Digital Transformation, Data Analytics, and Automation Driving Operational Performance in Pharmaceutical Manufacturing

Full text

Text extracted from the article PDF for reading and search. Tables, figures and equations are in the PDF, which remains the authoritative version.

1. Introduction

1.1. Background to The Study

Pharmaceutical manufacturing occupies an unusual position among industrial sectors. It is simultaneously among the most scientifically sophisticated and among the most operationally conservative. The reason is structural rather than cultural. A change to a validated process is not simply an engineering decision; it is a regulatory event with filing consequences, inspection exposure, and potential supply interruption for products on which patients depend. The rational response to that incentive structure, sustained over several decades, has been to freeze processes once approved and to absorb inefficiency as the cost of assurance.

The consequence is a productivity profile that lags comparable process industries. Pharmaceutical plants have historically operated at overall equipment effectiveness levels well below those achieved in petrochemicals, semiconductors, or automotive assembly, carried substantial inventory buffers to insulate against process variability, and devoted a large proportion of skilled labour hours to documentation rather than to production or investigation. Deviation investigations, batch record review, and the reconciliation of information held across disconnected systems consume capacity that adds assurance but not throughput.

By 2021 that equilibrium was under visible strain. The COVID-19 pandemic compressed development and scale-up timelines that would ordinarily have spanned years into months, exposed the fragility of concentrated active pharmaceutical ingredient supply, and made the responsiveness of manufacturing networks a subject of political and public attention. At the same time the regulatory posture had shifted. Regulators had spent more than fifteen years signalling that science-based, risk-based, data-rich manufacturing was not merely permitted but preferred, beginning with the Process Analytical Technology framework and continuing through the ICH quality trilogy on pharmaceutical development, quality risk management, and pharmaceutical quality systems.

Against that background, the term Pharma 4.0 came into common use to describe the application of Industry 4.0 concepts to the specific constraints of medicinal product manufacture. The published literature of the period documents both the technological promise and the persistent adoption gap. Arden and colleagues, writing from the United States Food and Drug Administration in 2021, set out how the internet of things, artificial intelligence, robotics, and advanced computing were beginning to challenge established approaches, practices, and business models for pharmaceutical manufacture, and argued that the extensive data and analysis enabled by Industry 4.0 would allow both manufacturers and regulators to move further toward performance-based regulation (Arden et al., 2021), an argument that builds on an earlier programme of work connecting quality by design to regulatory flexibility (Yu et al., 2014). Reinhardt, Oliveira, and Ring, surveying the Irish pharmaceutical and biopharmaceutical sector, found that knowledge and involvement in Industry 4.0 implementation remained notably deficient across the sector despite widespread rhetorical commitment (Reinhardt, Oliveira and Ring, 2020, 2021). Hole, Hole, and McFalone-Shaw, examining digitalisation in contract development and manufacturing organisations, observed that the industry had been comparatively resistant to digitalisation even though robotics, automation solutions, and computerisation offered cost reduction, efficiency and productivity improvement, and greater responsiveness to change (Hole, Hole and McFalone-Shaw, 2021). Digitisation and analytics have likewise been linked to reductions in the substantial downtime to which pharmaceutical plants are prone (Anthony Jnr and Abbas Petersen, 2021).

What the literature of this period established less clearly is the question of magnitude and mechanism. That digital technologies are associated with better operations is asserted widely and demonstrated narrowly, usually through single-site case studies or vendor-reported implementations. Which technology class contributes most, whether the contributions are additive or interdependent, and what organisational conditions determine whether investment converts into performance, remained substantially open questions. Evidence from adjacent domains sharpens the question. Work on health system transformation has argued that scientific and technological innovation fails to translate into outcome improvement wherever the surrounding delivery infrastructure and organisational capability are absent (Ilodigwe and Adesemoye, 2019b), and that sustainable transformation rests on organisational strategy, operational excellence, and workforce capability rather than on technology procurement. Parallel findings in supply chain digitalisation, where transformation programmes carrying embedded risk and compliance controls outperform those treating compliance as a downstream concern (Mbonu et al., 2020a), point toward the same conclusion. This study addresses that gap.

1.2. Statement of The Problem

Pharmaceutical manufacturers committed substantial capital to digital initiatives across the 2018 to 2021 period, yet the realised operational return was widely reported as disappointing relative to business case projections. Three explanations recur in practitioner accounts and each has different remedial implications.

The first is a technology selection problem: firms invested in the wrong technology class, typically favouring visible equipment automation over less visible data infrastructure. The second is an integration problem: firms deployed capable technologies as isolated islands, so that a vision inspection system, a manufacturing execution system, and a laboratory information management system each performed adequately in isolation while producing no integrated view of the process. The third is an organisational absorption problem: the technology functioned as specified but the surrounding decision rights, analytical skills, and quality system procedures did not change, so that the new information had no route into action.

These explanations are not mutually exclusive, and they cannot be distinguished by descriptive case study. The third has been documented in other sectors under the label of the adoption gap, in which technically sound systems fail because the organisation surrounding them was never prepared to receive them (Ilodigwe and Adesemoye, 2019b). Distinguishing them requires a model that specifies the technology classes separately, models the pathways between them, and includes the organisational variables through which effects are transmitted. That is what this study constructs and tests.

1.3. Research Objectives

The study pursues five objectives:

1. To assess the relationship between data analytics capability and operational performance in pharmaceutical manufacturing operations. 2. To assess the relationship between process automation intensity and operational performance.

3. To assess the relationship between intelligent decision support adoption and operational performance. 4. To determine whether data governance and integrity maturity mediates the relationship between data analytics capability and operational performance.

5. To determine whether organisational absorptive capacity

moderates the relationship between intelligent decision support adoption and operational performance.

1.4. Research Questions

RQ1. To what extent is data analytics capability associated with operational performance in GMP pharmaceutical manufacturing?

RQ2. To what extent is process automation intensity associated with operational performance, and does that association depend on the level of analytics capability? RQ3. To what extent is intelligent decision support adoption associated with operational performance? RQ4. Does data governance and integrity maturity function as a transmission mechanism between analytics investment and operational outcomes?

RQ5. Does absorptive capacity condition the operational benefit obtained from intelligent decision support systems?

1.5. Significance of The Study

For operations leadership, the study offers evidence on sequencing. If data governance maturity mediates the analytics to performance relationship, then investment sequencing that places advanced analytics ahead of governance foundations is predictably inefficient, and the finding provides a defensible basis for reordering a capital plan.

For quality and regulatory functions, the study connects digital maturity to compliance outcomes rather than treating compliance as an obstacle to digitalisation. Deviation rates, right-first-time performance, and investigation cycle time are modelled as components of operational performance, which reframes the digital agenda as a quality agenda. For regulators and policy bodies, the study provides evidence relevant to the transition toward performance-based oversight. If digitally mature sites demonstrate measurably superior quality outcomes, that supports the case for regulatory approaches calibrated to demonstrated performance.

For the academic literature, the study extends dynamic capabilities theory and the technology-organisation-environment framework into a highly regulated operating context in which the environment dimension is unusually binding, and it supplies a validated instrument for subsequent replication.

1.6. Scope and Delimitations

The study is bounded in four ways. It addresses manufacturing and supply operations only, excluding drug discovery, clinical development, commercial, and pharmacovigilance functions, although the instrument acknowledges interfaces with those functions. It addresses the 2021 operating period, capturing a moment shaped by pandemic response and therefore not fully generalisable to steady-state conditions. It measures perceptual operational performance reported by informed respondents rather than audited financial or regulatory data, a limitation addressed in Section 7. It samples GMP-licensed sites, both innovator-owned and contract manufacturing, and therefore does not generalise to compounding pharmacies, over-the-counter-only producers, or non-regulated nutraceutical manufacture.

2. Literature Review

2.1. Conceptualising Digital Transformation in

Operations Digital transformation is distinguishable from digitisation and digitalisation, and the distinction matters analytically. Digitisation converts analogue information into digital form, for example replacing a paper batch record with an electronic one that reproduces the same fields and the same workflow. Digitalisation uses digital technology to change how a process is executed, for example replacing end-product laboratory testing with in-line spectroscopic measurement. Digital transformation changes the operating model itself, altering what decisions are made, by whom, on what evidence, and at what cadence.

Much of what pharmaceutical firms described as transformation during the period under study was in fact digitisation. An electronic batch record that reproduces the sequence and the review burden of its paper predecessor delivers legibility and audit trail improvements but leaves the underlying operating model intact. The same distinction appears in process improvement work outside the sector, where benefit is found to depend on redesigning the underlying process rather than on placing an existing process onto a new platform (Ambali et al., 2021). This distinction underlies the study's decision to measure capability and integration rather than technology presence. Asking whether a site has a manufacturing execution system yields little discriminating information by 2021; asking whether process data from that system is routinely used to adjust process parameters yields a great deal.

