Artificial intelligence sounds like the newest technology arriving at the paper mill.

In reality, the mills most capable of using it have often spent years building everything that comes before it.

Sensors.

Process historians.

Distributed control systems.

Quality-control systems.

Manufacturing execution software.

Reliable material data.

Production models.

And, just as importantly, operators who understand when a recommendation makes sense and when it does not.

That distinction is becoming important as AI moves deeper into pulp, paper, tissue and packaging operations.

The industry certainly has problems that machine learning can help solve.

Fibre quality changes.

Energy prices move.

Demand shifts.

Paper machines suffer disturbances.

Grade changes consume time and material.

Finished-goods inventory ties up working capital.

Transportation decisions interact with mill schedules.

And a pulp mill consists of tightly connected processes where optimizing one department can simply move the bottleneck somewhere else.

But the evidence emerging from actual mills suggests that AI is not replacing conventional mill optimization so much as being layered on top of decades of automation and process knowledge.

That may be the more useful lesson for mills considering where to invest.

The first question should not be:

Which AI platform should we buy?

It should be:

Is our mill data good enough that an algorithm can make a better decision than the people already running the process?

Pulp and paper already had optimization long before generative AI

There is a risk that today’s AI discussion makes industrial optimization sound new.

It is not.

Paper machines have used model-based controls, quality-control systems and advanced process control for decades.

Pulp mills already optimize variables such as:

cooking,

oxygen delignification,

bleaching,

evaporation,

lime kilns,

recovery boilers,

steam balance,

refining,

basis weight,

moisture,

and sheet properties.

Modern AI adds additional tools to that stack.

Machine-learning models can estimate properties that are difficult to measure continuously.

Digital twins can simulate process behaviour.

Anomaly-detection systems can look for patterns preceding equipment problems.

Optimization software can continuously update production targets as constraints change.

But the basic industrial problem remains familiar:

keep the process as close as possible to its economic optimum without compromising quality, reliability or safety.

ABB’s current pulp and paper portfolio makes that progression particularly visible.

Its advanced process-control systems already use process models and machine-learning-based virtual measurements, while its newer Plant Optimizer coordinates material flows, production plans and process constraints across the entire pulp mill.

The technology may be becoming more intelligent.

The objective has not changed very much.

The real opportunity is between process areas, not just inside them

This is where newer optimization systems become interesting.

A pulp mill contains multiple operations that are technically separate but economically interdependent.

Increasing the fibre-line rate may look attractive.

But the additional black liquor still has to move through evaporation and recovery.

More pulp production may increase steam demand.

A lime kiln constraint can restrict chemical recovery.

Tank levels can become limiting.

A planned maintenance event in one department can change the optimal production rate somewhere else.

Optimizing each department independently therefore does not necessarily optimize the mill.

Valmet’s Mill-Wide Optimization system is built around this problem.

It models the relationships between unit operations and updates its representation of the mill using real-time operating data, allowing production targets to change as physical constraints move.

Mercer Rosenthal in Germany is one current industrial example.

The kraft pulp mill has implemented Valmet’s mill-wide optimization as an advisory layer using real-time information from its distributed control system. Production planners establish overall targets and the system distributes coordinated operating targets across departments.

That may be a better representation of where industrial AI is heading than the idea of an autonomous algorithm simply “running the mill.”

The real value is often coordination.

A mill rarely fails because it lacks another dashboard

This also exposes one of the biggest weaknesses in industrial digitalization.

Many plants already have enormous amounts of data.

The problem is that those data live in different places.

DCS databases.

PLC systems.

Laboratory systems.

Maintenance software.

ERP.

Quality-control scanners.

Energy meters.

Operator spreadsheets.

Supplier portals.

Transportation platforms.

A technically advanced forecasting model becomes much less useful if the production system does not know whether a machine is actually available.

Likewise, predictive maintenance becomes unreliable if equipment history is incomplete.

ABB describes the required foundation in three steps: ingest data from operational systems, contextualize those data, and make them available to applications and analytical models through a scalable architecture. Only then do higher-level applications such as AI assistants, predictive analytics and real-time optimization become practical.

That sounds less exciting than artificial intelligence.

It is probably more important.

Billerud shows why automation modernization still comes first

Billerud’s current modernization of Paper Machine 5 at its Gävle mill provides a useful example.

The Swedish producer is upgrading the machine’s distributed control system, quality-control technology and drives.

ABB describes the project partly as creating the platform for future AI-enabled optimization.

But the AI is not the first investment.

The DCS is.

The control hardware is.

The quality measurement system is.

