Artificial Intelligence in Support of National Land Administration and Build-Back-Better Policies: A Technical and Policy Assessment of the Hellenic Cadastre and the Cross-Sectoral Reuse of Geospatial Infrastructure (HEPOS)

Abstract

In April 2024, the Hellenic Cadastre became one of Europe’s first land registries to use a generative AI model (a large language model served through Azure OpenAI) for the legal review of property deeds. Unlike similar European initiatives using classical NLP, Greece applied state-of-the-art generative AI to a massive legacy issue: 390 historical mortgage registries holding an estimated 600 million to one billion paper pages. By April 2026, the system had processed 310,000 acts, reducing the average per-act review time from about thirty minutes to under ten; a very large per-act cost reduction is also reported by the implementation partner, which we treat as a vendor-stated figure. Additionally, the cadastre’s geodetic infrastructure found a second use following the 2023 Tempi rail disaster. In 2026, the Hellenic Positioning System (HEPOS), a 98-station GNSS reference network, began providing corrections for Greece’s real-time train tracking platform. While satellite-based train positioning is not novel in Europe, where consortia such as CLUG have run a decade of research and pilots, this marks its operational deployment in Greece. The Greek case is unique institutionally rather than technically: it repurposed a national CORS network for a citizen-facing train tracking platform as a short-term crisis response, alongside an incomplete ETCS rollout. This paper documents both deployments, measures their impact, maps them onto the nine FELA pathways, and identifies transferable practices. Greece is not presented as a technological frontier, but as an example of how a country can put existing geospatial infrastructure and AI to rapid use in delivering build-back-better policies for the public, in line with the UN 2030 Agenda.

1. Introduction

Secure, transparent and accessible land administration is a precondition for sustainable development. The 2030 Agenda for Sustainable Development explicitly references tenure security under Goals 1 (poverty), 2 (food security), 5 (gender equality), 11 (sustainable cities) and 15 (life on land) [1]. UNECE has tracked the evolution of land administration in the region for two decades; comparing its 2005 baseline review with the 2025 ecosystems-and-security study shows a discipline that re-tools itself in step with societal needs and progressively absorbs new technology [2,3]. In August 2020 the Framework for Effective Land Administration (FELA) was endorsed for the land sector. It operationalises the Integrated Geospatial Information Framework (IGIF) and identifies nine strategic pathways: governance, institutions and accountability; policy and legal; financial; data; innovation; standards; partnerships; capacity and education; and advocacy and awareness [4]. Within the European context, FAO’s Voluntary Guidelines on the Responsible Governance of Tenure of Land, Fisheries and Forests (VGGT, 2012) provide the normative anchor [5].

The Greek land administration system presents an instructive and unusually well-documented case. The country is in the final stages of a 30-year transition from a 19th-century, person-based deed registration system, operated through 390 historical mortgage registries (ypothikofylakia), to a unified, parcel-based, digital cadastre. The transition is governed by Law 2664/1998 and consolidated by Law 4512/2018, which created the present state agency Hellenic Cadastre as a legal entity governed by public law, originally supervised by the Ministry of Environment and Energy and, since Presidential Decree 3/2021 [6], by the Ministry of Digital Governance [7]. As of mid-2026, approximately 27.85 million property rights (about 71% of the national total) are registered in operational cadastral form, with an additional 10.82 million (28%) in the formal publication phase, and full national operational coverage is targeted for completion by the end of 2026 [8]. Greek scientific work linked to the National Technical University of Athens has long anticipated this convergence, proposing machine learning support for 3D crowdsourced cadastral surveys [9] and deep learning-assisted disaster response with VGI and human-in-the-loop design [10], methods whose crowdsourcing and citizen-participation logic the cadastre has now operationalised at scale.

The Greek case repays attention on two counts. One is fairly incremental: since 2024 the cadastre has run a generative AI pipeline on Azure OpenAI Service (GPT-class large language models; the specific model version is not publicly disclosed) for the substantive legal review of property deeds [11]. Greece is hardly alone here. Comparable work is well underway across Europe. The Netherlands’ Kadaster runs “Akte AI”, registered in the Dutch government Algorithm Register, which extracts structured data from notarial deeds using AI-based OCR and a BERT-based Dutch NER model [12,13]. Sweden’s Lantmäteriet uses machine learning for building and forest-road mapping; Spain’s IGN does the same for transport maps. Germany’s BKG works on AI for disaster mapping. Estonia’s Land Board applies machine learning to topographic updating, within the broader e-Estonia digital-government stack [14,15]. The UK’s Ordnance Survey has commercialised a series of GeoAI products, and the Catalan ICGC applies AI to vegetation monitoring. All of this is documented in the EuroSDR 2024–2025 workshop on “AI in Core Business Processes of NMCAs” [14]. The deployment’s particular character comes from how three features sit together. For the legal-review step it relies on a state-of-the-art, general-purpose generative model (a large language model), not classical NLP. Its per-transaction cost figures, from €15.00 down to €0.11, are detailed enough to benchmark against. Additionally, the legacy problem behind it is severe even by southern-European standards: 390 historical mortgage registries and an archive estimated at 600 million to one billion pages of paper. This distinction between deploying technology and embedding it institutionally is central to the international literature: a recent World Bank study of 85 jurisdictions finds that registry digitisation on its own did not reliably reduce the cost or time of registration and that gains materialise only when digital records are paired with interoperability, supporting regulation and integration with other data [16].

The other has more to do with institutions than with technology. In February 2026, on the third anniversary of the Tempi rail disaster of 28 February 2023 [17], in which 57 people died, the Ministry of Infrastructure and Transport launched railway.gov.gr (accessed on 15 June 2026), a national real-time train tracking platform using HEPOS, the cadastre’s 98-station Continuously Operating Reference Station network, as the source of GNSS corrections. The technology itself is not new. The pattern (GNSS, augmentation, inertial fusion, track-matching, virtual balises) has been the subject of European R&D for over a decade (3InSat, ERSAT EAV, ERSAT-GGC, STARS, GATE4Rail, CLUG and CLUG 2.0), with commercial pilots running in Italy on the Novara–Rho and Cagliari–Decimomannu lines and trials by DB Netz, SBB and SNCF [18,19,20]. Public real-time train tracking is also widespread; passengers in at least a dozen European countries can already follow their trains on a map. The distinctive part is institutional. The cadastre’s CORS network, built and operated for surveying, supplies the corrections; the platform is public-facing, not confined to operational signalling, and the whole arrangement was assembled as the principal short-term response to a national safety crisis, ahead of and alongside a still-incomplete ETCS rollout. What the Greek case illustrates is an implementation pathway rather than a technological frontier.

