When Production Becomes Intelligent - Who Really Makes the Decision?

PRODUCTION INTELLIGENCE • DECISION GOVERNANCE • PRINTING & PACKAGING

When the Press Succeeds but the Plant Fails

Imagine a modern printing plant. Its presses are stable, color is within tolerance, automated inspection catches defects in real time, and equipment dashboards display excellent performance. The production manager sees a reassuring line of green indicators. On the screen, everything appears to be working as intended.

But the laminator is waiting because the adhesive has not completed the required bond-development window. A material reel appears available in the warehouse even though its conditioning period has not ended. Finishing is accumulating work that it cannot process on time, and the customer's committed receipt event is now at risk.

The press has succeeded. The plant has not.

Industrial printing has spent decades learning how to measure and improve individual machines, departments, and processes. Yet a successful production operation is not the same as a successful customer order. A printed sheet is not a finished carton; a printed web is not a released pouch; a pallet at the shipping door is not necessarily a completed promise.

Artificial intelligence offers an important opportunity. It can combine vast amounts of data, detect weak signals, identify relationships across departments, and propose decisions that no single person could assemble as quickly. But a more difficult question follows: if a system can recommend an action convincingly, when should it be permitted to perform that action?

The answer is not determined by computational capability alone. It depends on the quality of the evidence, the boundaries of qualification, the consequences of error, and who remains accountable for the outcome.

1. The Plant Already Produces Data. Why Is It Not Yet Intelligent?

A modern printing plant generates information continuously. Enterprise resource planning (ERP) manages orders, purchasing, inventory, and commercial commitments. Manufacturing execution systems (MES) record production progress. Prepress and digital front ends (DFEs) retain artwork versions, separations, screening, imposition, and output instructions. Press and supervisory systems collect speed, temperatures, interruptions, calibration, and equipment states. Inspection systems retain defect classes and images; laboratories record color, adhesion, dimensional, bond, and other test results.

Each system knows something. None necessarily knows the whole job.

One material may be identified by a purchasing code, a supplier designation, a warehouse reel number, an abbreviated press-entry name, and a laboratory trial number. The records may refer to the same physical lot without a reliable digital connection. During a complaint investigation, an experienced person reconstructs the relationship through calls, samples, timestamps, and professional judgment.

This is the first limitation an AI program must confront. More data is not useful if the identity, provenance, conditions, and applicability of that data cannot be established. Connecting databases is an engineering requirement; interpreting whether their records describe the same valid production state is a separate task.

For industrial use, every meaningful event should retain its job and product identity, material lot, production route, process version, time, measurement method, operating conditions, and source. The broader enterprise-to-manufacturing integration problem is the subject of the ISA-95/IEC 62264 standards family [3]. An intelligent layer should use such integration to preserve context, not to turn ambiguous records into confident conclusions.

Data integration tells us what different systems recorded. Industrial intelligence must help us determine what those records mean together—and what they cannot establish.

1. Automation, Intelligence, and Autonomy Are Not the Same Thing

Conventional automation performs an action according to a defined rule. An AI-supported system may recognize patterns, estimate risk, compare alternatives, or propose a next step based on evidence and learned relationships. An autonomous function can execute a class of decisions without requiring an individual approval for every instance, but only within the authority deliberately granted to it.

These are different capabilities, not automatic stages of maturity.

A sophisticated AI application might assemble a complex complaint file and propose discriminating tests without being permitted to adjust a press or release a product. Conversely, a rule-based function may safely move an unreleased job between two previously qualified equivalent resources when all current conditions match the approved operating envelope.

The objective is not the highest degree of autonomy. It is the appropriate degree of delegated authority for each decision. Some actions are safe to automate because the engineering and commercial decision has already been defined within an approved rule. Others create a new decision every time they are considered.

A practical rule follows: no production parameter, material qualification, customer promise, regulatory claim, or product release should change merely because an algorithm assigns a high confidence score. Authority must be based on the decision's purpose, consequences, evidence, and approved boundaries.