The drives are.

The communications architecture is.

That sequence matters.

A mill running obsolete controls with inconsistent sensors does not become intelligent by attaching a machine-learning model to the top of the architecture.

If an algorithm receives poor measurements faster, it simply produces poor recommendations faster.

Industrial AI therefore tends to reward mills that have already invested in basic automation discipline.

Reliable measurement may be more valuable than sophisticated prediction

Visy Tumut provides another example.

The Australian integrated kraft pulp and paper mill uses online analyzers and automation across cooking, recausticizing, refining and wet-end operations.

Its technical team emphasizes high-frequency, reliable process measurements because those measurements allow a relatively small operating team to make faster decisions and support process optimization.

This is an important point for mills tempted to start with predictive algorithms.

The better question may sometimes be:

What important process variable are we still measuring only every few hours in the laboratory?

Machine-learning soft sensors can be valuable precisely because many important paper properties are difficult to measure continuously.

ABB, for example, uses machine-learning-based soft sensors to estimate paper properties online.

But even these models depend heavily on data conditioning, validation and maintenance. ABB explicitly warns that raw data have little value until they are properly collected and contextualized.

That is a much more realistic view of AI than the idea that a model independently discovers how a paper machine works.

Digital twins show what happens when the data layer is good enough

Smurfit Westrock provides one of the clearer documented examples of digital modelling producing a measurable result in paper manufacturing.

At the Townsend Hook recycled containerboard mill in the UK, the former Smurfit Kappa used digital-twin modelling while redesigning the drying process.

The company reported a 6% reduction in steam-energy consumption and a 5% reduction in total CO₂ emissions following implementation.

This is not principally a supply-chain forecasting example.

It is arguably more useful.

Paper drying is one of the largest energy consumers on a paper machine.

A digital model that helps engineers understand how changes affect steam consumption is dealing with a physical process where the economic value can be measured directly.

Less steam.

Lower energy cost.

Lower emissions.

The environmental and financial result come from the same operational change.

That convergence is one of the strongest reasons digital optimization fits pulp and paper particularly well.

AI does not have to choose between cost and sustainability

Papermaking is unusually resource-intensive.

Fibre.

Steam.

Electricity.

Water.

Chemicals.

Transport.

Improving process efficiency therefore frequently reduces both cost and environmental impact.

Consider off-spec production.

If a machine produces fewer rejected tonnes, the mill avoids consuming fibre, energy, chemicals and machine time on paper it cannot sell at full value.

The same logic applies to:

lower basis-weight variability,

fewer sheet breaks,

more stable moisture,

reduced steam consumption,

lower chemical overdosing,

fewer emergency shipments,

and better production sequencing.

This is why the sustainability business case for industrial optimization does not always need to be separated from the financial one.

ABB’s current Plant Optimizer marketing, for example, estimates potential benefits including higher output, lower inventory, less off-spec production and lower production costs.

Those figures are vendor expectations rather than independently audited results, so they should not be treated as universal performance benchmarks. But the categories of value are consistent with how mills actually make money.

AI creates value when it reduces physical inefficiency.

Not when it creates another dashboard.

Demand forecasting is one area where newer AI methods are clearly moving forward

The downstream side of the supply chain presents a different problem.

A mill can optimize its process perfectly and still manufacture the wrong product.

That is particularly relevant in tissue and consumer products, where changes in retail demand can appear much more quickly than changes in industrial paper markets.

Kimberly-Clark is one of the better-documented current examples.

In a 2026 presentation featuring Kimberly-Clark executive Misty Alexander, enterprise-planning company Board says the consumer-products manufacturer replaced heavily manual forecasting with AI-supported econometric models using more than 8 million external market signals.

The program reportedly achieved forecasts within 2% of actual demand in the application discussed, generated more than $10 million in business value, and associated each percentage-point improvement in forecasting accuracy with approximately $1.3 million of safety-stock reduction.

Those figures are still company- and vendor-presented rather than independent audit results.

They should be reported that way.

But the underlying approach is instructive.

The model does not simply look at yesterday’s shipment and extrapolate tomorrow’s.

It attempts to incorporate external signals that may explain why demand is changing.

That is an important difference.

Kimberly-Clark has actually been pursuing demand sensing for years

The current AI language can make this sound like a completely new capability.

Kimberly-Clark’s work with demand signals goes back much further.

As early as 2013, the company was publicly discussing the use of retail data to move beyond historical shipments.

At the time, Kimberly-Clark reported forecast-error reductions of as much as 35% over a one-week horizon and said improved forecasting allowed it to remove between one and three days of safety stock for certain SKUs.