In what follows, the AI legal-review pipeline and the HEPOS–railway reuse are documented from primary sources, situated in the European state of the art, and assessed for what would and would not travel to other countries.

Both deployments are examined against Greece’s recent fiscal history. A country that lost its investment-grade credit rating in 2010, went through three successive economic adjustment programmes, exited the last of them in August 2018, and regained investment grade only in 2023 has every incentive to modernise by reusing what it already owns instead of buying new systems from scratch.

The paper argues that AI-assisted deeds review and HEPOS train tracking are not unrelated tools. Instead, both show a new habit: using existing infrastructure and low-cost tools to quickly deliver public value under a tight budget.

Four questions organise the analysis:

• Technical: What is the architecture of the Hellenic Cadastre’s AI-assisted legal-review pipeline, and what verifiable performance gains has it produced?
• Infrastructure reuse: How, and to what extent, has the HEPOS reference network been repurposed as a positioning source for real-time railway monitoring, and what does that imply for the design of future national geospatial infrastructure?
• Comparative and normative: How does the combined Greek experience map onto the nine FELA pathways and the wider 2030 Agenda?
• Transferability: Which elements of the Greek model would transfer to other countries, and which are specific to the Greek context?

The account that follows moves from the historical and institutional background of Greek land administration to the technical architecture and measured impact of the cadastre’s AI-assisted legal review, then to the cross-sectoral reuse of HEPOS for railway safety, the wider applications that connect the cadastre to the 2030 Agenda, and finally to the limitations, risks and policy implications of the Greek model. The Materials and Methods (Section 2) set out the sources used and the boundary between documented fact and interpretation.

2. Materials and Methods

This is a documentary, two-case study; its empirical base is the public record, read against the European literature on AI in cadastres (notably the systematic review by Chen et al. [21]) and on GNSS for rail. Four kinds of sources were used. Primary institutional sources—the Hellenic Cadastre and HEPOS documentation, Greek and EU legislation (Laws 2308/1995, 2664/1998 and 4512/2018; Regulations (EU) 2016/679, 2016/919 and 2024/1689) [7,22,23,24,25,26], UN-GGIM (FELA) and FAO (VGGT) frameworks, and releases from GRNET and the EuroHPC Joint Undertaking—were treated as authoritative. Industry-partner documentation, principally the Microsoft customer story, was treated as a high-quality but interested source because the implementation partner is also the model provider; it was admitted only where corroborated, and the metrics drawn from it are flagged as vendor-reported in the text. Independent academic and journalistic sources were used for peer-reviewed findings and for date-stamped statements by named officials. Comparative European sources were used to situate the Greek deployment within the European state of the art.

Material was collected and verified through mid-2026. Selection favoured primary and official sources; any quantitative claim originating with the vendor was admitted only if corroborated by at least one independent source. Where sources conflicted, the official figure was preferred and the discrepancy is reported explicitly rather than silently reconciled—for example, the mortgage archive is reported as approximately 600 million pages, with the roughly one-billion-page figure in the Microsoft account differing in the scope of material counted.

To locate the Greek case within the European state of the art, we checked the Dutch government Algorithm Register entry for Kadaster’s “Akte AI” (a production, non-generative NER deployment), the EuroSDR 2024–2025 workshop report cataloguing comparable NMCA initiatives, and the European GNSS-for-rail record (3InSat, ERSAT EAV, ERSAT-GGC, STARS, GATE4Rail, CLUG and CLUG 2.0; the RFI–Hitachi pilots; the SBB, DB Netz and SNCF demonstrators), together with the public real-time train-tracking services already operating in a dozen or more European countries. On this basis, the priority claim is confined to the specific institutional configuration described in the sections that follow. Among figures that required reconciliation, the values adopted after verification were: DAEDALUS at approximately 89 PFlops (the delivered system); the TEN-T designation as the Orient/East-Med Core Network Corridor under Regulation (EU) 2024/1679 [27]; and cadastral full coverage targeted for late 2026, distinct from the summer-2026 ETCS target. The staff-training headcount could not be tied to a verifiable English-language source and is left general. Throughout, the boundary between documented fact and interpretation is maintained: where the paper advances a reading not directly supported by a source—notably the link between fiscal pressure and a “reuse-first” policy disposition—this is identified as interpretation.

3. Background: Greek Land Administration from a Historical Perspective

Greek property registration historically operated under a person-based transcription system, in place since 1836 and administered through approximately 390 mortgage registries (ypothikofylakia) [28]. Records were indexed by the names of contracting parties rather than by parcel, which produced systematic legal vulnerabilities: encumbrances on a property could not be reliably discovered without exhaustive searches across all transactions in which any past owner appeared; identical names produced ambiguity; and the geographical extent and boundaries of the parcel were not part of the legal register at all. Cadastral mapping, where it existed, was administratively separated from registration.

Modernisation began with Law 2308/1995 (cadastral survey procedure) and Law 2664/1998 (registration in the cadastre). Law 4512/2018 dissolved the previous National Cadastre and Mapping Agency S.A. (KTIMATOLOGIO S.A./NCMA) and established the Hellenic Cadastre as a unified statutory body [7], absorbing the 390 mortgage registries through a phased transition [29,30]. As of 2026, only a residual 15 stipendiary and 258 unpaid mortgage registries continue to operate provisionally as cadastral offices in areas where the transition is incomplete [29].

The early Hellenic Cadastre programme encountered serious fiscal, institutional and procedural difficulties in its first decade: cost overruns, delays, disputes with the existing notarial and registry ecosystem, and contested compensation arrangements [31]. Some of these difficulties were severe enough around the turn of the 2000s to prompt a partial restructuring of the implementing body and a search for ways to deliver the remaining work more quickly and at lower unit cost. The far deeper sovereign debt crisis that followed (Greece lost its investment-grade rating in 2010, accepted three successive adjustment programmes worth roughly €260 billion, exited the last in August 2018 and regained investment grade only in 2023) hardened that disposition into something like a standing policy preference: modernise by reusing and recombining assets the state already owns, rather than by commissioning new systems it can no longer easily afford. Researchers at the National Technical University of Athens captured this logic early: during the crisis years they proposed and field-tested a low-cost, fit-for-purpose model for cadastral surveying built on citizen participation, crowdsourcing and mobile-government services, explicitly aimed at completing the survey faster and more cheaply under fiscal pressure [32]. On that reading, the two innovations this paper documents look less like isolated technical choices than like build-back-better responses. (A) AI-assisted legal review at the cadastre and (B) the rapid cross-sectoral repurposing of HEPOS for railway safety after the Tempi disaster are what a state reaches for once it has paid dearly for past delays and overruns: cheaper and faster tools to discharge commitments it can no longer afford to keep deferring.