2.2. Pharma 4.0 and the regulated operating context

Industry 4.0 as originally articulated rests on cyber-physical systems, the internet of things, cloud computing, and cognitive computing, integrated so that physical processes and their digital representations remain continuously coupled. Arden and colleagues describe the enabling architecture for pharmaceutical application as resting on two foundational technology layers, data capture through advanced sensing and data storage through cloud infrastructure, which together create the data-rich environment that permits big data analysis and simulation, artificial intelligence and adaptive control, digital twins, and cyber-physical systems, and which in turn enable precision robotics and augmented reality applications on the production floor (Arden et al., 2021). Cyber-physical framings of process analytical technology develop the same architecture from the control side (Barenji et al., 2019). The pharmaceutical adaptation of this architecture faces constraints absent in discrete manufacturing. Computerised systems require validation under Annex 11 of the EU GMP guidelines (European Commission, 2011) and 21 CFR Part 11 in the United States. Changes to validated processes carry filing obligations. Records must satisfy data integrity expectations captured in the ALCOA+ principles, meaning that data must be attributable, legible, contemporaneous, original, accurate, complete, consistent, enduring, and available (United States Food and Drug Administration, 2018). Machine learning models that adapt in production create a particular difficulty, since a system whose behaviour changes after validation is difficult to reconcile with a regulatory framework built around demonstrating and maintaining a validated state.

Reinhardt and colleagues found that adoption in the pharmaceutical and biopharmaceutical sector lagged the ambition expressed by firms, with genuine knowledge and involvement in implementation concentrated in a small proportion of the workforce. Their finding that the sector's stated intent substantially exceeded its realised capability (Reinhardt, Oliveira and Ring, 2020) is one motivation for measuring capability constructs in this study rather than adoption declarations.

2.3. Data Analytics in Pharmaceutical Operations

Analytics applications in pharmaceutical operations divide usefully into four levels of ambition.

Descriptive analytics summarises what occurred and, in the pharmaceutical setting, is most familiar as the annual product quality review and routine statistical process control. Value here is real but incremental. Diagnostic analytics explains why an outcome occurred, and its highest-value pharmaceutical application is deviation root cause analysis. Multivariate methods including principal component analysis and partial least squares regression, whose progression from detection to diagnosis to prognosis is traced by Reis and Gins (2017), allow a site to identify which combination of raw material attributes and process parameters accounts for a yield excursion, a question that univariate trending cannot answer when the causal structure is interactive. Predictive analytics forecasts what will occur, supporting predictive maintenance on critical equipment, batch outcome prediction during processing, and demand forecasting across the supply network, where predictive models have been shown to improve forecast accuracy and reduce the inventory buffers held against forecast error (Aifuwa et al., 2020). Comparable analytics-to-efficiency relationships are reported in other production sectors, including data-driven models applied to production efficiency and resource optimisation in agricultural systems (Michael and Ogunsola, 2021c). Prescriptive analytics recommends action, and in the pharmaceutical setting this shades directly into the decision support construct treated separately in Section 2.5. Two implementation questions cut across all four levels. The first is where analytical logic should reside. Analysis run as a periodic offline exercise informs review; analysis embedded in the operational data flow informs control, and risk-based business intelligence architectures describe how that embedding is structured (Mbonu et al., 2021a). The second is how analytical output reaches decision makers, an interface problem in which dashboard design determines whether managers obtain genuine visibility or merely additional reporting (Sanni and Atima, 2021b). Comparable work on structured profitability analysis using general-purpose analytical tools shows that much of the practical benefit arises from disciplined data preparation rather than from analytical sophistication (Akin-Oluyomi et al., 2020), a finding of direct relevance to sites where process data must be reconciled manually before it can be analysed at all. Executive decision systems built on integrated financial and operational data describe the organisational end state (Lawal and Oduleye, 2019b).

Method transfer from other data-rich regulated domains is substantial. Predictive detection methods developed for financial surveillance, where the task is to separate genuinely irregular patterns from ordinary variation across very high transaction volumes (Atakpa, 2021), share their statistical foundation with out-of-trend and out-of-specification detection in process monitoring. Both settings face the same operating constraint, namely that a false positive consumes investigative capacity while a false negative carries regulatory consequence.

Forecasting method transfers similarly. Machine learning approaches to demand forecasting developed for revenue projection (Tonoyan, Dada and Ayivi-Donkor, 2021a) and working capital optimisation (Tonoyan, Dada and Ayivi-Donkor, 2021b) address the same structural problem as pharmaceutical demand planning, in which forecast error is absorbed by inventory that is expensive to hold and, for products with defined shelf life, cannot be held indefinitely. Forecasting models built on customer relationship data in multichannel settings illustrate the value of combining transactional and behavioural signals (Dada, Isiekwu and Oluwo, 2021b), an approach with parallels in combining process and materials data for batch outcome prediction. Process Analytical Technology occupies a distinctive position because it is both an analytics practice and a regulatory framework. The FDA guidance of 2004 defined a system for designing, analysing, and controlling manufacturing through timely measurement of critical quality and performance attributes of raw and in-process materials, with the goal of ensuring final product quality (United States Food and Drug Administration, 2004). PAT enables real-time release testing, in which product is released on the basis of in-process data rather than end-product testing, and Markl and colleagues document both the state of the art and the persistent obstacles in that transition (Markl et al., 2020). The operational significance is considerable: laboratory testing cycle time frequently exceeds manufacturing cycle time in solid dosage production, so eliminating it changes the plant's throughput arithmetic rather than merely improving it, a shift closely bound up with the wider move from batch toward continuous production (Lee et al., 2015).

2.4. Automation and Robotics in Pharmaceutical

Operations Automation in pharmaceutical operations spans several distinguishable categories, and treating them as a single construct obscures important differences. Process automation covers distributed control systems and programmable logic controllers governing unit operations, together with automated material handling and dispensing. Aseptic processing automation, including isolator and restricted access barrier technology combined with robotic manipulation, has a quality rationale that is largely independent of cost, since removing the human operator from the critical zone removes the dominant source of microbial contamination risk. Packaging and serialisation automation was driven in this period by regulatory mandate, notably the Drug Supply Chain Security Act in the United States and the Falsified Medicines Directive in the European Union, both of which required item-level identification infrastructure that incidentally created a substantial data asset. Laboratory automation covers robotic sample preparation, automated dissolution and chromatography, and laboratory information management system integration. Machine vision and automated classification, developed extensively for remote inspection of distributed assets (Dagodzo, 2018a, 2018b; Dagodzo and Ahiaeke Patrick, 2020, 2021b), extend automation into inspection tasks previously requiring human judgement, and geospatially referenced asset management systems (Dagodzo and Ahiaeke Patrick, 2021a) offer a model for maintaining equipment master data across a distributed manufacturing network. Robotic process automation, distinct from physical robotics, automates transactional workflows in the software layer, and by 2021 was being applied to batch record review, deviation triage, and regulatory submission assembly, Warehouse and materials automation has comparable application to pharmaceutical warehousing, where segregation, status control, and traceability requirements make manual material movement a recurring source of error.

A further category, condition monitoring and predictive maintenance, sits between automation and analytics. The transition from reactive to predictive maintenance regimes has been documented across mechanical and building systems (Sunday et al., 2020), with vibration-based condition monitoring of rotating equipment and sensor-based fault detection in air handling systems among the most developed techniques, the latter resting on established work relating control strategy to thermodynamic efficiency in air conditioning plant (Sunday et al., 2019). Interpretable machine learning approaches to early failure prediction in distributed assets (Komi and Adamolekun, 2021) are of particular interest in the GMP context, since a maintenance intervention triggered by a model that cannot be interrogated is difficult to justify in a validated environment, and hybrid modelling approaches developed for forecasting under physical constraints (Komi, 2021) address the related problem of combining mechanistic and empirical components. The HVAC case carries particular pharmaceutical significance, since cleanroom classification depends on continuous performance of air handling systems and an undetected drift in differential pressure or particulate load is a quality event rather than an engineering inconvenience. The organising principle is prioritisation of maintenance effort against criticality rather than against schedule, which aligns closely with the risk-based allocation of control effort required under ICH Q9 (International Council for Harmonisation, 2005).

The pharmaceutical-specific complication is that automation interacts with validation burden in a non-obvious way. Automating a process step reduces variability and human error, which reduces deviations. It also adds a computerised system that must itself be validated and maintained in a validated state, and it creates a new failure mode in which system malfunction affects an entire campaign rather than a single operator error affecting a single batch. Firms that automated without correspondingly maturing their computerised system lifecycle management sometimes exchanged one deviation category for another. Asset integrity inspection illustrates the same tension from the equipment side. Non-destructive testing regimes developed for process equipment and pressure systems (Atta et al., 2020; Atta, Tofah and Adenuga, 2021) show how inspection frequency, technique selection, and risk assessment interact, and the pharmaceutical parallel is direct: equipment qualification status, like weld integrity, is a property that must be actively maintained and evidenced rather than assumed to persist. Automation also extends the attack surface. Connecting previously isolated process control systems to enterprise networks creates the information technology and operational technology convergence problem documented at length in energy and utility settings, where the security architecture, the compliance regime, and the availability requirements of the two domains differ fundamentally (Adegbite, Adebayo and Ahmed, 2020). Zero trust architectures are the principal response (Ahmed, Adegbite and Adebayo, 2021), and they impose organisational requirements that an equipment-focused automation programme will not have anticipated. This reasoning motivates the study's hypothesis that automation intensity yields weaker unconditional returns than analytics capability and that its returns are conditional on data infrastructure.