1. Intelligent Scheduling Must Optimize the Route, Not the Next Available Press

Consider a scheduling system that finds a press that can start a job earlier. Its assessment of press capacity may be correct. But does the complete route exist? Are the die-cutting tools available? Is the correct film conditioned? Can lamination meet its cure requirement before slitting? Is the inspection method qualified for the artwork and construction? Can quality release occur before the customer's actual receiving cutoff?

A press-centered optimizer can accelerate the first operation and simply move waiting time downstream. That increases work in progress, creates urgent resequencing, and may worsen delivery reliability while reporting higher local utilization.

The intelligent planning question is therefore not, 'Which job should be printed next?' It is, 'Which permitted route best protects the customer's defined commitment across all required operations?'

Overall equipment effectiveness (OEE) remains useful for describing asset performance within defined conditions. But it does not, by itself, measure whether a customer order progressed through preparation, production, finishing, inspection, release, and delivery without interruption. Route-level effectiveness and governance must preserve readiness, flow integrity, and the validity of the final release decision.

Scheduling intelligence should distinguish physical movement from authorized progression. Material moved to the pressroom is not automatically ready. A printed reel is not automatically convertible. Finished goods staged at the dock are not automatically approved for shipment. A good model cannot optimize away these distinctions without changing what success means.

1. Root-Cause Investigation: From Symptom to Mechanism

Complaint investigation is a particularly promising application for AI because the relevant evidence is distributed across nearly every department.

Suppose a customer reports 'ink cracking' on folding cartons. Photographs show white lines along the closure creases. The complaint label may accurately describe the appearance, but it does not yet identify the failure mechanism. The damaged region may lie within the ink layer, the board coating, the fiber network, or at another interface. Crease geometry, board moisture, conditioning, folding temperature, and customer erection conditions may matter more than ink adhesion.

A useful AI system can collect press conditions, board identity, moisture records, coating details, crease tooling, finishing history, retained samples, and similar investigations. It can identify missing evidence and recommend tests designed to distinguish competing explanations. Microscopy through the failure, controlled folding under relevant conditions, and comparison with qualified references may be more informative than repeating a tape test away from the crease.

The investigation must preserve separate states: observation, hypothesis, discriminating test, supported mechanism, corrective action, and authorized closure. A customer term such as 'delamination' or 'migration' should not become a verified mechanism just because it appears in an email subject line.

An AI model may discover correlations that people would otherwise miss. It may also propose a causal structure. But correlation is not causation, and a plausible causal statement is not a confirmed root cause until evidence supports it. Controlled physical testing, expert interpretation, and explicit ownership remain essential.

The goal is not to make the machine declare a root cause faster. It is to shorten the time required to reach the first meaningful test, improve evidence completeness, and reduce repeated failures caused by premature conclusions.

1. Predictive Maintenance: Finding a Signal Is Not Diagnosing a Failure

Predictive-maintenance tools can examine vibration, thermal, current, acoustic, and operating-history data to identify patterns that precede mechanical problems. This can reveal deterioration earlier than a conventional alarm or periodic maintenance schedule.

A change in a bearing's vibration signature, however, might reflect damage, a different load, a speed change, or an altered operating condition. Detection is not diagnosis. The model can recommend targeted examination and provide a likely risk window, but the physical mechanism still requires appropriate verification.

The intervention decision is broader than the anomaly score. Repairing too early may waste production capacity and usable component life; intervening too late can damage the asset or introduce quality and safety exposure. The decision must consider qualified maintenance practice, spare-part readiness, scheduled orders, downstream commitments, and safe access to equipment.

Predictive advice must not weaken independent machine protections, safety interlocks, or established operational controls. Good intelligence strengthens preventive action without confusing its recommendation layer with the equipment's safety authority.

1. Materials Intelligence: Similar Is Not Equivalent

A model may identify a less expensive substrate with a nominal thickness, grammage, or tensile range close to the incumbent material. That can be a valuable procurement opportunity. It is not proof of interchangeability.

In printing and packaging, qualification belongs to a specific material state, construction, route, and end-use condition. Surface treatment, receptive coatings, primer, wetting, dimensional response, heat resistance, ink anchorage, adhesive compatibility, bond development, friction, sealing, and distribution exposure can all change the outcome.