We use “build back better” in the sense of the Sendai Framework’s fourth priority—recovery and reconstruction that leaves a system more resilient and better-serving than before, rather than merely restoring the prior state—but extended, as UN land-sector discourse allows, to administrative recovery. The HEPOS-to-rail case meets this definition directly, as a post-disaster reconstruction that improved public safety and transparency by reusing existing infrastructure. The AI legal-review case meets it in the extended, administrative sense, rebuilding registry throughput cheaply after a decade of fiscal constraint. The further claim that fiscal pressure produced a standing “reuse-first” disposition is our interpretation of the sequence of events; it is consistent with the record but not directly evidenced by a policy document, and we present it as such.

The legacy archive of the former mortgage registries comprises approximately 600 million paper pages of deeds, mortgages, attachments, claims and indices; Microsoft’s account of the project refers to a digitisation target of roughly one billion pages, the two figures apparently differing in the scope of material counted [11]. According to the Hellenic Court of Audit’s Report 5/2025, this digitisation is now essentially complete, and the broader transition to a fully digital cadastre is expected to be finished by mid-2026 [8,33]. Microsoft has reported that the resulting digital archive is projected to amount to approximately 8 petabytes of data, hosted on Azure infrastructure [11]. Lawyers and notaries searching deeds in Athens, Thessaloniki, Piraeus and 30 other localities can now perform online title searches via the portal archive.ktimatologio.gr (accessed on 15 June 2026) [34,35,36].

The Hellenic Cadastre is a legal entity governed by public law (NPDD), based at Mesogeion 288, Cholargos, Athens, with approximately 1173 staff and an annual budget of €65 million [37]. It is a full member of EuroGeographics, whose member profile lists HEPOS as the agency’s positioning service [37]. Its responsibilities span geodetic survey, topographic mapping, cadastre, land registration, thematic mapping, aerial photography, and satellite imagery [37]. The project has since received investment via the Greek Recovery and Resilience Plan (RRP, “Greece 2.0”), funded under the EU’s NextGenerationEU instrument.

A.

The Hellenic Cadastre’s AI programme should be understood as part of a European trend documented systematically by EuroSDR (the European Spatial Data Research network, which connects national mapping and cadastral agencies, academia and the private sector). EuroSDR’s November 2024 workshop on “AI in Core Business Processes of NMCAs,” held in Amsterdam, surveyed live and planned deployments across the continent [14]. The findings, summarised in Table 1, place Greece in a clear comparative frame and make explicit that Greece is neither isolated nor strictly first; rather, it is part of an identifiable cohort of European NMCAs using AI in production, with each jurisdiction emphasising different parts of the cadastral workflow.

Table 1. Comparative snapshot of AI deployment in European NMCAs, based on the EuroSDR 2024–2025 workshop report and primary sources.

Jurisdiction/Agency	AI Application in Production or Pilot (2024–2026)	Distinctive Feature
Netherlands—Kadaster	“Akte AI”: structured-data extraction from notarial deeds (OCR and named-entity recognition); registered in the Dutch government Algorithm Register; ~1 M deeds/year	Non-generative NLP (BERT-based NER and OCR) with human-in-the-loop design; mature governance via the Algorithm Register [13]
Sweden—Lantmäteriet	ML for mapping of buildings and forest roads from earth observation imagery	Strong focus on topographic update; broad ML pipelines [14]
Estonia—Land Board/Tartu LR	ML for topographic updating within the national digital-government infrastructure	Embedded in the e-Estonia stack and the X-Road interoperability layer [14,15]
Spain—IGN	AI for updating transport (road/rail) maps	Focus on linear-feature extraction [14]
Germany—BKG	AI for disaster-mapping and governance use cases	Cross-sectoral disaster-mapping focus [14]
United Kingdom—Ordnance Survey	GeoAI for productivity and topographic mapping	Operational scale; commercial product lines [14]
Catalonia (ES)—ICGC	AI for vegetation monitoring and forest segmentation	Strong environmental focus [14]
Greece—Hellenic Cadastre	Generative AI (Azure OpenAI/GPT-class LLMs) for legal review of notarial deeds and recommendation generation; citizen chatbot; integration of HEPOS CORS into railway.gov.gr public tracking platform	Use of generative LLMs (rather than classical NLP) for the legal-review step; integration of cadastral GNSS infrastructure into a citizen-facing rail-tracking platform deployed in response to a national safety crisis [11,17]

The comparison points to two conclusions. By 2026, AI in cadastral and mapping workflows is a European phenomenon rather than a Greek innovation, so any priority claim has to be read narrowly. Greece’s own contribution is the more specific thing. The legal-review step depends on a state-of-the-art generative model rather than a classical NLP framework of the kind Kadaster uses. A generative model can handle unstructured, heterogeneous and historical deed formats with less schema engineering than a fixed-schema NER pipeline; it also carries higher risks (hallucination, non-determinism, and sensitivity to prompt and model changes) and so places a greater burden on human oversight. Due to the absence of published comparative benchmarks, we do not claim it is more accurate than the classical approach; the two occupy different points on a capability-versus-controllability trade-off. Additionally, the way it has wired the cadastral CORS network into a public, citizen-facing rail-tracking platform, launched as the immediate response to a national disaster, is an institutional arrangement we have not found replicated elsewhere. The technology in it is well established (see Section 5); what is particular is the configuration. These are the two threads the rest of the paper follows.

B.

Embedded within Hellenic Cadastre is the Hellenic Positioning System (HEPOS), a national Continuously Operating Reference Station (CORS) network developed in-house by KTIMATOLOGIO S.A. with EU co-financing and operational since 2007/2008 [38]. HEPOS comprises 98 permanent reference stations geographically distributed across mainland Greece and the islands and a control centre located at the central facilities of the Hellenic Cadastre. The network was originally equipped with Trimble NetRS GPS receivers and Trimble Zephyr Geodetic antennas, with Trimble GPSNet/RTKNet software providing the network solution [39]; a subsequent technology refresh has extended multi-constellation support. The system today supports four full GNSS constellations: GPS (United States), GLONASS (Russian Federation), Galileo (European Union) and BeiDou (China), together with the SBAS and QZSS augmentation systems [38].