The technology has changed substantially since then.

The principle has not.

Actual consumer demand contains information that orders from wholesalers or retailers may hide or delay.

The modern AI system expands the number of variables that can be incorporated and the speed at which they can be evaluated.

This is another reminder that many industrial AI programs are not replacing an old process overnight.

They are the next stage of a much longer digitalization path.

Forecast accuracy only creates value when production can respond

This is where the paper industry becomes harder than consumer forecasting alone.

A forecast can change instantly.

A paper machine cannot.

Mills face physical constraints.

Minimum run lengths.

Grade-change time.

Trim optimization.

Coating changes.

Winder patterns.

Storage capacity.

Fibre availability.

Maintenance windows.

Customer specifications.

Transportation capacity.

That means demand sensing and mill scheduling cannot remain independent forever.

If an improved forecast says demand for one grade has fallen, production still needs to determine whether changing the schedule now saves more money than finishing the existing campaign.

An AI model that ignores the cost of a grade change can recommend economically irrational behaviour.

This is why hybrid systems are likely to remain important.

Forecasting models estimate what may happen.

Optimization engines determine what can physically and economically be done about it.

Operators decide whether unusual conditions make either conclusion unsafe or impractical.

Reinforcement learning is interesting—but paper mills should remain skeptical of headline numbers

Reinforcement learning receives increasing attention in manufacturing because it can learn scheduling policies under changing conditions.

Recent peer-reviewed work has demonstrated its potential for dynamic manufacturing scheduling and for responding to machine disruptions, changing orders and competing production objectives.

But there is a large difference between:

a research model,

a simulated factory,

and a running paper machine.

That distinction matters when evaluating claims such as double-digit reductions in trim waste or disruption recovery measured in seconds.

I could not independently verify the paper-industry-specific 11% trim-waste reduction and 47-minute-to-90-second recovery figures presented in the source article.

They should therefore not be published as established paper-mill benchmarks without the underlying study.

This is exactly where trade reporting on AI needs to become more disciplined.

The technology is developing quickly enough without inflating the evidence.

The WestRock case requires even more caution

The source article also describes a 2025 WestRock supply-chain resiliency report supposedly documenting:

an 18% reduction in mill finished-goods inventory,

service above 97%,

and a 12% reduction in transportation cost per tonne through AI freight consolidation.

I could not locate the cited report or independently confirm those figures.

There is also a corporate timeline problem.

Smurfit Kappa and WestRock completed their combination on July 5, 2024, creating Smurfit Westrock. WestRock has since operated inside the combined company rather than as the standalone public company described by the article.

Smurfit Westrock does publish sustainability reporting and has genuine digital-twin examples, including Townsend Hook.

But that is not the same evidence as the claimed 2025 WestRock AI supply-chain case.

For a professional paper-industry audience, I would omit those numbers unless the original report can be produced.

This does not weaken the AI argument

If anything, removing weak statistics makes the real argument stronger.

Current suppliers are openly developing mill-wide optimization systems.

Real mills are implementing them.

Paper manufacturers are modernizing automation specifically to enable more advanced analytics.

Digital twins have already produced measurable energy savings.

Large tissue and consumer-products companies are using external data to improve forecasting.

That is enough to establish that artificial intelligence and advanced analytics are becoming operational tools.

The industry does not need invented precision to make the case.

Operator trust is not a soft issue

One part of the source article is directionally right even if its exact acceptance percentages require better sourcing.

Industrial algorithms fail commercially when experienced operators do not trust them.

That is not necessarily resistance to technology.

Operators know things that may never have entered the historical dataset.

A particular grade behaves badly when humidity rises.

One pump vibrates differently before it fails.

A certain furnish mixture causes drainage problems.

A scheduled truck frequently arrives late despite what the transport database says.

A recommendation may therefore be mathematically rational and operationally wrong.

This is why advisory systems remain important.

At Mercer Rosenthal, for example, Valmet describes its mill-wide optimization as an intelligent advisory system supporting planners and operators rather than eliminating them.

The human role changes.

It does not disappear.

Overrides should become data

There is also a better way to interpret an operator rejecting a model recommendation.

Do not simply record:

operator ignored AI.

Record why.

Maintenance risk.

Quality concern.

Unavailable raw material.

Customer priority.

Unrecorded equipment limitation.

Safety constraint.

Bad sensor.

Incorrect model assumption.

Those overrides can reveal where the digital representation of the mill is incomplete.

This creates a feedback loop.