A candidate flexible-packaging film must be evaluated with the actual ink, adhesive, curing, converting, and customer-use sequence. A candidate folding cartonboard may match purchased thickness and weight while responding differently to creasing, folding, protective coating, storage, or customer erection. Food-contact structures may require additional migration and regulatory assessment appropriate to the final construction.

AI can identify candidates, organize historical evidence, recommend a qualification plan, and recognize promising results. It cannot turn a successful trial into unrestricted approval. Performance proven for one structure, process window, storage interval, and customer condition remains bounded by those conditions.

Only after equivalence is established for a defined scope may a system select qualified stock automatically under a preapproved rule. Before that, the recommendation is an invitation to test, not authority to substitute.

1. Color Intelligence: A Correct Measurement Can Still Lead to a Wrong Decision

Color management offers structured data that can support useful machine-learning applications: calibration history, spectral measurements, spot-color recipes, process conditions, substrate white, screening, and finishing results. A system may identify drift earlier, compare qualified correction paths, and anticipate when a change is likely to cause instability.

But the same measured result can look different when the substrate, gloss, viewing conditions, varnish, or lamination changes. A recipe that succeeded for one press and finished construction is not a universal definition of the brand color.

The master color target must remain distinct from the route-specific recipe used to achieve it. Likewise, an accepted historical sample shows a finished appearance under particular conditions, not necessarily a general instruction for a different material or print technology.

A trustworthy model should show why a previous recipe is considered relevant, which differences may invalidate that comparison, and what measurements or proofing remain necessary before production use.

1. Supply-Chain Intelligence: A Confirmation Is Not a Reliable Commitment

Procurement records may show that a supplier confirmed every line of a purchase order. But a confirmation may arrive before capacity has been allocated, before a production slot is secured, or before transport exists. The transaction is real. Its operational meaning may still be uncertain.

An intelligent supply-chain layer can compare promised quantities and dates with supplier behavior, capacity evidence, loading events, transit dependencies, warehouse readiness, and the sequence in which production needs the material. It can expose which customer commitment would fail if a particular load arrives late, and which recovery options remain both feasible and authorized.

The same problem appears at outbound delivery. 'Delivered' can mean departed the plant, reached the customer gate, entered inventory, passed inspection, or became available for use. These are different events, and the contractual meaning matters. AI cannot improve delivery reliability when the organization has not defined which event it has promised.

Effective supply-chain intelligence connects evidence to customer exposure. It does not confuse a commercially complete record with an operationally protected obligation.

1. The Silent Failure: Learning from Protected Success

One of the most consequential risks in industrial AI is unsupported generalization from unusually controlled conditions.

Imagine a new material that performs well during ten trial orders. Each trial had extra inspections, controlled conditioning, additional engineering attention, and more generous waiting time. If the learning system records only 'material succeeded,' it may conclude that the candidate is inherently superior and recommend removing the buffer or extending approval to untested families.

The missing variable is the work that made success possible.

This problem can occur across procurement, materials, maintenance, and quality. Equipment that was repaired and then monitored unusually closely may appear more reliable than equivalent equipment receiving routine attention. A pilot whose operators record every interruption may look less stable than a production line where minor events are not recorded at all. Cleaner digital documentation can bias supplier rankings even when physical performance is comparable.

Training and evaluation therefore need the full evidence envelope: material and route identity, qualification limits, trial versus routine status, temporary controls, monitoring intensity, intervention history, and unresolved post-use evidence. Eligibility and safety constraints must remain outside a purely statistical ranking when the model cannot legitimately relax them.

A successful controlled exception may justify another experiment. It does not, by itself, authorize a general operating rule.

1. The Interface Can Create Automation Bias Before Any Action Is Executed

It is tempting to consider an AI system safe when a person still gives final approval. That assumption can fail long before an approval is clicked.

A highly ranked option occupies the first position on the screen. The explanation looks complete. The recommendation is repeated across several shifts. Managers begin referring to it as 'the plan,' and dissent requires more effort than acceptance. The model has influenced the decision before receiving execution authority.