Reference-station data are continuously processed at the control centre and made available to authenticated users in two modes:

• Real-time services. RTCM corrections (RTCM SC-104 standard [40], 3.x messages) are streamed over the public internet using the Networked Transport of RTCM via Internet Protocol (NTRIP) standard 10410.1, originally developed by the German Federal Agency for Cartography and Geodesy (BKG) [41]. HEPOS supports the standard Network-RTK techniques: Virtual Reference Station (VRS), Master–Auxiliary Concept (MAC) and Flächenkorrekturparameter (FKP). VRS is available for both real-time kinematic and post-processing; MAC and FKP are used in real-time services [38].
• Post-processing services. Reference-station observations are delivered in the international RINEX (Receiver Independent Exchange) format, supporting cm-level positioning for surveying applications using either dual- or single-frequency receivers [42].

Achievable horizontal positioning accuracy is at the centimetre level (typically 2–5 cm) when the rover is within the network solution area, anywhere in mainland Greece, the Peloponnese, Evia, the Ionian islands, the Cyclades and Crete [38]. A HEPOS user permit costs €120, with per-minute pricing thereafter [38].

4. AI in the Hellenic Cadastre: Technical Architecture and Performance

Prior to AI integration, the legal review of a registrable act in Greek practice required a cadastre officer to read the underlying notarial deed (which routinely runs to 100–300 pages and may exceed 200 pages for complex cases), verify the identity and capacity of contracting parties, confirm that the property was correctly identified by the National Cadastre Code (KAEK), and check the presence and validity of attached supporting certificates (tax clearance, ENFIA real-property tax declaration, energy performance certificate, urban planning certificates, etc.). A second officer would re-read the file and re-key the data into the registry system [11]. Processing took on average about 30 min per act, though complex cases would run to one or two person-hours, and in many parts of the country the wait between submission and final registration would extend to months or years [11]. The national backlog was somewhere around half a million unprocessed cases.

The system that went into production in 2024 runs on Microsoft’s Azure OpenAI Service, a Microsoft-managed deployment of OpenAI’s GPT-class large language models with enterprise data-governance terms, and is integrated with the case-management environment that cadastre staff already use [11]. Neither the cadastre, the Ministry of Digital Governance nor Microsoft’s own customer story specifies the Azure deployment type or the region in which inference takes place, and the distinction is not a technicality. Azure OpenAI offers three deployment types with materially different residency properties: Global deployments, for which prompts and completions may be processed in any Microsoft Foundry region worldwide even though data at rest remains in the customer’s selected geography; Data Zone deployments, for which processing is confined to the EU Data Boundary; and Regional deployments, for which processing and storage remain within the resource’s own region, at the cost of the most limited model availability [43]. The EU Data Zone currently spans Azure regions in France, Germany, Italy, the Netherlands, Norway, Poland, Spain, Sweden and Switzerland; it does not include Greece, and there is as yet no operational Azure region in the country, Greece Central in Attica having been announced but not yet brought into service [43,44]. Under any of the three options, the deeds of Greek citizens are therefore processed outside Greek territory, and under a Global deployment they could be processed outside the European Union altogether. Which of the three the cadastre has selected is not publicly documented. That gap is worth recording: it is precisely the kind of information a public-sector deployment of a state-of-the-art model should disclose, and its absence weakens an otherwise strong transparency record. There is likewise no complete public technical specification of the pipeline, but from the public descriptions it reconstructs as five stages. We distinguish what the sources document from what we infer: the ingestion and OCR step, the extraction of the registry fields, and the human director’s retained statutory authority are documented, whereas the use of retrieval-augmented prompting, the page-level provenance and field-level citations, and the deterministic post-checks against the KAEK index, the AADE tax-clearance system and the ENFIA database are our reconstruction of a plausible implementation rather than features asserted by the sources.

It begins with ingestion and OCR: notarial deeds and supporting certificates, submitted as PDFs, are passed through optical character recognition adapted to Greek script—including the older typographic conventions of historical attachments—to yield a structured text representation with page-level provenance. The model then classifies the act type (sale, parental gift, donation inter vivos, inheritance acceptance, mortgage, and so on) and selects the relevant rule set. A retrieval-augmented prompt next extracts the structured data items the registry requires—the parties (full names, tax IDs, capacity), the KAEK and parcel description, transaction price or value, disclosed encumbrances, and references to attached certificates—with each extracted field carrying a citation back to its source page and span [11].

A rule-based compliance check follows: codified legal rules derived from Law 2664/1998, Law 4512/2018 and downstream ministerial decisions are applied as deterministic post-checks against the extracted fields and against authoritative external registers, including the cadastre’s own KAEK index, the AADE tax-clearance system and the ENFIA database. Finally, the model generates a structured recommendation (eisigisi) for the responsible cadastre director—summarising the act, listing the extracted fields with citations, flagging discrepancies and indicating an approve-or-reject suggestion; the director retains full statutory authority for the decision.

A second Azure OpenAI deployment powers a citizen-facing digital assistant on the cadastre’s web properties, answering procedural questions in natural Greek and reducing inbound load on local registry offices. The Ministry reports a programme of weekly knowledge-base updates as new question categories emerge [11].

The metrics are grouped in two tables. Table 2 reports the AI-assisted legal-review process and Table 3 the broader cadastral digitisation and modernisation. The legal-review metrics in Table 2 are reported by the cadastre and by the implementation partner (Microsoft, which is also the model provider) and are corroborated by the Hellenic Court of Audit and Greek press; they have not, however, been independently audited for extraction accuracy, and the cost figures in particular should be read as vendor-reported.

Table 2. Metrics of the AI-assisted legal-review process, April 2026.

Metric	Pre-AI Baseline	Post-AI (April 2026)	Δ (%)	Source
Mean review time per act	~30 min (1–2 h complex)	<10 min	~−67%	[11,45]
Cost per legal-review assessment	€15.00	€0.11	−99.27%	[11]
Acts processed via AI (cumulative)	0	~310,000	—	[8,33]

Note for Table 2: The €15.00 baseline is a vendor-stated figure whose derivation is not documented in the public sources (see the derivation below); the cost figures are vendor-reported and have not been independently audited for extraction accuracy.

Table 3. Broader cadastral digitisation and modernisation indicators.

Indicator	Value	Source
Digital transactions, 2025	560,000	[33]
Digital transactions, 2026 YTD	185,000	[33]
Pages digitised (mortgage archives)	~600 million *	[8]
Total archive size (Azure)	~8 PB target	[11]
Rights in operational cadastre	27.85 M (~71%)	[33]
Rights in publication phase	10.82 M (~28%)	[33]
New digital services since 2021	12	[33]

* Microsoft reports a ~1 billion-page digitisation target; the figures differ in the scope of material counted.