2.5. Intelligent Decision Support Systems

Intelligent decision support systems combine data, models, and interface design to improve the quality and speed of human decisions, and are distinguished from analytics by their orientation toward action and from full automation by their retention of the human decision maker. In the pharmaceutical operations setting the significant categories are advanced planning and scheduling, real-time process control advisory systems, deviation and CAPA triage support, supply chain control towers offering network-wide visibility with exception-based alerting, and digital twins. The control tower category draws directly on end-to-end visibility frameworks developed for transparency, compliance, and traceability across complex supply networks (Nnabueze et al., 2021), and on materials readiness models that treat supply visibility as a determinant of operational uptime rather than of holding cost (Okonkwo et al., 2021a). A distinct rationale for decision support, less often stated, concerns decision quality rather than decision speed. Structured decision models for capital allocation demonstrate that the value of an analytical system frequently lies in disciplining the reasoning behind a decision rather than in accelerating it (Lawal and Oduleye, 2021a), and integration models linking planning analytics to strategy make the same argument at the level of the management system (Lawal and Oduleye, 2021b). Predictive compliance monitoring, in which routine operational reporting is analysed to detect systemic risk before it manifests as an event, has been developed most fully in aviation safety oversight (Adeyelu, 2018) and maps closely onto the pharmaceutical use of deviation and complaint data as leading quality signals. Digital twins deserve separate mention because they represent the most theoretically ambitious category. A digital twin is a computational representation of a physical asset or process, maintained in synchrony with the physical system through live data exchange, and used for prediction, optimisation, and scenario evaluation. Chen and colleagues, reviewing digital twins in pharmaceutical and biopharmaceutical manufacturing, distinguish mechanistic, data-driven, and hybrid modelling approaches and observe that hybrid approaches are frequently the most practical in the pharmaceutical setting, where mechanistic understanding is partial and data volumes are insufficient to support purely empirical models for complex biological processes (Chen et al., 2020).

The regulatory question that intelligent decision support raises with particular force concerns the role of the human. When an advisory system recommends a process adjustment, the operator who accepts the recommendation bears the documented responsibility, yet may lack the information required to evaluate it independently. The resulting accountability ambiguity is a recognised barrier to adoption and one reason that decision support benefits are expected to be conditional on the organisation's absorptive capacity.

2.6. Operational Performance in Regulated

Manufacturing Operational performance in pharmaceutical manufacturing cannot be adequately captured by the cost, quality, delivery, and flexibility framework inherited from general operations management, because compliance performance functions as a constraint rather than as one dimension among several. A site that achieves low unit cost while accumulating regulatory observations has not traded off performance dimensions; it has failed, because the regulatory consequence can extinguish the site's licence to operate. This study therefore models operational performance as a construct with five dimensions. Cost efficiency covers unit conversion cost, yield, and scrap. Quality performance covers right-first-time rate, deviation frequency, and batch rejection. Delivery performance covers schedule adherence, cycle time, and order fulfilment. Flexibility covers changeover time, product mix responsiveness, and volume responsiveness. Compliance performance covers inspection outcomes, observation closure timeliness, and data integrity findings. Each dimension has an established measurement literature in adjacent operating contexts. Cost efficiency and operational continuity have been modelled jointly in supply operations research, where inventory availability is linked explicitly to plant uptime rather than to holding cost alone, so that a cost reduction compromising continuity counts as a failure rather than as a trade-off (Okonkwo, Ogunwole and Okeke, 2018b). Delivery and cycle time performance draw on lean practice, where waste elimination and flow improvement are the primary levers and where adaptations of Lean Six Sigma for resource-constrained organisations show that the method does not require large-scale deployment infrastructure to produce results (Ambali et al., 2021). Materials readiness has been modelled as a maintenance-driven rather than a purely procurement-driven determinant of operational performance (Okonkwo et al., 2021a), a framing that fits pharmaceutical operations closely, since a qualified material that is unavailable and an unqualified material that is present are equally blocking. Logistics performance carries a comparable dual character, in which delay costs are incurred at points of transfer rather than within the operation itself (Okonkwo et al., 2020). Process automation applied to procurement and administrative workflows has been shown to improve both efficiency and transparency simultaneously (Akinleye and Adeyoyin, 2021), the second of which matters more in a GMP setting than the first, and structured operations management frameworks demonstrate how performance measurement systems are built where outcomes are multidimensional and partly qualitative (Efobi, Akinleye and Fasawe, 2021).

Two additional considerations shape measurement. Overall equipment effectiveness, the standard composite productivity measure, behaves unusually in pharmaceutical manufacturing because cleaning and changeover requirements between products are dictated by cross-contamination control rather than by efficiency optimisation, so a low availability figure may reflect appropriate control rather than poor practice. Deviation counts are similarly ambiguous, since a rising count may indicate deteriorating process control or improving detection and reporting culture. The instrument therefore emphasises deviation recurrence and investigation cycle time alongside raw frequency. Two measurement problems recur. The first is attributing performance change to a specific intervention when several changes occur concurrently, a difficulty documented in analytical work on return measurement in regulated service settings (Sanni, Ajiga and Atima, 2020a) and one reason this study measures standing capability rather than project outcomes. The second is the differing time constants of the five dimensions: cycle time responds to process redesign within weeks, whereas compliance performance is observable only at inspection intervals measured in years, so a cross-sectional design necessarily captures them at different stages of response.

2.7. The Regulatory and Quality Architecture

Five instruments define the environment within which the constructs of this study operate.

The ICH Q8(R2) guideline on pharmaceutical development established Quality by Design, shifting the basis of quality assurance from end-product testing to designed-in process understanding, and introduced the design space within which movement is not considered a change requiring regulatory approval (International Council for Harmonisation, 2009; Yu et al., 2014). ICH Q9 established quality risk management as the basis for allocating control effort (International Council for Harmonisation, 2005). ICH Q10 established the pharmaceutical quality system, including continual improvement and management review obligations (International Council for Harmonisation, 2008). The FDA PAT guidance of 2004 provided the framework for in-process measurement and control. ICH Q12, finalised in 2019, addressed lifecycle management and established mechanisms including established conditions and post-approval change management protocols that materially reduce the regulatory friction associated with process change (International Council for Harmonisation, 2019).

Taken together these instruments describe a regulatory system that had, by 2021, largely removed the formal obstacles to data-rich, adaptive manufacturing. The persistence of conservative practice therefore reflects organisational risk aversion, sunk investment in existing validated processes, and capability gaps rather than regulatory prohibition. This observation is important for interpreting the study's findings: where digital maturity is low, the binding constraint is more likely to be internal than external.

The operational expression of this architecture is procedural. Standard operating procedures are the mechanism through which regulatory requirement becomes daily practice, and are better understood as operational assets than as administrative overhead. Comparable compliance-by-design reasoning appears in procurement frameworks for high-risk operating environments (Okonkwo et al., 2021b) and in the design of laboratory and diagnostic facilities operating under controlled environment requirements closely analogous to GMP classification (Ogbete, Aminu-Ibrahim and Ambali, 2019, 2021).

Corrective action is the second procedural mechanism. Regulatory findings are closed through corrective and preventive action, and the governance of that process determines whether closure is substantive or documentary. Risk-based allocation of assurance effort, the principle underlying ICH Q9, has parallel development in internal audit practice, where the same problem arises of directing finite verification capacity toward the exposures that matter most (Akomolafe and Agu, 2018, 2019a). The recurring lesson from that literature is that controls embedded in the systems generating the records produce more durable assurance than review layered on top of them, which is precisely the case for digital approaches to compliance. Regulatory frameworks in other safety-critical sectors formalise this in explicit corrective action protocols linking evidence, enforcement, and verified closure (Adeyelu, 2021), and adaptive safety management system models describe how such protocols are implemented where regulatory capacity is itself constrained (Adeyelu, 2019). Translating requirement into workflow is the practical difficulty. Frameworks for converting statutory and regulatory obligations into executable operational procedures (Eyetsemitan et al., 2021) address the same problem that GMP procedure design faces, namely that a requirement expressed in regulatory language must become a sequence of actions an operator can perform and an auditor can verify. Reconciliation control and workflow redesign in transaction-intensive environments (Dogbatsey, Ebhojie and Oyeleye, 2019) illustrate the mechanism, and cross-border compliance analytics describe the additional complexity where one organisation answers to several regulatory regimes at once (Lawal and Oduleye, 2018b).

2.8. Validation, Computerised System Lifecycle, And

Access Governance Automation and analytics both create computerised systems, and computerised systems in a GMP environment carry lifecycle obligations that the underlying technology literature rarely addresses. Five requirements recur.

Access control and attribution: Because every GMP record must be attributable to an identified individual, identity and access management is a regulatory control rather than merely an operational one, and its difficulty rises sharply once systems span on-premises and cloud environments (Mbonu et al., 2020b). Directory-based identity management supplies the practical mechanism in most enterprise estates (Ladapo et al., 2019), and assessment work on the security posture of such systems documents the failure modes that arise when they are treated as infrastructure rather than as controlled systems (Ladapo et al., 2018). Zero trust approaches to enterprise identity governance extend the same principle to environments where network location no longer implies trust (Ogbole et al., 2021). Configuration drift introduces an equivalent exposure, since a system validated in one configuration is not validated in another, which is the problem addressed by governance of infrastructure defined as code (Mbonu et al., 2020c).

