
Industrial AI is making enterprise knowledge more accessible and engineering workflows increasingly autonomous. But better information and faster execution do not necessarily produce better decisions. The opportunity lies in designing the right decision loops between people, AI and enterprise capabilities, with clear accountability, proportionate autonomy and continuous learning.
Better Information Is Not the Same as Better Decisions
Industrial organisations have spent decades investing in capabilities to capture, structure and govern information across the product lifecycle. PLM encompasses an interconnected ecosystem of people, processes and technologies, spanning engineering, manufacturing, supply chain, operations and service. Yet connecting these capabilities does not necessarily improve how decisions are made across them.
Digital threads, knowledge graphs, and AI are improving how organisations connect these capabilities, surface relationships, and access information. Increasingly, the ambition extends beyond retrieving data to preserving decision context: what was decided, why, under which assumptions and with what consequences. This is valuable progress, but it exposes a fundamental limitation. An organisation can retain extensive knowledge without becoming more effective at recognising changing circumstances, challenging assumptions or taking corrective action.
Better information creates the conditions for better decisions, but does not guarantee them. The value of industrial AI depends not simply on information accessibility, but on how effectively that information influences decisions, actions and outcomes.
Context Is Not Static
The growing interest in product memory highlights a genuine weakness in enterprise information management. Engineering decisions rarely exist within a single application. Their rationale may be distributed across PDM records, ERP transactions, manufacturing data, technical reports, emails and individual expertise. Preserving this context and maintaining its relationship to authoritative information is important, particularly when decisions must be revisited years later.
However, context should not become an ever-expanding historical record. Its relevance changes as products evolve, operating conditions shift, and new evidence emerges. A supplier qualification may require reassessment following a manufacturing change, while a validated design assumption may become questionable after observing field performance.
The challenge is not simply preserving context, but maintaining its relevance, reliability and usefulness for subsequent decisions. This requires distinguishing historical evidence that must remain traceable, assumptions that require reassessment, and decision criteria that should evolve through experience. Materiality, risk and lifecycle consequences should determine what context is retained, which dependencies matter and when reassessment is justified.
Product memory provides a foundation, but it does not determine how decisions should be made. That responsibility belongs to the operating model.
Designing the Right Decision Loops
As AI progresses from assisting engineers to executing complex workflows, the familiar concept of human-in-the-loop deserves closer examination. Human involvement is often reduced to reviewing AI-generated results, approving recommendations or checking predefined criteria. These controls remain important, but they do not necessarily constitute an effective decision-making model.
The greater challenge is designing the right loops between people, AI and enterprise capabilities. Four questions should shape their design:
- Decision authority: Who or what can decide, and within which boundaries?
- Autonomy and verification: Which activities can AI execute independently, and when is technical verification required?
- Escalation: What evidence or level of risk warrants human intervention?
- Feedback: How should outcomes influence subsequent decisions and the operating model?
Consider a manufacturing process where AI identifies a deterioration in product quality. The same underlying evidence could justify very different responses depending on its significance and consequences.
A minor process deviation might warrant an automated adjustment within approved operating limits. A recurring defect could require investigation by manufacturing and quality engineers. Evidence of a design-related failure might trigger formal engineering change management, involving additional stakeholders and potentially affecting products already in service.
The information may be connected, but the appropriate decision loop is not predetermined by that connection. It depends on risk, decision authority, technical evidence and the consequences of acting—or failing to act.
The objective is neither maximum AI autonomy nor human approval at every stage. It is proportionate decision governance, with clear accountability and appropriate intervention.
This is an operating-model challenge, not simply a technology architecture problem.
Verification Must Extend to Assumptions
Industrial AI introduces an important distinction between validating a result and examining the assumptions on which it depends. A simulation may satisfy its acceptance criteria while relying on assumptions that no longer reflect operating conditions. A manufacturing optimisation may be technically valid against outdated constraints. An approved engineering change may become inappropriate following new regulatory requirements.
Verification establishes whether specified criteria have been satisfied; it does not necessarily establish whether those criteria remain appropriate. As AI assumes greater responsibility for engineering analysis and execution, this distinction becomes increasingly consequential.
Connected enterprise context can expose dependencies between decisions, assumptions and operational evidence. AI can help identify inconsistencies and conditions warranting reassessment, but detection alone does not improve a decision.
An effective loop must establish whether a change is material, what evidence is required, who has authority to act and whether the response should be automated, independently verified or escalated. Equally, not every new signal warrants intervention. Excessive reassessment creates delays, consumes expertise and can undermine automation benefits.
The right loop is not necessarily the one with the most controls, but the one applying appropriate governance when it matters.