The first-order economic benefit can be expressed as the avoided legal-review cost over the volume of acts processed. Let n denote the number of acts processed via the AI pipeline, ~c0~ the per-act baseline cost (a vendor-stated figure whose derivation is not set out in the public sources) and ~c1~ the post-AI per-act cost. The gross avoided review labour cost is S = n(c~0~ − c~1~). With n ≈ 310,000, c~0~ = €15.00 and c~1~ = €0.11, this yields direct cumulative savings of approximately €4.616 million—a gross figure that does not net out digitisation, cloud and compute, model access, integration and training costs, which the public record does not disaggregate, and so should be read as avoided review labour rather than a net programme saving by April 2026 [11,33]. Indirect benefits (faster transaction completion, reduced opportunity cost of capital frozen during registration delays, lower legal risk for purchasers, and induced increase in the number of transactions) are larger but harder to quantify with public data.

The 67% reduction in review time is realised by cadastre directors as well, which expands the institution’s effective decision-making capacity without a proportionate increase in staffing. Training has been rolled out to staff in cadastral offices nationwide to support the new AI-assisted workflow [11].

5. Cross-Sectoral Reuse of HEPOS: From Surveying to Real-Time Railway Monitoring

On 28 February 2023, IC62, an Athens–Thessaloniki passenger train operated by Hellenic Train collided head-on with an intermodal freight train south of the Tempi Valley near Larissa, killing 57 people and injuring dozens more in Greece’s deadliest rail accident [17]. A criminal trial that opened in March 2026 involves 36 defendants, of whom 33 face felony charges [46]; public investigations have concluded that the European Train Control System (ETCS) Level 1 specified for the Patras–Athens–Thessaloniki–Idomeni axis (the Greek section of the Orient/East-Med Core Network Corridor, now part of the Western Balkans–Eastern Mediterranean Corridor under the revised TEN-T Regulation 2024/1679) had been scheduled for operation since 2016 but was not in service at the time of the accident [22,47]. As of the official ministry briefing of February 2026, the Athens–Thessaloniki main axis is being brought to full (100%) signalling, remote control and ETCS by summer 2026, with onboard ETCS already installed on 100 trainsets [17].

Before describing the Greek system, it is necessary to position it within an active European research and deployment programme that long predates the Tempi accident. The use of GNSS for train positioning has been a strategic priority of the European Union and the European Union Agency for Railways (ERA) since at least 2012, on the rationale articulated in the ERA’s 2015 Report on ERTMS Longer Term Perspective: that satellite-based positioning would allow a significant reduction in trackside equipment and improve the performance of train localisation and would thereby reduce the cost of bringing ERTMS to regional lines [18]. The principal programmes are summarised in Table 4.

Table 4. Principal European programmes on GNSS-based train positioning, 2012–2026.

Programme	Years	Lead Actors	Outcome Relevant to the Greek Case
3InSat	2012–2014	ESA, Ansaldo STS (now Hitachi Rail STS), RFI	Established feasibility of satellite-based ERTMS
ERSAT EAV (H2020)	2015–2017	Ansaldo STS, RFI, ESA, GSA, Italian Space Agency	First live European rail test using Galileo positioning (Cagliari–Decimomannu, Sardinia, 24 Feb 2017); introduced “virtual balises” every 50 m
STARS, ERSAT-GGC	2016–2019	Multi-stakeholder consortia under GSA/EUSPA	Refined GNSS performance models for rail; route classification methodology
CLUG, CLUG 2.0	2019–present	SBB (CH), DB Netz (DE), SNCF (FR), Siemens, CAF, Airbus DS, Naventik, FDC, ENAC, Navcert	Certifiable onboard localisation unit with GNSS + IMU + odometer fusion; designed for ERTMS integration
GATE4Rail, SBS RailS	2019–present	Shift2Rail Joint Undertaking	Digital test platforms and tools for GNSS-ETCS certification
RFI/Hitachi Rail commercial pilots	2020–present	RFI (IT), Hitachi Rail, Alstom Ferroviaria	First commercial-scale satellite-based ERTMS deployment, Novara–Rho line; multi-billion-euro tenders for nation-wide rollout in Italy under PNRR/NextGenerationEU

The standard architecture that emerges from this programme is the combination of (i) multi-constellation GNSS receivers, (ii) GNSS augmentation (in Europe typically EGNOS for SBAS, plus Network-RTK from a national CORS network), (iii) Inertial Measurement Units (IMUs) and odometry for dead reckoning where satellite visibility is degraded, and (iv) onboard high-resolution digital track maps with map-matching algorithms [19,48]. This is the architecture used in CLUG and ERSAT, and it is the architecture used at railway.gov.gr. The Greek system is, in technical terms, an application of this European stack to the Greek national network.

Public-facing real-time train tracking is, by 2026, a routine service of European national rail operators: national railways such as Deutsche Bahn, SNCF, SBB, Trenitalia and Nederlandse Spoorwegen publish live train positions through their own apps or open APIs, and Europe-wide aggregator services have existed since the mid-2010s. The integration of real-time train position data with public-facing maps is therefore not, in itself, distinctive.

On 25 February 2026, the Ministry of Infrastructure and Transport launched railway.gov.gr, a national real-time train tracking platform [49,50]. The platform’s design rationale, as articulated by Deputy Minister of Infrastructure and Transport Konstantinos Kyranakis at the launch briefing, is to provide an additional, independent layer of operational safety and public transparency on top of the deploying ETCS infrastructure: it strengthens rather than replaces existing safety systems, functioning as “an additional safety layer, a digital safety net operating continuously in real time” [17]. Initial deployment covers the Athens–Thessaloniki corridor, with full network coverage of approximately 2000 km targeted by April 2026 [49].

The technical architecture of the railway tracking platform draws together four components. Two are documented specifically for railway.gov.gr—the onboard GNSS receivers (i) and the HEPOS correction feed (ii); the inertial-navigation fusion (iii) and track-matching (iv) layers are described here from the general European reference architecture (ERSAT, CLUG) and are not separately documented for the Greek platform:

(i) Onboard GNSS receivers support multi-constellation positioning, in particular GPS and Galileo [49], and feed raw observations into the onboard processing module.

(ii) HEPOS Network-RTK corrections are streamed from the Cadastre Control Centre to onboard receivers using NTRIP-distributed RTCM 3.x messages [38,41]. The choice of network technique (VRS, MAC or FKP) determines how the rover combines observations from neighbouring reference stations to model atmospheric and orbital errors at the rover position; under nominal conditions Network-RTK with double-differenced carrier-phase ambiguity resolution yields horizontal accuracies of 1–3 cm and vertical accuracies of 2–5 cm, with convergence times of seconds [38].

(iii) Inertial Navigation System (INS) integration. Because GNSS signals are unavailable in tunnels and degraded in cuttings and dense urban canyons, the onboard system fuses GNSS-derived positions with inertial measurements (typically MEMS- or fibre-optic gyros plus accelerometers) using a Kalman or extended-Kalman filter, providing dead-reckoning continuity across signal outages [48].