Testing and regression control: Validation evidence must be regenerated whenever a system changes, and the cost of that regeneration is the principal reason firms avoid changing validated systems at all. Automated testing and regression control, including machine learning assisted approaches (Mbonu et al., 2021b), attack that cost directly and are therefore of greater strategic significance in pharmaceutical operations than their technical description suggests. Continuous integration and deployment practice addresses the same cost from the delivery side, since a pipeline that regenerates test evidence automatically converts change control from a project into a routine (Badmus, Dosunmu and Ozowara, 2018), and comparative evaluation of the available tooling shows that the governance capability of the pipeline matters more than its technical features (Badmus et al., 2020).

Interface governance: Integration between manufacturing execution, laboratory, enterprise resource planning, and quality systems is where most pharmaceutical data integration effort is spent, and where uncontrolled interfaces create data integrity exposure. Extract, transform, and load design determines whether data crossing a system boundary retains its meaning and its audit trail (Badmus, Dosunmu and Ozowara, 2019a), and platform governance frameworks developed specifically for regulated healthcare environments demonstrate how configuration control, access control, and change control are held together as one discipline rather than three (Badmus, Dosunmu and Ozowara, 2019b).

Log analytics and monitoring: Audit trail review is a GMP requirement that most sites discharge by sampling, because complete manual review is infeasible. Enterprise log analytics and automated incident response architectures (Mbonu et al., 2019b) provide the technical basis for exhaustive rather than sampled review. Comparative analysis of data protection obligations across jurisdictions (Mbonu et al., 2019a) becomes material for manufacturers operating multi-jurisdiction networks, since the same record may attract different obligations at different sites in one network, and privacy-centric approaches to security engineering show that these obligations are more cheaply met by design than by retrofit (Okoruwa et al., 2020).

Detection and response: A validated system that has been compromised is no longer in a validated state, which makes security monitoring a data integrity control rather than solely an information technology one. Enterprise risk assessment and security audit frameworks establish the baseline (Dosunmu and Ogundele, 2019), intrusion detection and prevention supply the monitoring layer (Dosunmu and Ogundele, 2020), and incident response and forensic capability determine how quickly a site can establish what was affected and whether product quality decisions relied on compromised data (Dosunmu and Ogundele, 2021).

2.9. Data Governance and Integrity as An Operational

Construct Data governance is treated here as a construct distinct from analytics capability, because the two are separable in practice and because governance carries an independent regulatory consequence in the pharmaceutical setting. The architectural case for treating governance as foundational has been made most directly in health system work describing interoperability, data quality, and analytics governance as a single data backbone on which every downstream analytical capability depends (Ilodigwe and Adesemoye, 2021). Three governance sub-capabilities follow from that framing. The first is ownership and classification, meaning the ability to identify what data an organisation holds, how sensitive it is, and where regulatory obligation attaches to it, a question that acquires additional complexity where a manufacturer operates across jurisdictions with differing data protection regimes (Mbonu et al., 2019a). The second is access governance, since the attribution requirement central to GMP record-keeping cannot be met unless identity and permission are controlled consistently across every system holding those records (Mbonu et al., 2020b). The third is control assurance, meaning the mechanisms by which an organisation demonstrates that its governance operates as designed, which corresponds directly to the pharmaceutical requirement for periodic audit trail review and draws on the same technology-enabled assurance methods developed in internal audit practice (Akomolafe and Agu, 2018).

Governance failure carries legal and ethical exposure beyond the operational (Mbonu et al., 2018). The design question is therefore not whether to constrain analytical use of operational data but how to reconcile analytical utility with regulatory obligation architecturally rather than by restriction, which is the same reconciliation that risk-based control models attempt in financial services (Akomolafe and Agu, 2019a). This reconciliation is of evident relevance to innovator and contract manufacturer relationships, in which process data is commercially sensitive on both sides yet must be jointly governed.

2.10. Human Factors, Deviation Investigation, And

Organisational Learning Deviation investigation is the largest single consumer of quality function capacity at most pharmaceutical sites, and its effectiveness determines whether a site's deviation rate declines over time or merely persists. The industrial safety literature has developed this problem in greater analytical depth than the pharmaceutical quality literature has, and the mapping is close: an incident is to safety what a deviation is to quality, and the investigation and prevention machinery is structurally identical.

Four findings transfer. First, human error is not an explanation but a category requiring decomposition, and causation frameworks distinguishing skill-based, rule-based, and knowledge-based failure modes yield different corrective actions than an undifferentiated attribution to operator error does (Obriki and Arumosoye, 2020). Second, leading indicators derived from near-miss and observation data give earlier signal than lagging event counts, but only where the reporting systems that generate them are deliberately designed rather than passively maintained (Arumosoye and Obriki, 2019; Obogo, Arumosoye and Obriki, 2021). Third, systematic hazard identification, conducted in advance rather than reconstructed after an event, determines how much of the risk landscape is visible at all (Obogo, Uduokhai and Ozobu, 2019). Fourth, the capacity to convert investigation output into changed practice is an organisational maturity property rather than a procedural one, and maturity models built on organisational learning describe the progression explicitly (Arumosoye and Obriki, 2021), with institutionalisation across project-based organisations posing a distinct difficulty where teams reform continually (Obriki and Arumosoye, 2021).

The organisational conditions supporting that conversion are also documented. Audit systems that verify whether controls operate as designed (Obogo, Arumosoye and Obriki, 2020a), together with the reporting culture determining whether events reach those systems at all, constitute the social subsystem through which technical monitoring capability becomes performance improvement. Leadership behaviour is the principal lever on that culture (Obogo, Ozobu and Uduokhai, 2019a), and the visible presence of senior personnel in the operating area has been linked directly to reporting and culture outcomes (Obriki and Arumosoye, 2019), a mechanism with an exact pharmaceutical analogue in management participation in gemba walks and self-inspection. Earlier conceptual work on data-driven risk control in high-consequence operations anticipates the same structure (Obriki and Arumosoye, 2018). This is the sociotechnical argument of Section 3.1 in applied form, and it supplies part of the justification for treating absorptive capacity as a moderator rather than as a control. Oversight of external parties raises the same questions across an organisational boundary. Contract manufacturing relationships resemble multi-contractor project environments in which the principal carries regulatory responsibility for work it does not directly supervise, and governance frameworks developed for contractor and subcontractor performance management (Arumosoye and Obriki, 2020; Obogo, Uduokhai and Ozobu, 2021) address precisely that structure, as do risk governance models for environments in which several contractors work concurrently under one principal (Obogo, Ozobu and Uduokhai, 2019b).

2.11. The Pandemic as Accelerant and Stress Test

The 2020 to 2021 period functioned as an unplanned natural experiment. Site access restrictions forced remote work in functions that had been presumed to require physical presence, including batch record review, deviation investigation, and, significantly, regulatory inspection. Remote and hybrid inspection approaches adopted by several agencies during this period required inspectable digital records and remote access to quality systems, which converted electronic documentation from a convenience into a condition of regulatory continuity.

Simultaneously, vaccine manufacturing scale-up compressed technology transfer timelines to an unprecedented degree. Transfers that would conventionally require twelve to eighteen months were executed in a fraction of that time, and the sites that achieved this were disproportionately those with mature digital process descriptions, because a process defined in structured digital form transfers faster than one defined in documents requiring human interpretation. Supply networks were stressed simultaneously. Resilience frameworks developed for critical infrastructure supply (Agbabiaka et al., 2019; Ogunwole et al., 2021) describe the design properties, namely redundancy, visibility, and substitutability, that determined which networks absorbed the shock,. Constraints were felt furthest downstream, where pharmacy-led delivery and access models became the binding determinant of whether manufactured product reached patients (Ilodigwe and Adesemoye, 2020). The pandemic therefore both accelerated adoption and revealed which forms of digital investment carried real operational weight. This is directly relevant to interpreting 2021 data: the cross-sectional associations observed in this period may be stronger than those observable in steady-state conditions, because the period placed unusual demands on responsiveness.

2.12. Cross-Sector Evidence from Adjacent Regulated

Operating Environments Because multi-site quantitative evidence within pharmaceutical manufacturing is limited, this study draws supporting evidence from adjacent sectors sharing the defining features of the pharmaceutical operating context: a binding regulatory constraint, a high consequence of failure, and a documentation burden that is itself a significant consumer of operational capacity.

Energy and infrastructure supply operations provide the closest analogue on the supply side. Risk management and resilience frameworks developed for these sectors address the same problem of coordinating procurement, materials, and maintenance under regulatory scrutiny (Agbabiaka et al., 2019; Ogunwole et al., 2021), and models linking materials readiness to plant uptime rather than to holding cost describe the operational logic that also governs pharmaceutical material availability (Okonkwo, Ogunwole and Okeke, 2018b).

Healthcare infrastructure and diagnostic laboratory operations provide the closest analogue on the facility side. Diagnostic laboratories operate under environmental control, contamination control, and regulatory design requirements closely paralleling GMP classification (Ogbete, Aminu-Ibrahim and Ambali, 2018, 2019), and risk-managed construction strategies for highly regulated healthcare facilities address the asset-side questions that manufacturing digitalisation eventually reaches (Ogbete, Aminu-Ibrahim and Ambali, 2021). Work connecting infrastructure investment to diagnostic access and delivery capability in constrained health systems (Aminu-Ibrahim, Ogbete and Ambali, 2018, 2019, 2020) provides a methodological precedent for the outcome-linked framing adopted here, and materials selection and energy efficiency strategies for laboratory facilities (Ogbete, Aminu-Ibrahim and Ambali, 2020) address the operating-cost dimension that facility decisions determine for the life of the asset. Industrial safety management provides the closest analogue on the leading-indicator question. The shift from lagging incident counts toward leading signals derived from near-miss and observation reporting (Arumosoye and Obriki, 2019; Obogo, Arumosoye and Obriki, 2021) parallels the pharmaceutical move from end-product testing to in-process signal detection, and encounter the same interpretive difficulty in which improved detection raises recorded event counts before it lowers them.

