New Value-Chain Framework Reveals Where AI Creates and Destroys Public Value

Since ChatGPT burst into public consciousness in late 2022, governments around the world have raced to invest in generative artificial intelligence, convinced that systems capable of simulating human thought will transform the delivery of public services. A new open-access study published in the journal Global Public Policy and Governance argues that this rush risks repeating the mistakes of previous technology hypes unless governments adopt a more disciplined way of thinking about where, exactly, value is created and lost when AI enters the machinery of the state. Researchers Karl Löfgren of Victoria University of Wellington and C. William R. Webster of the University of Stirling propose that a value-chain approach, borrowed from commercial production analysis and adapted to public service logic, can serve as a diagnostic roadmap for understanding how AI both enhances and threatens public value.

The core insight of the study is deceptively simple. Rather than treating AI as a monolithic technology that is either a savior or a menace, the authors break the deployment of AI in government into a series of interlinked stages: data collection, data storage, algorithm training, and data analysis and usage. At each link in this chain, different values are created, different actors derive benefit, and different problems emerge. Crucially, the authors insist that the AI value chain cannot be represented as a simple linear sequence. Because generative AI systems continuously recycle their outputs back into the chain, feeding new data into model calibration in an ongoing learning process, the appropriate mental model is cyclical rather than linear. This cyclical structure, they argue, is what fundamentally distinguishes generative AI from earlier waves of digital government technology, from voice recognition systems to automated decision-support tools, which have been part of the digital government trajectory since the 1950s.

The analysis begins where every AI system begins: with data. The authors identify four distinct categories of data feeding public sector AI. Historical administrative datasets, such as census records, tax files, welfare benefits and health data, are compulsory, regularly updated and broadly accurate, making them reliable building blocks. Voluntarily shared data, from library card registrations to social media activity, is far more uneven. Data surrendered by non-public agents, including banks, insurers and private contractors delivering day-care or social services, introduces yet another set of provenance problems. Finally, governments increasingly procure data collected by private actors, such as location data from telecom providers used for transport demand modelling or pandemic mobility tracing. Each category carries different expectations of quality, and the latter two are far more prone to duplicates, gaps, inconsistencies and outright errors, which propagate silently into downstream algorithmic decisions.

Bias is the shadow that falls across this first stage of the chain. The study catalogues biases of measurement, representation, aggregation and sampling, many rooted in the digital divide itself. Older citizens are frequently absent from digital datasets, yet the authors caution that younger users are not reliably representative either, lacking the skills to be fully digital natives in the way popular narratives suggest. Socio-economic inequalities in device access persist even amid the proliferation of mobile phones, and commercial data collectors systematically ignore less prosperous groups because they are unattractive customers, precisely the populations public policymakers most need to understand. The authors also warn against the seductive aggregation fallacy, the belief that with enough data the numbers will speak for themselves. Data never speaks for itself, they note; someone always formulates the questions, organizes the material and interprets the results, and correlation is not causation. Big data, they argue, must be supplemented with small data and basic social science.

Privacy challenges emerge at the very start of the chain and, according to the study, never let go. Vast quantities of data are routinely collected without informed consent, despite the GDPR providing a global minimum standard. The authors argue that the contemporary privacy paradigm of notice-and-consent, in which individuals manage their own data permissions, is simply unattainable in an AI context given the multitude of sensory devices, data sources and indefinite possibilities for future repurposing. Once collected, data creates risks of breach, re-identification and profiling, a problem long observed with video surveillance and now magnified by the Internet of Things. Anonymization offers false comfort: removing names does little when residency, workplace, age and shopping habits are sufficient to identify an individual.

The storage stage brings its own hazards. Larger data volumes and reliance on remote cloud storage amplify vulnerabilities to tampering, malicious insiders, data loss and large-scale breaches, and the question of how to isolate and protect sensitive personal information within heterogeneous datasets remains one of the hardest problems in the field. The authors also probe the blurred boundary between public and private responsibility. When public service data is stored or processed by social media companies, data brokers, cloud providers or telecoms firms, does the public agency renounce responsibility for future breaches? And can private organizations whose datasets emerged from public service sources claim copyright, commercialize that data and sell it back to the state, a question that intersects uneasily with the global open government data movement?