(iv) Track-matching (map-matching) algorithms. The fused GNSS + INS solution is constrained to a high-resolution digital map of the track network (track-constraint imposition). This matters because distinguishing adjacent parallel tracks—with track-centre separations on the order of 4 m—is at or beyond the accuracy of stand-alone GNSS; RTK-level corrections combined with track-matching bring the train onto a specific track with the accuracy a public tracking display requires. Achieving the safety-integrity level of track discrimination needed for train control (ETCS) remains an active research and certification problem, as the ERSAT, RHINOS and CLUG programmes attest [19,20,48,51].

This is the standard European architecture for satellite-based train positioning (discussed above). What is more specific to the Greek system is the institutional choice to use the national cadastral CORS network as the source of GNSS corrections, rather than commissioning a dedicated railway reference network as Italy did under the ERSAT programme. This is a legitimate design option: cadastral CORS networks are operated by national mapping/cadastral agencies in most European jurisdictions and are routinely used as horizontal infrastructure across multiple sectors (precision agriculture, construction, machine control, asset management). It is, however, a configuration choice rather than a technical novelty. Its principal advantages in the Greek context are speed of deployment and avoidance of new physical infrastructure; its disadvantages, discussed in Section 7, include institutional dependency on a service designed for surveying rather than safety-relevant use. Several factors plausibly favoured HEPOS over a dedicated railway reference network: it already provided national, multi-constellation, centimetre-level coverage over the relevant corridors; it was operable within the short post-crisis timeframe with no new reference infrastructure; it avoided fresh capital expenditure under fiscal constraint; and cadastral CORS networks are, in most European jurisdictions, routinely reused as horizontal infrastructure across sectors. The internal decision record is not public, so this rationale is reconstructed from the documented properties of the two systems rather than from a published justification.

At launch, the Ministry reported a set of functional capabilities for railway.gov.gr [49,50]. The platform provides a public-facing live map of train positions, speeds, headings, estimated arrival times and delays; it raises operational alerts for deviation from designated routes and for unsafe proximity between trains; it maintains a continuous data feed to the OSE (Hellenic Railways Organisation) control centre; and it keeps an append-only log of position data that supports subsequent forensic reconstruction of incidents.

These capabilities should be read in the same register as the platform itself (discussed below): they are advisory monitoring and public-information functions, not safety-certified train-control functions. The route-deviation and unsafe-proximity alerts are decision-support signals to the OSE control centre and to the public, operating at the integrity level of a positioning-and-monitoring service; train separation with safety-of-life integrity remains the responsibility of ETCS and conventional signalling, not of railway.gov.gr. As with track discrimination discussed above, obtaining safety-integrity performance from GNSS-based positioning is the open problem that the ERSAT, CLUG and RHINOS programmes exist to address [19,20,51].

It is worth being explicit about why. Centimetre-level positioning accuracy is necessary but not sufficient for railway safety: suitability for train control additionally requires integrity (bounded, quantified confidence in the reported position, with timely alerts when it cannot be guaranteed), availability, continuity, bounded latency, and fault detection and exclusion. It is the demonstration of these RAMS and GNSS-integrity properties, rather than accuracy alone, that the ERSAT, RHINOS and CLUG programmes exist to establish and that railway.gov.gr, as a positioning-and-monitoring service, does not claim.

It helps to be precise about what is, and is not, particular to the Greek case. The technology behind railway.gov.gr is not itself new: it is the same GNSS + IMU + map-matching architecture worked out under ERSAT, CLUG and the rest. Nor is public real-time train tracking novel, with comparable services running in at least a dozen European countries. The distinctive element, as the European background above makes clear, is institutional.

Consider first where the GNSS corrections come from. The Greek railway operator (OSE) draws them not from a dedicated railway reference network and not from a commercial provider, but from HEPOS, the national cadastral CORS network. That is a cross-sectoral data-reuse decision and an institutional one at least as much as a technical one: the cadastre had to be willing, on a short timeline, to run a service designed for surveyors as a feed into a safety-relevant monitoring setting.

A second choice concerns who sees the position data: citizens, not just signallers. Most European operators publish real-time train positions, but the framing in Greece is different. Here the platform is openly an instrument of public accountability following a national disaster. Operational alerts (route deviation, unsafe proximity to another train) and forensic logging of every position appear as headline features rather than as internal traces.

Timing matters as well. The platform was deployed as the principal short-term institutional response to a national safety crisis and is explicitly framed by the Greek government as a complement to a still-incomplete ETCS rollout (with the Athens–Thessaloniki axis targeted for full signalling, remote control and ETCS by summer 2026 [17]). The point should not be pushed too far: railway.gov.gr is a positioning, monitoring and alerting platform; it is not a Train Control System in the regulatory sense. The legally mandated mechanism for European interoperability and safety-critical train control remains ETCS, now deployed under Commission Implementing Regulation (EU) 2023/1695 (TSI “CCS”, which repealed Regulation 2016/919 in September 2023) [52]. The HEPOS-based platform is correctly described as a complementary safety and transparency layer that compensates for the slow rollout of ETCS, not a substitute for it.

One further advantage of the satellite-based approach tends to be underrated: it is resilient to common ground-infrastructure failures such as outages of trackside signalling [53]; in the Greek context this matters particularly because theft of copper signal cables has been a recurring and widely reported problem of the network. This resilience is inherent to any satellite-based positioning architecture and is therefore not specifically Greek, but it is more salient in a network with the documented infrastructure problems of OSE.

The upshot is that the Greek case is most usefully read as a worked example of one institutional assembly (cadastral-agency CORS, public-facing platform, post-crisis deployment) rather than a technological frontier. What travels to other countries is the configuration, not an invention.

Post-Crisis Infrastructure Reuse: A Comparative Reading

The Greek response is usefully read against three other European post-disaster episodes (Table 5). After the Morandi bridge collapse in Genoa (2018), Italy rebuilt rapidly and expanded structural monitoring, but the response was largely sector-specific reconstruction. After the Alvia derailment at Santiago de Compostela (2013), Spain’s response centred on signalling and automatic train protection on the affected line—reform within the rail sector. After Grenfell (2017) [54], the United Kingdom created new building-safety institutions and registers (culminating in the Building Safety Act 2022 [55])—an institutional rather than cross-sectoral-infrastructure response. Greece is distinctive in that the principal short-term response repurposed an existing national geospatial asset (HEPOS) across an institutional silo, from cadastre to rail, rather than building or reforming within the affected sector alone. This is an analytical typology, not an exhaustive account; a systematic multi-case study of what makes such cross-sectoral reuse possible is left for future work.