Aviation provides the closest analogue on the regulated-safety-oversight question, since it shares with pharmaceutical manufacturing a licensing regime, mandatory occurrence reporting, and an inspectorate with the power to halt operations. Predictive compliance monitoring built on routine reporting streams (Adeyelu, 2018), adaptive safety management system implementation under constrained regulatory capacity (Adeyelu, 2019), and structured corrective action protocols linking evidence to enforcement (Adeyelu, 2021) each address problems with direct pharmaceutical counterparts. Quantified risk modelling for low-frequency high-consequence hazards (Adeyelu, 2020) parallels the contamination and cross-contamination risk assessments that govern facility design. Financial audit and risk analytics provide the closest analogue on the assurance question. Technology-enabled risk assessment in internal audit (Akomolafe and Agu, 2018) and risk-based internal control models (Akomolafe and Agu, 2019a) address the same underlying problem as pharmaceutical quality assurance, namely how to obtain assurance over a high-volume process without inspecting every instance of it.

Procurement and vendor management extend the supply-side parallel. Strategic procurement optimisation in high-consequence operating environments (Okonkwo, Ogunwole and Okeke, 2018a), vendor governance and contract negotiation as cost levers (Okonkwo et al., 2019), data-driven vendor evaluation and bid selection (Ahiaeke Patrick et al., 2021), and supplier relationship management treated as a source of resilience rather than merely of price (Ike et al., 2021) map onto the pharmaceutical problem of qualifying suppliers of active ingredients and excipients, where switching cost is elevated by regulatory filing obligation. Enterprise resource planning integration with spatially referenced procurement data (Ahiaeke Patrick et al., 2020) and analytical models of enterprise value creation from financial data (Lawal and Oduleye, 2018a) describe the systems and the appraisal logic respectively. Small and medium enterprise operations research contributes on the implementation side, where resource constraints force sequencing questions that larger organisations can defer. Governance alignment across multiple stakeholders in regulated joint operations (Eyetsemitan et al., 2020) resembles the contract manufacturing structure, in which the principal must coordinate a process it does not wholly control, and lean adaptations for resource-constrained organisations demonstrate process benefit at a scale where the confounding effects present in large-firm studies are absent (Ambali et al., 2021).

Agricultural production systems provide a more distant but instructive parallel on adoption specifically. Work on extension services as the mechanism converting available innovation into adopted practice (Michael and Ogunsola, 2021e), alongside evidence that data-driven models improve production efficiency and resource use (Michael and Ogunsola, 2021c), describes an adoption problem structurally similar to the one examined here, in which technology availability substantially exceeds technology use and the binding constraint sits with the receiving organisation. These parallels are offered as supporting rather than substituting evidence. Differences in regulatory regime, product risk profile, and capital structure limit direct transfer, and the study's conclusions rest on its pharmaceutical sample.

2.13. Identified Gaps

Four gaps emerge from the reviewed literature. First, most pharmaceutical digitalisation research is conceptual, descriptive, or single-case, with limited multi-site quantitative modelling of technology-to-performance relationships. Second, studies typically treat digital technology as a single aggregate construct, preventing comparison of relative contribution across technology classes. Third, transmission mechanisms are under-theorised, so that the organisational conditions under which technology converts into performance remain largely unspecified. Fourth, compliance performance is generally treated as a constraint or an obstacle rather than modelled as a performance outcome that digital capability may improve. This study addresses each.

3. Theoretical Framework and Hypothesis Development

3.1. Theoretical Foundations

The study draws on three complementary theoretical traditions.

The resource-based view and dynamic capabilities: The resource-based view holds that sustained competitive advantage derives from resources that are valuable, rare, imperfectly imitable, and non-substitutable (Barney, 1991).

Digital technologies in isolation satisfy few of these conditions, since a manufacturing execution system or a vision inspection unit is purchasable by any competitor. The dynamic capabilities extension resolves this by locating advantage not in the assets but in the organisational capacity to sense opportunity, seize it through resource commitment, and reconfigure the resource base accordingly (Teece, Pisano and Shuen, 1997). Applied here, the argument is that analytics capability, understood as the socially complex combination of infrastructure, skills, and routines, is imperfectly imitable in a way that analytics software is not. This underpins the expectation that capability constructs will outperform adoption constructs as predictors.

The technology-organisation-environment framework: The TOE framework holds that innovation adoption and outcome are determined jointly by technological context, organisational context, and environmental context (Tornatzky and Fleischer, 1990). The pharmaceutical setting makes the environmental dimension unusually salient and unusually binding, since regulatory requirements constrain the design space of permissible implementations. The framework informs the inclusion of absorptive capacity as an organisational moderator and the treatment of compliance performance as an outcome rather than a control. Empirical support for the organisational dimension comes from work showing that delivery infrastructure and organisational capability, rather than the innovation itself, determine whether a technical advance reaches its intended outcome (Ilodigwe and Adesemoye, 2019b), and from adoption research outside the sector identifying structured knowledge transfer as the mechanism converting available innovation into adopted practice (Michael and Ogunsola, 2021e).

Sociotechnical systems theory: Sociotechnical systems theory holds that performance is a joint function of the technical and social subsystems, and that optimising either independently produces inferior results to joint optimisation (Trist and Bamforth, 1951). This provides the theoretical basis for the mediation hypothesis. Data governance and integrity maturity is precisely a social subsystem construct, comprising ownership, standards, stewardship, and procedural discipline. The prediction that analytics capability operates partly through governance maturity is a direct application of the joint optimisation principle.

3.2. Construct Definitions

Data Analytics Capability (DAC): The organisation's ability to collect, integrate, analyse, and act upon operational data, comprising data infrastructure and integration, analytical skills and personnel, analytical routines embedded in operational decision cycles, and process analytical technology deployment.

Process Automation Intensity (PAI): The extent to which manufacturing, packaging, laboratory, and transactional processes are executed by automated systems with reduced human intervention, encompassing process control automation, physical robotics, laboratory automation, and robotic process automation.

Intelligent Decision Support Adoption (IDS): The extent to which model-based systems, including advanced planning and scheduling, advisory process control, digital twins, and supply chain control towers, inform operational decisions. Data Governance and Integrity Maturity (DGM): The maturity of the organisation's arrangements for data ownership, quality standards, master data management, and adherence to ALCOA+ data integrity principles across computerised systems.

Absorptive Capacity (AC): The organisation's ability to recognise the value of new external information, assimilate it, and apply it to operational ends, operationalised through prior related knowledge, internal knowledge sharing, and cross-functional integration (Cohen and Levinthal, 1990; Zahra and George, 2002).

Operational Performance (OP): A higher-order construct comprising cost efficiency, quality performance, delivery performance, flexibility, and compliance performance, as defined in Section 2.6.

3.3. Hypotheses

H1. Data analytics capability is positively associated with operational performance.

Rationale. Analytics capability improves the informational basis of operational decisions across all five performance dimensions. It supports yield optimisation and waste reduction, deviation root cause identification and recurrence prevention, cycle time reduction through bottleneck identification and real-time release, changeover optimisation, and proactive detection of quality signals before they become inspection findings.

H2. Process automation intensity is positively associated with operational performance.

Rationale. Automation reduces process variability and human error, improves throughput consistency, and in aseptic operations directly reduces contamination risk. The effect is expected to be positive but smaller than that of analytics capability, because automation without data integration produces islands of local optimisation and because it introduces offsetting validation and computerised system maintenance burden.

H3. Intelligent decision support adoption is positively associated with operational performance. Rationale. Decision support systems improve decision quality and reduce decision latency at the points where operational performance is determined, including scheduling, process adjustment, deviation triage, and supply allocation. H4. Data analytics capability is positively associated with data governance and integrity maturity.

Rationale. Analytics deployment surfaces data quality deficiencies that were tolerable when data was used only for record-keeping, creating both the diagnostic visibility and the organisational motivation to formalise governance. This ordering is consistent with accounts in which classification, traceability, and storage-layer governance are prerequisites that analytical demand forces an organisation to confront. H5. Data governance and integrity maturity is positively associated with operational performance. Rationale. Governance maturity determines whether analytical outputs are trustworthy enough to act upon, and directly affects compliance performance through data integrity inspection exposure. The characterisation of interoperability, data quality, and governance as the backbone on which analytical capability rests (Ilodigwe and Adesemoye, 2021) supplies the theoretical form of this relationship.

H6. Data governance and integrity maturity mediates the relationship between data analytics capability and operational performance.

Rationale. Follows from H4 and H5 and from the sociotechnical joint optimisation principle. Partial rather than full mediation is expected, since analytics capability also acts directly.

H7. Absorptive capacity positively moderates the relationship between intelligent decision support adoption and operational performance.