This is a practical manifestation of automation bias: people place too much weight on a machine-generated answer, especially when it arrives quickly and appears objective. A green status, confidence score, or persuasive narrative can conceal a missing qualification boundary.

A human approver matters only when that person can access the underlying evidence, identify uncertainties, understand alternatives, has sufficient time, has the necessary competence, and can genuinely refuse or override the proposal. Approval without these conditions is little more than a procedural signature.

Interfaces should distinguish findings from recommendations, routine executable options from learning proposals, and current evidence from expired assumptions. A proposed causal explanation should not share the same visual treatment as an approved root cause. The design itself is part of the governance system.

1. A Practical Permission Architecture

Instead of asking whether a plant should 'use autonomous AI,' define the authority appropriate to each individual decision. The following matrix is an engineering proposal, not a mandatory classification specified by any standard.

This approach is consistent with risk-based organizational governance, including the functions described by the NIST AI Risk Management Framework and the management-system requirements of ISO/IEC 42001:2023 [1, 2]. These publications do not prescribe the specific printing examples above; those are proposed applications of general governance principles.

Authority is not merely a user permission. It is the relationship between a decision, the evidence supporting it, its current operating context, the applicable limits, and the owner responsible for the result. The system must also know when an action's enabling condition has expired.

Not every loop needs an individual human decision. Every loop needs defined ownership.

1. When the Intelligent Layer Goes Silent

A plant that depends on AI-supported decisions must remain capable of safe operation when the intelligence layer is unavailable, degraded, or unable to verify current authority.

The failure might involve network communications, application software, identity services, inconsistent records, or a cybersecurity incident. Not every interruption requires a full plant shutdown. But a system that cannot prove that permissions, identities, and qualification states remain current should not continue making new automated decisions as though nothing changed.

Degraded-mode planning should preserve the minimum truth required for safe continuity: active customer commitments, material and route qualifications, protected holds, critical time windows, named escalation owners, and the authority to restart or suspend affected decisions. Human teams must rehearse how to use that information without relying on the missing dashboard.

Safety systems should remain independent of the AI recommendation layer, with clearly established recovery criteria. NIST SP 800-82 Rev. 3 provides publicly available guidance on operational-technology security and its distinctive reliability, availability, and safety considerations [4].

Resilience is not simply how quickly a server comes back online. It is whether the organization can still make defensible decisions while its most convenient decision-support system is unavailable.

1. What Should an AI Investment Actually Measure?

Accuracy, response time, dashboard adoption, and the number of accepted recommendations can be useful indicators. None is sufficient proof of operational value.

A planning system that raises press utilization while increasing finishing queues has not necessarily improved the business. A procurement model that lowers purchase price while raising qualification cost or customer complaints may be destroying value. A complaint assistant that accelerates closure by overstating causal certainty creates a different risk from the one it was intended to reduce.

Success should be evaluated at the level of the complete route and the customer commitment. For complaints, relevant measures include time to evidence assembly, time to the first discriminating test, the proportion of proposed mechanisms revised or rejected after investigation, and recurrence of confirmed failures. For planning, useful measures include schedule stability across the route, incomplete work awaiting finishing, emergency resequencing, and reliability at the customer-defined event. For maintenance, track useful warnings, false alarms, avoided breakdowns, intervention costs, and safety outcomes.

AI governance itself must be measured. How often are recommendations reversed after new evidence appears? How often is a constraint found to be missing? Can the system retract a conclusion? Does it preserve the distinction between trial success and routine performance? Are operators' challenges changing the evidence or merely becoming a new metric for management?

Finally, include the full cost of ownership: integration, data cleaning, validation, model maintenance, controlled dual-running, workforce training, cybersecurity, continuity practice, and preserving human competence. A business case based only on license cost and theoretical productivity gain is incomplete.

1. The Human Question No Algorithm Can Resolve Alone

Introducing AI changes the relationship between knowledge, experience, and authority. Experienced operators may notice shifts in sound or material behavior before a recorded alarm exists. Specialists may recognize that a test passed under the wrong conditions. Planners may know that two dates that look equivalent on a screen protect very different customer obligations.

This knowledge should not remain forever isolated in individual memory, but capturing it should not turn every human observation into a disciplinary metric. A system gains better evidence when employees can challenge it safely and see their corrections improve the operating record.