Table 5. Post-crisis institutional responses compared.

Crisis (Year)	Sector	Dominant Response	Cross-Sectoral Reuse of Existing Infrastructure?
Tempi, Greece (2023)	Rail	Public real-time tracking platform on HEPOS as immediate response, ahead of ETCS	Yes: National cadastral CORS reused for rail monitoring
Genoa/Morandi, Italy (2018)	Road	Rapid rebuild; expanded structural monitoring	Limited: Mainly within-sector reconstruction
Santiago de Compostela, Spain (2013)	Rail	Signalling/ATP scrutiny and reform on the line	No: Within-sector safety reform
Grenfell, UK (2017)	Buildings	New building-safety regulator and registers (Building Safety Act 2022)	No: Institutional/regulatory reform

6. Wider Applications and Linkages with Agenda 2030

This section investigates the wider AI applications in land administration in Greece and their relevance to the 2030 Agenda. The research shows that the overall Greek experience of introducing AI in the broader land administration sector maps readily onto all nine strategic pathways of the Framework for Effective Land Administration.

On governance, institutions and accountability, Law 4512/2018 unified the cadastre and absorbed the 390 mortgage registries, a transition now largely completed; the policy and legal pathway is correspondingly mature, resting on Laws 2308/1995, 2664/1998 and 4512/2018 together with recent fee regulations (KYA 2/2026) and rules on digital deeds. Financing is stable, drawing on EU co-financing and Greece 2.0 recovery funds via NextGenerationEU against an annual budget of some €65 million. The data pathway is well advanced—parcel-based registration keyed to the KAEK identifier, the HEPOS CORS network, national ortho-imagery and the INSPIRE Geoportal operational since 2020—while innovation sits at the frontier, with the Azure OpenAI legal-review system, the citizen chatbot and the Daedalus HPC facility for geospatial analytics.

The remaining pathways are equally well served. Standards compliance is solid, spanning INSPIRE, the RTCM/NTRIP/RINEX stack for HEPOS and the ETRS89/ITRF reference frames [56]. The partnerships pathway is equally active, and it is worth setting out who the partners are and what each contributes. Microsoft supplies the Azure OpenAI Service on which both the deed-review system and the citizen chatbot run [11]. On the computing side, the European High-Performance Computing Joint Undertaking (EuroHPC JU), the EU body that procures and co-finances Europe’s supercomputers, is co-funding DAEDALUS, the 89-petaflop national machine, with 35% of its €36 million acquisition cost against 65% from the Greece 2.0 Recovery and Resilience Plan [57]. The machine itself is supplied and installed by Hewlett Packard Enterprise (HPE) under a €41 million contract, while MainSys S.A. is building the electromechanical infrastructure and the Lavrion data centre that will house it under a further €17.9 million contract; both contracts were awarded by GRNET, the national research and technology infrastructure operator, which will manage and operate the system [58]. DAEDALUS is run as a regional consortium: Cyprus, North Macedonia and Montenegro participate as EuroHPC member states with in-kind contributions and share access to the machine in proportion to their investment, extending Greek computing capacity into a South-East European partnership rather than confining it to national use [57,58]. The Hellenic Centre for Defence Innovation (ELKAK) channels public funding into geospatial AI start-ups and into the indigenous ISR-satellite programme [59], and NTUA contributes research and training capacity. Capacity-building is ongoing through staff training across the cadastral offices, NTUA high-performance-computing training and the ELKAK Innovation Challenges, and advocacy and awareness are strong, through the citizen chatbot, the archive.ktimatologio.gr and railway.gov.gr portals and the INSPIRE Geoportal. Rather than treat every pathway as uniformly strong, Table 6 sets out, for each of the nine FELA pathways, the documented evidence, an assessed maturity, and the remaining limitations; on the accuracy-and-disclosure grounds discussed in Section 7, the Data (IV) and Standards (VI) pathways are the least well-evidenced, since the substantive accuracy of the AI extractions is undisclosed and there is no published conformance or independent-verification evidence for either the extractions or the HEPOS-to-rail feed. The importance of interoperability seen here is not incidental: the same World Bank analysis identifies connectivity, interoperability and integration with spatially explicit data as the three channels through which digitisation actually improves registry performance, and the Greek reuse of HEPOS corrections and of cadastre–DEDDIE–E9 cross-referencing is a concrete instance of all three [16].

Table 6. FELA pathway assessment for the Greek case (criteria-based).

FELA Pathway	Documented Evidence	Assessed Maturity	Remaining Limitation/Gap
I. Governance, Institutions & Accountability	Law 4512/2018 unified the Cadastre; 390 registries absorbed; supervision under the Ministry of Digital Governance (PD 3/2021)	Mature	Accountability of the AI-assisted decision is not independently audited
II. Policy & Legal	Laws 2308/1995, 2664/1998, 4512/2018; fee rules (KYA 2/2026); digital-deeds rules	Mature	No published legal basis/DPIA for cadastre–DEDDIE–E9 cross-referencing; AI-Act high-risk boundary unresolved
III. Financial	EU co-financing; Greece 2.0/RRP; ~€65 M annual budget	Advanced	Reported savings are gross and vendor-sourced; no published net cost–benefit
IV. Data	Parcel-based KAEK; HEPOS CORS; ortho-imagery; INSPIRE Geoportal	Advanced but under-evidenced	Substantive extraction accuracy undisclosed; performance on handwritten/pre-1980 deeds unknown
V. Innovation	Azure OpenAI legal review; citizen chatbot; DAEDALUS HPC	Frontier	Single-vendor dependency; no published open-weight benchmarking
VI. Standards	INSPIRE; RTCM/NTRIP/RINEX; ETRS89/ITRF	Solid but unverified	No published conformance/interoperability testing for the HEPOS → rail feed; no reported LADM (ISO 19152 [60]) alignment
VII. Partnerships	Microsoft; EuroHPC JU/GRNET/HPE/MainSys; ELKAK; NTUA; regional HPC consortium	Active	Implementation partner is also the model provider (independence concern); governance of the safety-relevant reuse undocumented
VIII. Capacity & Education	Staff training across cadastral offices; NTUA HPC training; ELKAK challenges	Ongoing	Training headcount not verifiable in an English-language source; no evaluation of oversight readiness
IX. Advocacy & Awareness	Citizen chatbot; archive.ktimatologio.gr; railway.gov.gr; INSPIRE Geoportal	Strong	Limited public communication of data-protection guarantees; no published engagement evaluation