Rationale. Decision support systems produce recommendations that require evaluation before adoption. Organisations with greater absorptive capacity can evaluate, contextualise, and act on those recommendations, while organisations lacking it will either ignore or blindly follow them, neither of which realises the intended benefit. Comparative evidence on organisational readiness for artificial intelligence integration in health operations supports this specification directly.

H8. Data analytics capability positively moderates the relationship between process automation intensity and operational performance.

Rationale. Automated equipment generates data as a by-product. Where analytics capability exists to exploit that data, automation contributes both direct process benefit and informational benefit. Where it does not, only the direct benefit is realised, and automation functions as an island.

3.4. Conceptual Model

The conceptual model specifies DAC, PAI, and IDS as exogenous constructs; DGM as a mediator on the DAC to OP path; AC as a moderator on the IDS to OP path; DAC as a moderator on the PAI to OP path; and OP as the endogenous higher-order construct. Site size, ownership type (innovator versus contract manufacturer), product modality (small molecule, biologic, mixed), and regional regulatory jurisdiction are specified as control variables. Table 1: Conceptual model Data Analytics Capability (DAC) H1 → Operational Performance (OP)

Cost efficiency Quality performance Delivery performance Flexibility Compliance performance ↓ H4 Data Governance and Integrity Maturity (DGM)

H5 → Process Automation Intensity (PAI) H2 → H8: DAC × PAI Intelligent Decision Support Adoption (IDS) H3 → H7: AC × IDS Controls: site size, ownership type, product modality, jurisdiction.

4. Research Methodology

4.1. Research Philosophy and Design

The study adopts a post-positivist philosophical position. It assumes that operational reality exists independently of the observer but is apprehended imperfectly through fallible measurement, which justifies the use of multi-item reflective constructs, formal validity assessment, and probabilistic inference rather than claims of direct observation. The design is explanatory, quantitative, and cross-sectional. It is explanatory because the objective is to test theoretically derived relationships rather than to generate theory. It is cross-sectional because the constructs were measured at a single point in the 2021 operating year. The cross-sectional choice carries a well-recognised cost in causal inference, addressed in Section 7.

4.2. Population, Sampling Frame, And Sample

The target population comprises operations, quality, engineering, technical services, and supply chain professionals employed at GMP-licensed pharmaceutical manufacturing sites, holding roles that provide direct visibility of both digital systems and operational outcomes at their site.

Sampling was purposive with a snowball component. The initial frame was constructed from professional association membership directories, industry conference delegate lists, and professional network searches filtered by employer, job title, and site location. Respondents were screened on three criteria: employment at a GMP-licensed manufacturing site, a minimum of two years in the current site, and a role providing operational visibility. Screening questions were embedded at the start of the instrument and failure at any criterion terminated the response.

Purposive sampling is appropriate here because the population is specialised, no probability sampling frame exists, and construct validity depends on respondent competence rather than on population representativeness alone. The limitation is that generalisation must be theoretical rather than statistical.

Minimum sample size was determined by two methods. The inverse square root method (Kock and Hadaya, 2018), applied to a minimum expected path coefficient of 0.15 at 5 percent significance and 80 percent power, indicates a requirement of approximately 155 cases. The ten times rule applied to the maximum number of structural paths directed at any single construct indicates a lower figure. A target of 200 was set to exceed both comfortably and to permit multi-group comparison.

The achieved sample was 214 usable responses from an initial 312 responses, after removal of 51 screen-outs, 32 incomplete responses exceeding the 15 percent missing data threshold, and 15 responses failing attention checks or exhibiting straight-line response patterns. The effective response rate against contacted eligible individuals was 31.2 percent. Responses originated from 68 distinct manufacturing sites, with a mean of 3.1 respondents per site.

4.3 Instrument Development

The instrument was developed in four stages. Stage one: item generation. Items were adapted from validated scales in the information systems and operations management literatures, then modified for pharmaceutical context. Analytics capability items were adapted from established firm analytics capability scales with the addition of PAT-specific indicators. Automation items were adapted from advanced manufacturing technology scales. Absorptive capacity items were adapted from the standard potential and realised absorptive capacity instrument (Zahra and George, 2002). Operational performance items were adapted from competitive priorities scales with the addition of the compliance dimension developed for this study from GMP inspection outcome categories. The known difficulty of attributing performance change to specific interventions in regulated service settings (Sanni, Ajiga and Atima, 2020a) informed the decision to measure standing capability rather than project-level outcomes.

Stage two: content validity. A panel of nine experts, comprising three qualified persons or equivalent quality leaders, three manufacturing operations directors, and three academic researchers in pharmaceutical engineering or operations management, assessed each item for relevance and clarity. The content validity ratio was computed for each item and items falling below the critical value for a nine-member panel were removed or rewritten. Eleven items were removed at this stage and seven were reworded.

Stage three: pilot testing. The revised instrument was administered to 34 respondents drawn from the target population but excluded from the main study. Cronbach's alpha was computed for each construct and item-total correlations were examined. Two items with corrected item-total correlations below 0.40 were removed. Stage four: final instrument. The final instrument comprises 43 reflective indicators across six constructs plus demographic and control items. All substantive items use a seven-point Likert scale anchored at 1 (strongly disagree) and 7 (strongly agree), with performance items anchored relative to the respondent's assessment of comparable sites. Construct coverage is summarised in Table 2.

Table 2: Construct measurement summary Construct Code Items Dimensions Primary source of adaptation Data Analytics Capability DAC 9 Infrastructure, skills, routines, PAT Firm analytics capability literature Process Automation Intensity PAI 7 Process, physical robotics, laboratory, RPA Advanced manufacturing technology literature Intelligent Decision Support Adoption IDS 7 Planning, advisory control, digital twin, control tower Decision support systems literature Data Governance and Integrity Maturity DGM 6 Ownership, standards, master data, ALCOA+ Data governance and GMP data integrity guidance Absorptive Capacity AC 6 Prior knowledge, assimilation, application Absorptive capacity literature Operational Performance OP 8 Cost, quality, delivery, flexibility, compliance Competitive priorities literature, extended

4.4. Data Collection Procedure

Data were collected during 2021 through a web-hosted questionnaire. Invitations included a participant information sheet describing purpose, voluntariness, anonymity arrangements, and data handling. Consent was recorded at the first screen. Two reminders were issued at two-week intervals. Median completion time was 14 minutes. No incentive was offered, on the reasoning that incentives risk attracting responses from individuals outside the competence criteria.

4.5. Common Method Bias

Because all variables were obtained from a single respondent through a single instrument, common method bias was addressed both procedurally and statistically. Procedural remedies followed established recommendations (Podsakoff et al., 2003) and included separating predictor and criterion items within the instrument, using different anchor formats for performance items, guaranteeing anonymity to reduce evaluation apprehension, randomising item order within blocks, and including two attention check items. Statistical assessment used Harman's single-factor test, in which the first unrotated factor accounted for 31.4 percent of variance, below the 50 percent threshold. Kock's full collinearity assessment was also applied (Kock, 2015), with all variance inflation factors from the full collinearity test below 3.3, indicating that the model is unlikely to be contaminated by common method bias.

4.6. Analytical Approach

Partial least squares structural equation modelling was selected over covariance-based SEM for four reasons. The model is complex, incorporating a higher-order construct, mediation, and two moderation terms. The study is predictive and exploratory in orientation, extending theory to a new context rather than confirming an established model. Several indicator distributions depart from normality. Sample size, while adequate, is modest for covariance-based estimation of a model of this complexity.

Analysis proceeded in the standard two-stage sequence recommended for variance-based structural equation modelling (Hair et al., 2019). The measurement model was assessed first for indicator reliability, internal consistency reliability, convergent validity, and discriminant validity. The structural model was assessed subsequently for collinearity, path coefficient magnitude and significance, coefficient of determination, effect size, and predictive relevance. Bootstrapping with 5,000 subsamples and no sign changes was used for significance testing. The higher-order operational performance construct was modelled using the two-stage approach with reflective-reflective specification. Mediation was assessed using the bootstrapped indirect effect with bias-corrected confidence intervals. Moderation was assessed using the two-stage approach with standardised interaction terms.

4.7. Ethical Considerations

The study involved no vulnerable populations, no clinical data, and no patient information. Informed consent was obtained electronically. Responses were anonymised at collection, with site identity captured only at the level of country and site size band. No proprietary process information was requested. Respondents were advised not to disclose confidential inspection outcomes and performance items were framed as relative self-assessments to avoid eliciting such disclosure. Data were stored encrypted with access restricted to the research team, and will be retained for five years before secure destruction.

5. Results

5.1. Sample Profile

Table 3: Respondent and site characteristics (n = 214) Characteristic Category n % Functional role Manufacturing or production 61 28.5 Quality assurance or quality control 54 25.2 Engineering or automation 38 17.8 Supply chain or planning 33 15.4 Technical services or MSAT 28 13.1 Seniority Operator or analyst 24 11.2 Supervisor or specialist 71 33.2 Manager 78 36.4 Director or above 41 19.2 Site size (headcount) Under 250 47 22.0 250 to 999 89 41.6 1,000 and above 78 36.4 Ownership type Innovator or originator 118 55.1 Contract manufacturer (CMO or CDMO) 71 33.2 Generic manufacturer 25 11.7 Product modality Small molecule only 92 43.0 Biologic only 51 23.8 Mixed portfolio 71 33.2 Jurisdiction North America 63 29.4 Western Europe 68 31.8 South Asia 47 22.0 West Africa 36 16.8 Tenure at site 2 to 5 years 82 38.3 6 to 10 years 76 35.5 Over 10 years 56 26.2 The sample is weighted toward supervisory and managerial respondents, which is appropriate given the visibility requirement, and spans four regulatory jurisdictions with differing inspection regimes, supporting the multi-group analysis reported in Section 5.6.