People also need to maintain their ability to work without the model. If routine judgments are always outsourced to a dashboard, practical competence can erode just when an exceptional condition requires it most. Training must therefore cover not only how to follow the system, but when and how to question it.

There is no responsible reason to promise that automation will never change jobs. Some repetitive tasks may shrink; other roles may require new capabilities; not everyone will benefit in the same way. Management owes people accurate information, participation, time to develop competence, and honest treatment of consequences—not a comforting story that ignores the changes.

A genuinely intelligent organization uses technology to distribute trustworthy evidence and meaningful authority. It does not quietly remove human judgment while still advertising human accountability.

1. A Practical Starting Point: Build Trust Before Delegating Action

A printing plant does not need an enterprise-wide autonomous platform on day one. Start with a bounded, high-value problem in which the evidence and consequences can be inspected.

Complaint intelligence is often a useful first application. Begin by assembling records, identifying missing links, preserving source identity, and proposing tests. Do not permit the application to change a press parameter, approve material, issue a compliance statement, or close a customer complaint on its own.

Validate the system against cases where the mechanism is known, where the initial hypothesis was wrong, and where essential physical evidence was missing. Test whether it distinguishes customer observations from verified mechanisms. Review whether it maintains context across material lots, production routes, coating states, and final-use conditions.

Only after the application demonstrates reliable evidence handling should its authority expand, and then only for a clearly described action class. Define the owner, permitted scope, validation evidence, override pathway, monitoring requirements, and rollback conditions before activation.

The aim is not simply to generate more recommendations. It is to create a production organization capable of learning from supported decisions, rejected hypotheses, revised rules, and outcomes that arrive months after the original action.

Conclusion — The Future Is Not a Plant Without People

Artificial intelligence can make industrial printing and packaging more consistent, better connected, and faster to learn. It can bring together data previously isolated in different departments, identify emerging failures, improve scheduling, reveal promising materials, and help specialists reach meaningful tests sooner.

But it can also propagate a weak assumption with extraordinary speed. A model trained on incomplete context may generalize beyond the evidence. An optimizer focused on a local metric may transfer waste to another department. And an organization that grants authority before defining its boundaries may convert a persuasive recommendation into a decision nobody can properly explain or own.

The decisive question, then, is not how many decisions AI can make instead of people. It is how to design a production system in which evidence is traceable, competence remains active, authority is explicit, and each action is connected to its real consequence.

The intelligent plant of the future is not the one from which people have disappeared. It is the one in which people make better decisions, intervene where judgment matters, and confidently delegate only what has been demonstrated safe to delegate.

The quality of intelligence is measured not only by what the system knows, but also by whether the organization understands what the system does not yet know.

Professional References

[1] National Institute of Standards and Technology (NIST). Artificial Intelligence Risk Management Framework (AI RMF 1.0), 2023.

[2] International Organization for Standardization / International Electrotechnical Commission. ISO/IEC 42001:2023 — Information technology — Artificial intelligence — Management system.

[3] International Society of Automation (ISA). ISA-95 Series of Standards — Enterprise-Control System Integration; related IEC 62264 standards family.

[4] National Institute of Standards and Technology (NIST). NIST SP 800-82 Rev. 3 — Guide to Operational Technology (OT) Security, 2023.

About the Book

The themes explored in this article are developed further in Ronen Dolev's book When Production Becomes Intelligent: The Industrial Reality of AI, Automation and Decision-Making in Printing (2026). Through a human-centered industrial novel set in a fictional printing and packaging company, the book explores intelligent scheduling, predictive maintenance, materials, color, complaint investigation, supply-chain risk, automation bias, and the responsibility that remains with people when systems grow more capable.

The book is available to purchase on Amazon. Visit the Books section at ronendolev.com and select When Production Becomes Intelligent to follow the Amazon purchase link: .

Editorial note: The industrial scenarios in this article are illustrative rather than reports of verified customer cases. Standards and guidance are cited for their relevant general principles; specific AI authority levels and printing applications are the author's proposed engineering framework. The book's characters and plant are fictional.