the rest of the world—it took until the last few years for the industry to start embracing AI full-scale. What changed? Data quality.

Most of the basic computations that modern AI systems perform have been around for decades. The newly impressive effectiveness of modern industrial AI is due largely to our increased ability to collect, store, and manage huge quantities of high-quality data to feed into those AI systems. If you don’t have high-quality data to give to your AI, you may as well be back in the 1990s.

The need for good input data goes back to the first computer invented by Charles Babbage in the 1800s. The workings of these new machines seemed so magical at the time that people asked its inventor, “Pray, Mr. Babbage, if you put into the machine wrong figures, will the right answers come out?" His reply: “I am not able rightly to apprehend the kind of confusion of ideas that could provoke such a question.”

Even though new AI systems may seem just as magical to us as the first computers did to the Victorians, the same adage applies to both: garbage in, garbage out. Feed your AI poor-quality data, and you’ll get poor quality—even harmful—results.

Why poor data causes AI projects to fail

When businesses give AI poor-quality data, the results can be catastrophically expensive. In 2022, Unity Technologies lost out on over $110 million and its stock fell 37% after it fed bad data into the AI it used to target advertisements. Around the same time, Zillow lost $881 million and had to lay off a quarter of its employees after it used an AI system to buy houses and try to resell them at a profit. One of the main problems was that the AI didn’t have access to good enough data on the houses to accurately predict what buyers would pay for them.

These case studies aren’t anomalies. More than 25% of companies contending with bad data estimate it causes them to lose more than $5 million annually, with 7% reporting they lose $25 million, according to a report by Forrester.

Verdantix reports that some 75% of industrial firms run into significant challenges implementing AI analytics because of the poor quality of their data. Verdantix recommends that before firms invest in AI technology, they first make sure they have a good industrial data management system that integrates data from all of their systems: operational technology (OT), information technology (IT), and engineering technology (ET).

How pulp and paper mills are implementing AI successfully

The news isn’t all doom and gloom, though. IBM notes that “organizations with strong data management can quickly move from pilot projects to deployment.” When pulp and paper companies give AI good-quality data, the results are impressive.

ANDRITZ has unveiled new technology that integrates and organizes data from across engineering, operations, maintenance and production management. AI can use that data to make pulp and paper plants become more fully autonomous—not only detecting anomalies, but estimating the remaining useful life of machinery components, and then recommending which interventions will be most effective in terms of both production and cost.

ANDRITZ VP of Digital Products and Solutions for Pulp & Paper, Daniel Schuck, gives a concrete example of the power of the system: “If I know that the refiner plate has a useful lifetime of 20 days, and my shutdown is in 30 days, I maybe need to adjust the process to lose efficiency, but run to the next shutdown.”

AI-enabled Interventions like these can both reduce costs and increase production significantly. This June, McKinsey started a new approach to reducing costs and improving performance at pulp and paper mills, called “site sprints.” So far, it has notched between 8% to 20% cost savings at individual mills, totaling several hundred million dollars.

AI optimization of manufacturing processes is a key component of those savings. For example, AI optimization of digester performance improved wood fiber yield, which increased output several percentage points while also reducing spending.

In another example, a board mill reduced chemical waste by 18% by using an AI system that adjusts chemicals in real time. Rather than waiting days for laboratory tests, it adjusts chemicals continuously in response to real-time data from sensors monitoring changes in raw materials, temperature, machine performance and other parameters across the entire manufacturing process.

AI needs accurate, complete data

That example shows why accurate, contextualized and complete data is so crucial for successfully deploying AI. Adjusting chemicals based on traditional lab testing relies on just one data point: the test result. What puts the intelligence in artificial intelligence is that AI can treat a paper mill more like a doctor treats a patient: by looking at how all the symptoms fit together before recommending the best course of action.

But, just like a doctor, AI will give you bad advice if it doesn’t have a clear picture of how all those symptoms fit together. If your smartwatch shows your heart is racing every night at bedtime, your doctor might prescribe you a Beta-blocker or some other medicine—even if she knows you drink a cup of coffee every day.

But if you tell her you’re downing your coffee every night at 10pm rather than in the morning, the advice might be very different: forget the pills and lay off the late-night caffeine! Understanding how your smart-watch readings and caffeine consumption fit together are crucial for an accurate diagnosis.

In the same way, if your production plant data isn’t organized and accurate, you’re liable to get unhelpful or even harmful recommendations from an AI. The AI needs to know how the temperature, vibration levels and moisture readings on different machines are sequenced in time just like your doctor needs to know the timing of your heart rate and coffee habits.

The more complete and comprehensive data you have on your operations, the better advice you can get from an AI. The best way for pulp and paper mills to enter the AI era is prepared with a data-management plan that:

1. Comprehensively integrates IT, OT and engineering data
2. Accurately contextualizes the relationships between that data
3. Ensures that data is accurate and trusted

Schuck says: “Something that most people neglect is that data comes from instruments. I have seen maybe just a few projects where people are really checking: if the instruments are reliable; if they have a maintenance plan; if the maintenance people are skilled enough to maintain those instruments. This is the foundation.”

AI can seem so magical, it’s hard for businesses to hear that getting useful results from it might take months of work and investment. But, it’s just not worth investing in AI if you’re not first investing in reliable data collection and data management.

Hannu Ojasalo, Sales Manager at Roima Intelligence, relates how his company helped a forest and paper products company update its real-time data collection at three different plants: “[At] first, the customer really thought, ‘We can do everything with our Azure platform.’ But for actually doing real-time data collection and utilization, they needed something else.”

In the end, though, that investment in data management will help you build an AI system that pays dividends well into the future. The alternative may be a flashy new AI that doesn’t generate revenue.

Schuck says, “I have seen so many projects where there is a lot of investment, the people build a really nice room, they take a photo they put on LinkedIn—six months later, that room is empty.”

A strong foundation in data management is what separates AI that actually increases production and reduces costs from an AI that’s ultimately just for show.