5.2. Descriptive Statistics

First, data governance and integrity maturity records the highest mean of the digital constructs. This is consistent with the regulatory environment: data integrity has been an enforcement priority for regulators since the middle of the previous decade, and firms have invested in governance under compliance pressure rather than under performance motivation. Governance maturity in this sector is therefore comparatively advanced relative to analytical exploitation of the governed data.

Second, intelligent decision support records the lowest mean and the highest dispersion. This is the least diffused technology class in the sample and the one on which sites differ most, which is consistent with an early adoption phase in which a minority of sites have moved substantially ahead. Third, automation intensity exceeds analytics capability, which is consistent with the sector's historical investment pattern of preferring capital equipment to data infrastructure. The gap between these two means is a compact statement of the integration problem described in Section 1.2.

5.3. Measurement Model Assessment

Table 6: Discriminant validity, heterotrait-monotrait ratio DAC PAI IDS DGM AC OP DAC PAI 0.612 IDS 0.694 0.581 DGM 0.703 0.489 0.556 AC 0.588 0.437 0.621 0.594 OP 0.671 0.548 0.647 0.639 0.602 All HTMT ratios fall below the conservative 0.85 threshold, establishing discriminant validity. The highest ratio, 0.703 between DAC and DGM, is theoretically expected given the hypothesised relationship between them and remains well within acceptable bounds. All HTMT ratios were additionally assessed against the bootstrapped confidence interval criterion (Henseler, Ringle and Sarstedt, 2015). Fornell-Larcker criterion assessment (Fornell and Larcker, 1981) and cross-loading examination, not tabulated here, produced consistent conclusions.

5.4. Structural Model Assessment

Collinearity was assessed before path estimation. All inner model variance inflation factors fall between 1.31 and 2.47, below the conservative 3.0 threshold, indicating no problematic collinearity among predictor constructs.

5.5. Mediation Analysis

Table 8: Mediation of the DAC to OP relationship through DGM Effect Coefficient SD t p 95% BC CI Conclusion Direct effect DAC → OP 0.284 0.071 4.00 < 0.001 [0.146, 0.423] Significant Indirect effect DAC → DGM → OP 0.124 0.047 2.64 0.008 [0.038, 0.221] Significant Total effect DAC → OP 0.408 0.068 6.00 < 0.001 [0.274, 0.541] Significant Variance accounted for (VAF) 30.4% Partial mediation Both direct and indirect effects are significant and of the same sign, indicating complementary partial mediation. The variance accounted for figure of 30.4 percent falls within the conventional range for partial mediation. H6 is supported. The substantive reading is that roughly three tenths of the operational benefit of analytics capability travels through improvements in data governance and integrity maturity, while the remaining seven tenths acts directly. Governance is therefore a material transmission mechanism but not the sole one, and analytics investment retains direct value even where governance is unchanged.

5.6. Moderation Analysis and Multi-Group Comparison

The interaction of absorptive capacity with intelligent decision support adoption is positive and significant (β = 0.132, p = 0.021). Simple slope analysis indicates that at one standard deviation above the mean of absorptive capacity the IDS to OP slope is 0.373, at the mean it is 0.241, and at one standard deviation below the mean it is 0.109 and not statistically distinguishable from zero (p = 0.164). The practical implication is direct: in organisations with below-average absorptive capacity, decision support adoption showed no detectable operational benefit in this sample. The interaction of analytics capability with automation intensity is likewise positive and significant (β = 0.118, p = 0.032). At high analytics capability the PAI to OP slope is 0.314; at low analytics capability it is 0.078 and not significant (p = 0.238). Automation without analytics capability shows no detectable independent benefit in this sample.

Multi-group analysis by ownership type, conducted after establishing partial measurement invariance through the MICOM procedure, found one significant difference between innovator and contract manufacturing sites. The IDS to OP path is significantly stronger for contract manufacturers (β = 0.318) than for innovator sites (β = 0.201, difference p = 0.038), which is plausibly attributable to the greater scheduling complexity and higher product mix variability characteristic of contract operations, conditions under which planning and scheduling decision support has more to optimise.

5.7. Summary of Hypothesis Testing

Table 9: Hypothesis testing summary Hypothesis Statement Outcome H1 DAC positively associated with OP Supported H2 PAI positively associated with OP Supported H3 IDS positively associated with OP Supported H4 DAC positively associated with DGM Supported H5 DGM positively associated with OP Supported H6 DGM mediates DAC to OP Supported (partial, VAF 30.4%) H7 AC moderates IDS to OP Supported H8 DAC moderates PAI to OP Supported

6. Discussion

6.1. The Primacy of Analytics Capability

The strongest direct predictor of operational performance in this model is data analytics capability, not automation intensity. This inverts the sector's historical investment ordering, in which capital equipment has consistently attracted funding more readily than data infrastructure, and it corresponds to the gap visible in the descriptive statistics where automation intensity (mean 4.71) exceeds analytics capability (mean 4.38).

The interpretation offered by dynamic capabilities theory is that automation is a purchasable asset while analytics capability is a socially complex organisational competence. Two sites may install identical process control systems, but the site that has developed the routines to interrogate the resulting data, the personnel able to conduct that interrogation, and the decision rights permitting action on the findings, extracts value that the other does not. The asset is imitable; the capability is not.

This finding aligns with the argument advanced by Arden and colleagues that the defining characteristic of the pharmaceutical smart factory is the data-rich environment created by capture and storage technologies, upon which every higher-order capability depends (Arden et al., 2021). Equipment that operates without contributing to that environment does not participate in the architecture that generates the benefit.

6.2. Automation as A Conditional Rather Than an

Unconditional Benefit H2 is supported but with the smallest effect size of the three direct paths, and H8 explains why. The moderation analysis shows that at low analytics capability the automation to performance slope is statistically indistinguishable from zero. Automation, in this sample, does not deliver measurable operational benefit on its own.

This is the islands of automation problem stated quantitatively. Automated equipment generates two forms of value: direct process value through reduced variability and labour, and informational value through the data it produces.

The direct value is partially offset in the pharmaceutical setting by the validation and computerised system lifecycle burden that automation imposes, as discussed in Section 2.4. The informational value is realised only where analytics capability exists to exploit it. Where that capability is absent, the offsetting burden consumes much of the direct benefit and the informational benefit is forgone entirely. The practical consequence is a sequencing rule. Automation investment appraised in isolation will systematically overstate its return unless the appraisal accounts for the analytics capability that must accompany it. This is the conclusion reached in maintenance research, where the move from reactive to predictive regimes depends on instrumentation and analytical interpretation rather than on the equipment itself (Sunday et al., 2020), and it holds equally in supply operations, where materials systems deliver uptime benefit only when the readiness data they hold is actually interrogated (Okonkwo et al., 2021a).

6.3. Decision Support and The Absorptive Capacity

Threshold The moderation result for H7 is the study's most operationally pointed finding. At below-average absorptive capacity, intelligent decision support adoption showed no significant relationship with operational performance. The mechanism is straightforward once stated. A decision support system produces recommendations. A recommendation creates value only if the receiving organisation can evaluate whether to accept it. An organisation lacking the prior knowledge to assess a scheduling optimisation, the internal knowledge sharing to circulate what was learned when a recommendation proved wrong, and the cross-functional integration to implement a recommendation that spans production and quality, will do one of two things. It will ignore the system, in which case the investment is dead capital. Or it will follow the system uncritically, in which case it has substituted an unexamined algorithmic judgement for an unexamined human one and has additionally created a regulatory exposure, since the operator remains documented as the responsible decision maker. This has particular weight in the pharmaceutical context because of the accountability structure discussed in Section

2.5. Regulatory frameworks assign responsibility to

identified persons. A system that produces recommendations the responsible person cannot evaluate does not merely fail to add value; it creates a gap between documented and actual accountability that would be difficult to defend under inspection. Explainability is therefore a condition of regulatory defensibility rather than a presentational refinement. The organisational reading of the same result, that capable systems fail where the receiving organisation was never prepared for them, matches the pattern reported in other sectors, where delivery capability rather than technical merit governs realised benefit (Ilodigwe and Adesemoye, 2019b; Michael and Ogunsola, 2021e).

6.4. Governance as The Transmission Mechanism

The mediation result establishes data governance and integrity maturity as a partial transmission mechanism carrying approximately 30 percent of the analytics effect. Two readings follow.