The algorithm learns from plant behaviour.

The plant’s data architecture learns from operator knowledge.

And management learns which constraints exist only in people’s heads.

That may be one of the most valuable outcomes of industrial AI adoption.

Not replacing tacit knowledge.

Making tacit knowledge visible enough to preserve it.

That matters because pulp and paper is losing experience

Workforce demographics make this especially relevant.

Pulp and paper mills depend heavily on accumulated operating knowledge.

Retiring control-room operators, maintenance specialists and process engineers can take decades of practical experience with them.

ABB explicitly identifies workforce turnover as one of the problems its mill-wide optimization technology is intended to address.

A well-designed digital system can preserve some of that expertise through:

historical operating patterns,

standardized grade recipes,

diagnostic logic,

maintenance histories,

process models,

and documented responses to disturbances.

But again, the technology has to capture the knowledge before the knowledgeable people leave.

AI cannot reconstruct everything after the fact.

Procurement may eventually be where the models become most controversial

Using machine learning to predict maintenance is relatively easy to justify conceptually.

Letting a model influence purchasing decisions is harder.

Recovered paper prices move.

Wood costs change.

Chemical markets fluctuate.

Energy contracts differ.

Freight capacity tightens.

An algorithm may eventually recommend:

buying additional OCC,

delaying a pulp purchase,

switching fibre suppliers,

changing the production mix,

or consolidating transportation.

At that point, AI moves from operational advice into financial authority.

Governance becomes more important.

What data were used?

What assumptions produced the recommendation?

Who can override it?

Was a supplier excluded for quality or compliance reasons?

Can the company reconstruct the decision six months later?

Those questions matter increasingly as purchasing algorithms influence real spend.

Autonomy should therefore expand only after accountability does.

Sustainability reporting adds pressure—but CSRD needs updated wording

The original argument also links supply-chain AI with Scope 3 reporting under the EU Corporate Sustainability Reporting Directive.

There is a valid connection.

Better supplier and logistics data can help companies understand upstream and downstream emissions.

But the regulatory description needs updating.

The EU’s 2026 Omnibus I reforms significantly narrowed the population of companies subject to mandatory CSRD reporting, while the European Commission adopted a substantially simplified set of ESRS in July 2026. The revised standards cut mandatory datapoints by more than 60%.

For companies remaining in scope, climate disclosures can still include Scope 3 emissions from significant categories where climate is material.

So the accurate conclusion is not that every paper supplier now faces ever-expanding CSRD Scope 3 reporting.

It is that the largest companies remaining in scope still have a reason to improve value-chain data, while suppliers outside mandatory reporting may continue receiving sustainability-data requests through customer relationships.

That is another data-infrastructure problem.

Carbon information becomes more useful when attached to operational decisions

This is where AI could eventually make sustainability reporting more valuable.

Consider freight.

A conventional sustainability system may calculate transportation emissions after the shipment has occurred.

A decision system could instead evaluate:

cost,

load utilization,

delivery timing,

distance,

transport mode,

and emissions

before assigning the shipment.

The same principle can apply to fibre sourcing.

A procurement decision can potentially incorporate:

price,

quality,

yield,

transport distance,

availability,

certification,

and carbon information.

The value is not simply producing a better ESG report.

It is moving environmental data upstream into the decision itself.

That is much harder.

It is also where sustainability information begins influencing plant economics instead of merely describing them.

Mills should probably start with boring problems

This may be the most practical lesson.

A mill beginning its AI program does not need an autonomous digital twin of the entire supply chain.

It needs one expensive problem with good data.

For example:

predict sheet breaks,

reduce steam consumption,

identify abnormal drive behaviour,

improve grade-change consistency,

estimate a difficult-to-measure paper property,

optimize winder performance,

or reduce inventory imbalance.

ABB’s current winder optimization service, for instance, uses continuous performance data to adjust acceleration profiles and says installations can improve winder productivity by roughly 3% to 8%.

Those again are vendor-reported potential benefits, not guaranteed results.

But the application is narrow enough to evaluate properly.

Did output improve?

Did quality remain stable?

Did operators use the recommendation?

What was the payback?

That is a much better AI pilot than launching an enterprise chatbot and searching afterward for a problem it can solve.

The second project should be harder than the first

Once one application earns trust, the next step can connect decisions.

A quality predictor can feed process control.

Process-control data can feed production scheduling.

Production plans can inform raw-material purchasing.

Demand forecasts can update production plans.

Logistics optimization can use the resulting finished-goods schedule.

This is how a genuine digital supply chain develops.

Not through one giant implementation.

