Europe’s federated democratic learning state
Thiemo Fetzer is Professor of Economics at Universität Bonn; Theme Leader for Globalisation and Global Crises at the Centre for Competitive Advantage in the Global Economy (CAGE); and Professor of Economics at the University of Warwick
Innovations like electronic payments and e-invoicing are making it much easier to trace economic activity (OECD 2024). At the same time, AI-assisted knowledge retrieval is increasingly substituting humans in a broad array of information intermediation tasks. Taken together, these developments can shift the informational boundaries of the state: what the state can observe, the knowledge it can derive, and how it may deploy that knowledge in the management of the economy.
The US and China have responded to the digitisation of economies and societies in radically different ways. In the American model, private players intermediate the economy’s data trace, causing an extreme concentration of power. Payment service providers, for example, benefit from large network externalities, steady intermediation fees from cash flows, and information rents that accrue outside the ordinary social contract due to widespread profit shifting. The same applies to search, compute, advertising, platform work, and digital identity.
In the Chinese model, the state retains much more of that capacity directly, but at the cost of political hierarchy and coercion. Europe needs digital state capacity without that societal form, and digital sovereignty without platform dependence. By 2050, Europe should have built an alternative model: a federated democratic learning state that leans on public-data-derived evolving knowledge graphs and AI as an accountability and coordination layer with built-in privacy.
Digitisation of tax systems to build a new informational architecture
The unsung hero through which this may become concrete is the electronic VAT invoice. Europe and much of the G20 – except the US – is moving towards invoice-level VAT. This means every firm-to-firm and firm-to-consumer economic transaction becomes technically legible to the tax system in near real time. If every invoice contains structured or semi-structured line items, AI-assisted processing can make the tax system more than a compliance and revenue-raising instrument.
Such a system can become a protected intelligence map of the economy: who trades what with whom, which capabilities exist, where, and who produces what value added. A ‘VAT firm-to-firm transaction network’ represents a knowledge graph that measures the granular economic ties between essentially all economic units and agents, providing knowledge of what firms can produce through linking the observed outputs and measured inputs with a production-network view.
Granular network measurement is key to understanding aggregate resilience and the propagation of economic shocks (Acemoglu et al 2012). AI is already helping shape even more granular production-network representations of the tangible economy (Fetzer et al 2024b).
Putting this public-stewarded VAT-derived knowledge graph, combined with AI, to work can fundamentally reshape the role of the state and hence, redefine the boundaries of the firm. For Europe, a unique opportunity emerges to evolve the Single Market to become an intelligence and orchestration platform of a modern state that leans on real-time, high-authority administrative data.
AI is integral to ensure that underlying data can be operationalised in a way that is compatible with strict privacy, to protect sensitive data, and to maintain a market-based economic organisation. It can ensure that Europe moves to being governed by (trusted) statistics rather than stories. Present discussions on the future of the EU, formulated by landmark reports by Draghi (2024) and Letta (2024), have not emphasised the importance of the information architecture that may make these agendas administratively feasible.
I sketch one application in some detail: public procurement. Similar arguments can be made for innovation policy, technology transfer, and many other areas of public good provision.
For Europe, a unique opportunity emerges to evolve the Single Market to become an intelligence and orchestration platform of a modern state that leans on real-time, high-authority administrative data
Procurement as an example to incorporate knowledge graphs
Public procurement is often slow and inefficient. Bureaucracy, process, and due diligence are costly by-products of past corruption episodes. Even where there is no favouritism, its mere possibility can generate stories, litigation risk, and cycles of distrust.
The administrative state’s response to any (perceived) scandal is more bureaucracy, raising participation costs, with the side effect of skewing public contracts to larger firms that have the long breadth alongside the required legal and administrative capacity.
A knowledge graph of the physical and material world, combined with an interoperable and digitally legible administrative state, consisting of company registries, and a capability graph anchored on production network knowledge fed with granular invoice-level VAT data, can vastly increase public procurement effectiveness. They can make the flow of physical goods through the economy visible in a way that Bill Phillips could only have dreamed of when designing his hydraulic computer of macroeconomic flows.
The process could look as follows: if the state wants to buy a good or service, it can use protected VAT-derived knowledge to suggest and invite a set of potential suppliers that – based on their economic classification, the value added, the observed inputs, and outputs – may possess the capability to provide that good or service.