The first concerns trust. Analytical outputs are acted upon only if decision makers believe the underlying data. In an environment where master data is inconsistent across systems, where the same material carries different identifiers in the enterprise resource planning system and the manufacturing execution system, and where audit trails are incomplete, analytical findings are contestable and contested findings do not change behaviour. Governance maturity converts analytical output into actionable evidence. The second concerns compliance directly. Data integrity has been among the most consistent sources of regulatory observation and enforcement action internationally over the past decade. Because compliance performance is a component of the operational performance construct, governance maturity improves measured performance through this channel independently of any analytical benefit. The mechanisms are those identified in Section 2.9: ownership and classification determining what obligation attaches to which records (Mbonu et al., 2019a), access governance supporting the attribution requirement (Mbonu et al., 2020b), and control assurance corresponding to periodic audit trail review (Akomolafe and Agu, 2018). The descriptive finding that governance maturity records the highest mean among the digital constructs suggests a sector that has built governance foundations under regulatory compulsion without fully exploiting them for operational purposes. That is an unusual and favourable position: the expensive prerequisite is largely in place and the analytical superstructure is what remains to be built.

6.5. Theoretical Implications

The study makes three theoretical contributions. It extends dynamic capabilities theory into a regulated operating context by demonstrating that the capability-asset distinction holds with unusual force where regulatory constraints limit the design space of permissible implementations. Where every firm faces the same technical constraints, differentiation must come from organisational capability rather than technical choice. It specifies a transmission mechanism between digital investment and operational outcome, addressing the under-theorisation identified in Section 2.9. The finding that governance mediates the analytics effect is a concrete application of sociotechnical joint optimisation and suggests that similar mediation structures may operate in other regulated industries.

It demonstrates the analytical cost of aggregating digital technology into a single construct. Had DAC, PAI, and IDS been combined, the aggregate would have shown a strong positive association with operational performance and would have concealed the two conditional relationships that constitute the study's most actionable findings.

6.6. Managerial Implications

Six implications follow for pharmaceutical operations leadership.

Sequence data foundations ahead of advanced applications: The mediation result and the descriptive pattern together suggest that analytics investment layered onto weak governance underperforms. Firms with immature master data and incomplete audit trails should expect poor returns from advanced analytics until those foundations are addressed. Integrated frameworks treating procedure design, automation, compliance, and change management as one programme rather than as sequential projects offer a practical template (Ambali et al., 2021).

Appraise automation with its analytics complement: Business cases for automation should include the analytics capability required to exploit the resulting data, and should not claim the informational benefit unless that capability is funded.

Assess absorptive capacity before decision support investment: Organisations should establish whether they possess the analytical skills, knowledge sharing routines, and cross-functional integration to evaluate machine-generated recommendations before deploying systems that produce them. Where they do not, capability building should precede system procurement.

Treat compliance performance as a digital benefit, not a digital constraint: Framing digital investment as improving quality outcomes and inspection readiness, rather than as a compliance risk requiring mitigation, changes both the internal business case and the regulatory conversation. Distinguish digitisation from transformation in the investment portfolio: Projects that replicate existing workflows in electronic form should be appraised on documentation and audit trail benefits alone, and should not be credited with operational transformation benefits they will not deliver.

Contract manufacturers should prioritise decision support: The multi-group finding that the decision support to performance path is stronger for contract manufacturers suggests that planning and scheduling optimisation carries higher marginal value where product mix and scheduling complexity are greatest.

6.7. Regulatory and Policy Implications

The results support continued movement toward performance-based regulatory approaches. If digitally mature sites demonstrate superior quality and compliance outcomes, oversight calibrated to demonstrated performance rather than to uniform procedural prescription would allocate regulatory attention more efficiently and would create an incentive for digital maturity that current approaches do not. Two specific areas warrant regulatory attention. The validation of adaptive systems remains unresolved: frameworks designed to demonstrate and maintain a validated state sit awkwardly with machine learning models whose behaviour changes with continued training, and the absence of settled guidance is a material adoption barrier. The parallel question of securing converged information and operational technology estates, addressed at length in energy sector regulation (Adegbite, Adebayo and Ahmed, 2020), has no equivalent settled treatment in GMP guidance. A third area concerns proportionality. Validation frameworks were designed around process control systems whose failure directly affects product, yet they are applied with comparable weight to low-risk analytical applications whose output informs a human decision. The resulting burden discourages exactly the exploratory analytical work that would build the capability this study finds most consequential. Separately, the accountability structure for machine-generated recommendations requires clarification, since current frameworks assign responsibility to identified persons without addressing the case in which the person cannot independently evaluate the recommendation they are documented as accepting.

For emerging pharmaceutical manufacturing economies, including those represented in the West African and South Asian portions of the sample, the finding that governance maturity mediates analytics benefit has direct policy relevance. Investment in data governance capability, being less capital intensive than equipment automation, may offer a more efficient path to operational and compliance improvement than equipment-led modernisation programmes. This is consistent with findings on infrastructure-driven expansion of diagnostic capacity in resource-constrained health systems, where organisational and design decisions determined realised capability more than capital intensity did (Aminu-Ibrahim, Ogbete and Ambali, 2020), and with evidence from other production sectors in emerging economies that data-driven models improve efficiency and resource utilisation without requiring frontier capital equipment (Michael and Ogunsola, 2021c).

7. Limitations and Directions for Future Research

7.1. Limitations

Cross-sectional design: The study observes association at a single point in time and cannot establish causal direction. It is possible that high-performing sites acquire digital capability rather than digital capability producing high performance, and reciprocal causation is plausible. The theoretical argument and the mediation structure are consistent with the proposed direction but do not establish it. Perceptual measurement: Operational performance is measured through informed self-assessment rather than through audited data. Respondents may over-report performance, and respondents at digitally mature sites may hold different reference standards. The relative anchoring of performance items mitigates but does not eliminate this concern.

Single respondent method variance: Although procedural and statistical remedies were applied and results indicate that common method bias is unlikely to be substantive, single-source designs remain vulnerable.

Purposive sampling: Generalisation is theoretical rather than statistical. The sample may over-represent respondents engaged with digital initiatives, since such individuals are more likely to be reachable through the frames used and more likely to respond.

Period specificity: The 2021 collection period was atypical. Pandemic conditions placed unusual demands on operational responsiveness and may have amplified the observed associations relative to steady-state conditions. Construct scope: Cybersecurity maturity, workforce digital skills, organisational culture, and leadership commitment are excluded from the model despite plausible relevance, in the interest of model parsimony given sample size.

7.2. Future Research Directions

Longitudinal designs tracking digital maturity and performance across multiple periods would address the causal inference limitation and would permit estimation of the lag structure between investment and realised benefit, a question of considerable practical interest that cross-sectional work cannot address.

Objective performance measurement using audited operational and regulatory data, including inspection observation counts, batch rejection rates, and right-first-time figures drawn from site records, would substantially strengthen the outcome measurement.

Multi-informant designs collecting technology measures from engineering respondents and performance measures from quality or finance respondents would eliminate single-source variance concerns.

Cross-organisational model development presents a further opportunity. Analytical approaches that train models across sites and across firms without centralising commercially sensitive process data would permit the pooled analysis that competitive sensitivity currently prevents, and would be of particular value in innovator and contract manufacturer relationships. A parallel line of enquiry concerns whether digital twin maturity predicts regulatory as well as operational outcomes. Interpretability is likely to be the binding constraint on any such work, since models whose reasoning cannot be reconstructed sit poorly with validation requirements, and approaches developed for interpretable failure prediction in distributed assets (Komi and Adamolekun, 2021) indicate one route through that difficulty. Standardisation offers a further direction. Evidence that standardised infrastructure and operational design models govern realised capability in regulated facility networks (Aminu-Ibrahim, Ogbete and Ambali, 2018) raises the question of whether digital process definitions standardised across a manufacturing network produce measurable transfer and scale-up advantage, which the 2021 vaccine experience suggests but does not establish.

Comparative sectoral work testing the same model in medical devices, food manufacturing, and aerospace would establish whether the governance mediation mechanism is specific to pharmaceutical regulation or generalises to regulated manufacturing broadly.

Qualitative case study work examining sites that invested substantially in digital technology without realising performance benefit would illuminate the failure mechanisms that this study identifies statistically but does not describe in process terms.

Finally, replication in subsequent periods, and particularly work incorporating advanced machine learning applications in deviation investigation, batch record review, and regulatory documentation, would test whether the relationships identified here hold as the technology landscape changes.

8. Conclusion

This study set out to determine how data analytics, automation, and intelligent decision support relate to operational performance in pharmaceutical manufacturing, and through what organisational mechanisms. The evidence supports four conclusions. Data analytics capability is the strongest direct contributor to operational performance among the three technology classes examined, exceeding automation intensity despite the sector's historically greater investment in the latter. Automation intensity contributes positively but conditionally, showing no detectable independent benefit where analytics capability is absent, which is the islands of automation problem expressed as a moderation effect. Intelligent decision support contributes positively but likewise conditionally, showing no detectable benefit where organisational absorptive capacity is below average, because recommendations create value only for organisations able to evaluate them. Data governance and integrity maturity functions as a partial transmission mechanism, carrying approximately 30 percent of the analytics effect and reinforcing the sociotechnical principle that technical and social subsystems must be optimised jointly.

The unifying implication is that digital transformation in pharmaceutical operations is not a technology procurement problem. Every firm in this sample could purchase the same systems. What differentiated performance was the organisational capability to exploit them: the governance to trust the data, the analytical competence to interrogate it, and the absorptive capacity to act on what it showed. The pharmaceutical sector enters this transition holding an underappreciated advantage, in that regulatory pressure has already compelled investment in the data governance foundations that other sectors must build from scratch. The task that remains is to exploit foundations already paid for.