Training the algorithm exposes what the authors call one of the key challenges of AI in the public sector: transparency. Computer and data scientists retain the privilege of formulating the purpose of any data process, and algorithms inevitably embed subtle institutional and historical biases reflecting the context of their designers. The mushrooming literature on algorithmic bias offers sobering empirical evidence, from racial bias in facial recognition software used in policing, to predictive policing tools that disproportionately target the usual suspects, to risk-of-reoffending assessments that reproduce discriminatory patterns. At the analysis and usage stage, the study highlights the persistent difficulty of combining disparate datasets, a core premise of government AI that frequently exceeds existing data integration technology, complicated by technical, organizational, semantic and legal incompatibilities, along with unresolved intellectual property questions about who owns learning-based datasets built on public information.

Accountability emerges as perhaps the most consequential value at the end of the chain. In manual government processes, decision-makers are expected to explain and justify their decisions within codified chains of command. When generative AI replaces or shields the human decision-maker, the authors ask, why was this decision made, and who is to blame when a machine decides wrongly, particularly where human oversight is weak or absent? Accountability, they stress, means being called to account, something that cannot be delegated to software. Documented failures cluster in social welfare interventions and justice-system profiling, with recurring complaints of opaque reasoning, transferred human bias, and responsibility floating between developers, suppliers, administrative users and political masters. The authors invoke cautionary precedents such as Australia’s Robodebt scheme and the British Post Office Horizon scandal as evidence of what follows when automated systems fail without clear lines of answerability.

The study’s conclusion is neither a rejection of AI nor an endorsement of it. The challenges identified, the authors write, point instead to the need for governance structures and risk assurance frameworks capable of managing AI in the public interest within a highly fragmented space where public and private actors are intimately meshed. They observe that existing regulation, from data protection law to intellectual property rights and impact assessments, leaves significant gaps around responsibility and accountability, and that current policy responses remain embryonic and dominated by soft instruments. Overemphasizing the technical side of generative AI, they warn, distracts from problems embedded in existing structures and processes of government. It took electronic government roughly a decade to abandon utopian rhetoric for a sober discussion of online service delivery, and the authors argue the AI discourse must make the same transition: understanding the technology in its institutional environment, alongside its impacts and consequences, if individual and societal value is genuinely to be realized.

Beyond its diagnostic framework, the study situates the current wave of adoption within a longer history of technological enthusiasm in government, noting that artificial intelligence has become a symbol for a more productive, accurate and resource-efficient public administration much as big data, smart cities and blockchain did before it. The authors point out that there is still no consensus on what actually constitutes AI, and they adopt the OECD definition of a machine-based system that infers from received inputs how to generate predictions, content, recommendations or decisions, with varying levels of autonomy and adaptiveness after deployment.

A notable addition to the familiar narrative is the authors’ discussion of agentic AI, in which systems move beyond generating content toward making multi-step decisions and reasoning autonomously with other agents. This transition, they suggest, raises the stakes for public sector deployment because machine sense-making begins to replace not only acquired skills but interpretive judgment itself. They balance this against expected benefits drawn from the literature, including more accurate and less biased decisions, cost savings, reduced corruption and improved policy analysis, while stressing that past automated failures, such as flawed welfare debt recovery systems, offer predictable lessons.

The article also emphasizes that governance responses to date have relied largely on soft instruments, such as ethical principles, awareness-raising and assurance frameworks, in the expectation that industry will develop technical standards voluntarily. Meanwhile, cloud-based networked practices erode the capacity of national regulators to supervise AI effectively. The authors repeatedly caution that roll-out is deeply context-specific and shaped by norms surrounding public services, warning that the technology should not be divorced from the institutional environments in which it operates. Their proposed value-chain diagnostics, covering data quality, intellectual property, surveillance, privacy and transparency, are intended as tools for forecasting and evaluating generative AI systems more systematically than the prevailing promises-versus-pitfalls debate allows.

Subject of Research: A value-chain analysis of the opportunities and risks of artificial intelligence in public service delivery and policymaking.

Article Title: A value-chain perspective of artificial intelligence in public services

Article References: Löfgren, K., & Webster, C. W. R. (2026). A value-chain perspective of artificial intelligence in public services. Global Public Policy and Governance. https://doi.org/10.1007/s43508-026-00152-0

Image Credits: AI Generated

DOI: 10.1007/s43508-026-00152-0

Keywords: artificial intelligence, public services, value chain, public value, generative AI, algorithmic bias, data privacy, accountability, digital government, data governance, surveillance, transparency