AI-assisted retrieval is essential as it ensures that no human official can browse the underlying sensitive raw data that is built on production-network knowledge and VAT-transaction data. AI doesn’t magically guarantee neutrality; rather, it makes the process auditable, query-bound, and outsize influence-blind in a way that can be regularly and automatically audited.
This changes the role of public authority. Industrial policy, for example, as often implemented through procurement, becomes less a matter of discretionary selection and more a process of structured self-learning and discovery based on broad, rule-based access across a range of tournaments. This reduces the risk of rentierism through ‘revolving doors’ and investments in proximity to power.
Once capability-eligible suppliers have been identified, access to bidding opportunities can be opened for a (stratified) random sample. Firms that are not registered and licensed, use coercive labour market practices, or infract in other ways on the social contract could be excluded from participation, creating an indirect demand for formalisation. It may also indirectly help to level the playing field.
By incorporating randomisation, the political economy of state demand changes. Firms can spend less time lobbying for contracts and focus more on delivery. Business development costs decrease. Further, if firms enter an auditable stochastic process, then not being selected becomes closer to an idiosyncratic risk, against which firms may, in equilibrium, find it cheaper to borrow, compared to a world where access to finance is also politicised.
In such a world, the policymaker’s role shifts from managing to overseeing a near automated process. Procurement outcomes are shaped through sample stratification by factors such as region, firm size, labour standards, employment-structure, or firm-age. Each of these dimensions may reflect political decisions, but corruptive self-selection has been neutralised.
In essence, the policymaker’s role is to develop the prompt for an AI agent to identify candidate suppliers from the publicly stewarded knowledge graph. Stochastic fairness can be guaranteed and audited, and – thanks to the steady flow of data – the underlying data evolves with the economy and, as a result, so does the knowledge graph.
Privacy, democratic accountability, and trust
A European learning state cannot be built on raw-data exposure but must guarantee privacy. Citizens and firms will only support a pivot to extensive, public-stewarded data if they know that data shared with public bodies will only be used for lawful, bounded, transparent purposes with efficiently challenging processes. This is the core concern in work on the informational boundaries of the state (Fetzer et al 2024a).
Public institutions should be able to fully utilise the informational resources they acquire, computing more while seeing less. A robust public data architecture paired with AI for trusted private knowledge retrieval is precisely the combination that can move societies towards an equilibrium with maximal knowledge sharing, minimal data exposure, and, crucially, without running the risk of turning into a Chinese- or American-style equilibrium.
Having transparency and auditability over how underlying data are being used is one pillar that can ensure that economic agents supply accurate high-quality data. The VAT has a built-in monitoring logic because transactions are relational. A reported sale is someone else’s reported purchase. Dual reporting makes manipulation more costly, as false reporting requires coordination.
The threat of (automatic) exclusion, for instance from access to public demand, can become an additional disciplining device. None of this removes the need for audit, but it means the system is not based simply on self-reported claims; it is a network of mutually constraining records. Naturally, audit costs can be significantly lower in such a setting.
The development of a modern learning state can improve public accountability and repair democratic trust in a world of perceived polarisation. As rewards to proximity to government decline, the returns to lobbying fall. This can recalibrate the firm-size distribution by improving small and medium-sized companies’ access to public sector demand.
Faster procurement can improve performative state capacity. Auditable, human-not-in-the-loop retrieval processes can deprive malign actors of narrative ammunition around non-transparent allocation. Fewer resources are wasted on pork-barrel politics and defensive compliance.
Most importantly, policymaking becomes more future-oriented as the short-term returns to using procurement to divert public resources decline. Political competition can then move towards the battle over ideas: what visions for individual and collective wellbeing should Europe build?
References
Acemoglu, D, VM Carvalho, A Ozdaglar, and A Tahbaz-Salehi (2012), “The network origins of aggregate fluctuations”, Econometrica 80(5): 1977–2016.
Draghi, M (2024), The future of European competitiveness, European Commission.
Fetzer, T, C Shaw, and J Edenhofer (2024a), “Informational boundaries of the state”, CEPR Discussion Paper 18773.
Fetzer, T, PJ Lambert, B Feld, and P Garg (2024b), “AI-generated production networks: Measurement and applications to global trade”, CEPR Discussion Paper 19708.
Letta, E (2024), Much more than a market.
OECD (2024), “Digital public infrastructure for digital governments”, OECD Public Governance Policy Papers No. 68.
Author’s note: I would like to thank Olivier Blanchard, Jakob Schneebacher and Jacob Edenhofer who provided helpful comments that significantly helped streamline the exposition. This article was originally published on VoxEU.org.