Learning Must Improve Decisions and Context
Perhaps the greatest opportunity lies in connecting decisions to their outcomes. Enterprise applications record transactions, configurations, approvals and process execution, but these records do not necessarily establish whether the assumptions behind a decision proved correct or whether intended benefits were realised.
Consider an engineering change intended to reduce manufacturing cost. It passes technical validation, follows the required approval process and is implemented successfully. Months later, quality issues increase or operational savings fall short of expectations. The decision may remain fully traceable, yet traceability alone does not establish whether it was effective.
Closing the loop requires connecting the original rationale to subsequent evidence, examining contributing factors and determining what should change. Importantly, this is not simply a question of adding information to an existing knowledge base.
Learning can operate at three distinct, interconnected levels:
- Context: Refining the evidence, assumptions and dependencies informing decisions.
- Decision criteria: Improving engineering requirements, validation methods and risk thresholds.
- Operating model: Adapting decision rights, autonomy boundaries, governance and escalation mechanisms.
Better context does not automatically produce better decision criteria, nor does it necessarily change how an organisation operates. Continuous context tuning should therefore be an outcome of effective learning loops, not an isolated information-management objective.
This does not imply rewriting historical decisions or automatically modifying authoritative records. Evidence and decision history must remain traceable, while changes to governed product or business information follow established controls. Nor should every outcome generate a new rule. Industrial environments involve complex dependencies and uncertain conditions. Learning requires evidence, judgement and sufficient confidence that observed outcomes justify changing future practices.
The objective is to use experience to improve decision-making, not merely explain what happened.
The Operating Model Determines the Value
Much of the industrial AI discussion concentrates on technical architecture: knowledge graphs, semantic layers, digital threads, agents and increasingly capable models. These technologies can expose dependencies, connect information and support activities across fragmented enterprise applications. But technology alone cannot determine which decisions matter, who owns them, how much autonomy is appropriate or how outcomes should influence future practices.
This is where the cognitive data thread becomes relevant. Beyond connecting lifecycle information, it can provide the contextual relationships between decisions, assumptions, dependencies and outcomes needed to support intelligent action. Its value is not simply in making knowledge accessible, but in enabling the right decision loops across the PLM ecosystem while continuously refining the context that informs future decisions.
Those responsibilities belong to the operating model.
Within the broader PLM ecosystem, PDM, ERP, MRP, MES and other capabilities contribute different dimensions of product and operational information. Their responsibilities remain distinct, even as AI supports decisions spanning functional and technological boundaries. The challenge is not to replace this ecosystem with another overarching application or intelligence layer, but to improve how its capabilities support lifecycle decisions.
This requires more than defining approval workflows. Organisations need explicit decision rights, autonomy thresholds, exception-handling mechanisms and feedback processes. These should reflect the materiality of decisions and be reassessed as evidence accumulates, operating conditions change, and AI capabilities mature.
For example, an automated manufacturing adjustment that consistently performs within validated limits may justify greater autonomy. Conversely, repeated quality escapes may indicate that existing thresholds, escalation rules or validation criteria are inadequate. In both cases, the operating model should evolve through governed assessment rather than informal changes to automation behaviour.
This also changes how success should be measured. The number of connected applications, deployed agents, or automated workflows may indicate technical progress, but it does not necessarily demonstrate business value. Decision quality, engineering rework, time to resolution, compliance performance, operational reliability and realised lifecycle value provide more meaningful evidence.
The right loops should make these outcomes measurable and provide feedback to improve the operating model over time.
From Remembering to Learning and Unlearning
Industrial AI offers an opportunity to improve how enterprises use knowledge across the product lifecycle. Better-connected information, preserved decision context and increasingly autonomous workflows provide important foundations, but they are not the destination.
The greater opportunity lies in designing an operating model that can learn from experience and adapt. The right decision loops should determine where AI can act, where human judgement remains essential, when assumptions require reassessment and how outcomes influence subsequent decisions.
This also requires organisations to unlearn. Established assumptions, decision criteria and operating practices may have been appropriate when introduced but become less effective as technologies, products and business conditions evolve. Unlearning does not mean discarding historical knowledge. It means recognising when existing practices should be challenged, revised or retired, while preserving the evidence behind those decisions.
The right loops must improve more than the decisions themselves. They must also improve how decisions are made, who makes them and where autonomy is appropriate.
An enterprise that remembers can explain what happened and why. An enterprise that learns and unlearns can challenge established assumptions, adapt its operating model and improve what happens next.
The real measure of industrial AI will not be how much an organisation knows or automates, but how effectively it improves its decisions, context and operating model through experience.
What are your thoughts?
Disclaimer: articles and thoughts published on v+d do not necessarily represent the views of the company, but solely the views or interpretations of the author(s); reviews, insights and mentions of publications, products, or services do neither constitute endorsement, nor recommendations for purchase or adoption.