Through increasingly connected decisions.

The challenge is that integration raises the cost of mistakes.

A bad prediction shown on a dashboard is inconvenient.

A bad prediction automatically changing a machine setpoint is more serious.

A bad prediction placing a multi-million-dollar procurement order is more serious again.

The appropriate level of autonomy therefore depends on consequence.

Paper manufacturing may be unusually well suited to this technology

Despite all these qualifications, pulp and paper has several characteristics that make advanced optimization attractive.

The processes are continuous.

The plants generate large quantities of time-series data.

Energy consumption is high.

Raw materials are variable.

Small quality improvements can create large annual savings.

Downtime is expensive.

Production stages are highly interconnected.

And physical constraints are sufficiently well understood that process knowledge can often be combined with statistical models.

That last point may prove especially important.

Pure black-box AI is not necessarily the most useful approach in industrial manufacturing.

A model that combines physics, process knowledge and operating data can be more robust.

Recent research into continuous manufacturing digital twins is already emphasizing interpretable, physics-informed models rather than optimizing only for prediction accuracy.

A paper machine is not a social-media feed.

The governing physics matter.

The competitive advantage will not be the algorithm

Algorithms become widely available surprisingly quickly.

Cloud computing can be purchased.

AI models can be licensed.

Optimization platforms can be installed.

The difficult assets are harder to copy.

Ten years of clean operating history.

Reliable instrumentation.

Well-maintained automation.

Consistent product codes.

Accurate supplier records.

Integrated logistics data.

Engineers who understand the process.

Operators willing to challenge—and teach—the model.

Those are the capabilities that determine whether AI produces operational value.

This is why the mills beginning digitalization today should not feel that they have already missed the AI transition.

But they should understand the sequence.

Installing AI before fixing the data layer is like installing a new quality-control system on a paper machine whose scanners do not work.

The sophistication at the top cannot compensate for weakness underneath.

The real transformation is from reactive to predictive operations

For most mills, the destination is probably not full autonomy.

At least not soon.

The more realistic transformation is subtler.

Today:

a quality problem happens, then someone investigates it.

Tomorrow:

the model identifies conditions associated with the problem before off-spec tonnes accumulate.

Today:

maintenance responds after vibration exceeds a limit.

Tomorrow:

the system detects a pattern indicating that the equipment is moving away from normal behaviour.

Today:

production discovers a downstream constraint after inventory builds.

Tomorrow:

mill-wide optimization adjusts rates before the bottleneck develops.

Today:

procurement reacts to demand changes after the sales forecast is revised.

Tomorrow:

external signals identify the shift earlier.

That transition—from reacting to conditions toward anticipating them—is where AI can genuinely change pulp and paper operations.

But the mill still has to know what good looks like

Every model ultimately needs an objective.

Maximum tonnes?

Lowest cost?

Minimum energy?

Highest contribution margin?

Lowest inventory?

Best service level?

Lowest carbon?

Maximum equipment life?

These objectives can conflict.

Running harder may increase maintenance risk.

Reducing inventory may increase stockout risk.

Choosing cheaper fibre may increase chemical consumption.

Minimizing energy may constrain production.

A model cannot resolve those trade-offs until management decides how they should be valued.

That is why AI implementation is ultimately not just an IT project.

It exposes operating strategy.

The algorithm cannot optimize a mill whose business priorities have not been defined.

That is the lesson the industry should take from AI

Artificial intelligence will become increasingly normal in pulp and paper.

Some applications will be called AI.

Others will continue to be described as advanced process control, predictive analytics, digital twins or mathematical optimization.

The terminology matters less than the operating result.

The best systems will help mills:

see problems sooner,

coordinate departments better,

use less fibre and energy,

carry less unnecessary inventory,

respond faster to changing demand,

and preserve knowledge as experienced workers retire.

But none of those benefits begin with the algorithm.

They begin with trustworthy measurements and a process that is understood well enough to model.

Smurfit Westrock’s documented digital-twin energy savings came from applying modelling to a well-defined drying problem.

Mercer Rosenthal’s mill-wide optimization sits on top of an existing DCS and years of process automation.

Billerud is modernizing basic controls before moving toward AI-supported optimization.

Kimberly-Clark’s forecasting program depends on millions of external data signals and an enterprise planning environment capable of using them.

Those examples point in the same direction.

AI is not the foundation of the intelligent mill.

It is what becomes possible after the foundation is good enough.

For pulp and paper companies deciding where to start, that may be the most valuable insight of all.

Last Updated on 16/09/2026 by Jota Machinery