Beyond this mapping to the FELA framework, the cadastre’s spatial data underpin a number of concrete, SDG-linked applications. In terms of specific targets, the programme bears most directly on SDG 1.4 and indicator 1.4.2 (secure tenure rights), 9.1 (resilient infrastructure), 11.3 and 11.b (participatory planning and disaster-risk reduction), 13.1 (resilience to climate-related hazards), and 16.6 and 16.10 (effective, accountable institutions and public access to information). These applications share a common thread—each reuses the cadastre’s spatial data, and for compute-intensive cases the national HPC capacity, across an institutional boundary—and are treated here at a depth proportionate to their documented dependence on that data, with the least-connected uses flagged as prospective. The cadastre’s geospatial data (parcel boundaries, digital terrain models, ortho-imagery and 3D building footprints) support two energy use cases. First, automated screening of utility-scale renewable energy sites by combining cadastral attributes (parcel ownership, area, slope) with environmental constraint land-use layers for protected lands (forest maps, Natura 2000 zones, archaeological zones), reducing pre-feasibility timelines for site identification. Second, building-level solar potential analysis using digital surface models, which is being developed in line with the EU’s Energy Performance of Buildings Directive (recast) of 2024 and the Greek Energy Law [61].

Cross-referencing cadastral records with electricity-consumption data from the Hellenic Electricity Distribution Network Operator (DEDDIE) and tax declarations (the E9 statement of real-estate holdings filed with the Independent Authority for Public Revenue, AADE) is a methodologically standard approach to identifying vacant housing stock and planning for affordable-housing policy. The Greek government has used variants of this approach to design the Anakainizo–Noikiazo (Renovate-and-Rent) programme launched in late 2023/early 2024 and relaunched in enhanced form (60% subsidy, up to €8100) [62]. The cadastre also enables the identification of disused publicly owned land that the government has signalled could be released for social-housing development, such as military installations reported as candidates for transfer. The specific sites and timelines, however, have not been formally confirmed. Such data linkage, and the use of machine learning for fraud detection and tax enforcement in land records, is precisely the emerging frontier that the international literature now identifies for property institutions [16].

Disaster risk reduction is a further domain, and here new national computing capacity is central. The DAEDALUS supercomputer is a EuroHPC mid-range system delivering more than 89 PFlops (FP64 double precision) and integrating over 2000 NVIDIA Grace Hopper (GH200) superchips with direct liquid cooling [57,58,63]. It was commissioned by GRNET (the Greek National Infrastructures for Research and Technology), built by Hewlett Packard Enterprise with electromechanical works by MainSys S.A., and hosted in the historic 19th-century Electric Power Station of the National Technical University of Athens’ Lavrion Technological & Cultural Park. The procurement contract was signed in March 2025 between EuroHPC JU, GRNET and HPE. The total acquisition cost is €36 million, with the EuroHPC JU contributing 35% via the Digital Europe Programme and Greece contributing 65% via the RRP. Cyprus, Montenegro and North Macedonia are consortium partners [58,64]. Daedalus is integrated into Pharos, the Greek EuroHPC AI Factory [64].

DAEDALUS provides the computational substrate for high-resolution wildfire-spread simulation, urban flood-inundation modelling and Digital Twin development [65,66], all of which depend on the cadastre’s high-resolution terrain and land-use data. This kind of integration, with national land-administration data feeding national HPC for civil-protection modelling, is precisely the cross-sectoral synergy that the IGIF anticipates and that FELA’s Innovation pathway encourages [4].

Further, more loosely connected areas—prospective or adjacent, rather than demonstrably dependent on cadastral data or the AI pipeline—include: (i) road safety, where the National Technical University of Athens (NTUA) Road Safety Observatory provides crash data analytics that, together with cadastral and infrastructure data, are used in the placement and operation of automated enforcement cameras [67]; (ii) defence applications via the Hellenic Centre for Defence Innovation (ELKAK), which channels public funding into geospatial AI start-ups and into the indigenous ISR-satellite programme [59]; and (iii) education, where NTUA’s high-performance computing capabilities (in addition to its hosting of DAEDALUS) support student and faculty research on geospatial AI.

7. Discussion: Limitations, Risks and Open Questions

Relying on a single proprietary cloud and model provider (Microsoft, in this case, via Azure OpenAI Service) raises an obvious question about long-term data sovereignty and procurement independence. The Microsoft customer story emphasises enterprise data governance and the absence of training on customer data, and Azure’s European regions and EU Data Zone offer credible residency guarantees [11,43]. As noted in Section 4, however, the deployment type and region actually used by the cadastre have not been disclosed, so the strength of those guarantees in this particular case cannot be verified from public sources; and since no Azure region is yet operational in Greece, the processing takes place abroad in any event [44]. The operational dependency is real in a second sense as well: a future change in commercial terms, in model availability, or in the regulatory environment could materially affect whether the cadastre can keep operating the system as it is. This is not a specifically Greek problem. It applies to every national-scale public-sector deployment of a state-of-the-art LLM. The mitigations also generalise: publication of the deployment type and region, parallel internal benchmarking against open-weight models such as Llama or Mistral running on DAEDALUS, and clear contractual provisions for data portability.

Regulation (EU) 2024/1689 of the European Parliament and of the Council (the AI Act), which entered into force on 1 August 2024 and is being applied progressively through to August 2026, classifies AI systems used by public authorities to evaluate the eligibility of natural persons for benefits and services as high-risk (Annex III, point 5(a)) [23]. The cadastre’s legal-review system is at the boundary of this classification: it does not directly determine eligibility for a public benefit, but it does generate the substantive recommendation on which the registry director’s decision is based. Best practice under the AI Act should be applied irrespective of the formal classification: a documented risk-management system, technical documentation, transparency to the data subject, human oversight, accuracy, robustness and cybersecurity. The cadastre’s preservation of full statutory authority with the human director is a necessary but not sufficient condition for compliance.

The substantive accuracy of the AI extractions in routine production is still not disclosed in any public source we have found: there are no published figures for false-positive or false-negative rates on rule violations, for the accuracy of party-name extraction across diacritical and transliteration variants, or for performance on handwritten or pre-1980 deeds. A serious published evaluation (ideally by an independent auditor with access to a stratified random sample of pre- and post-AI cases with ground-truth annotations) is overdue. Until one exists, the cost-reduction figures need a caveat: a per-act cost of €0.11 looks rather different if the error rate is materially higher than the human baseline.

Cross-referencing of cadastral records with electricity-consumption, tax and other administrative data (useful as it is for housing policy and fraud detection) expands the state’s ability to infer behaviour from administrative traces. This cross-referencing is governed by the General Data Protection Regulation (Regulation (EU) 2016/679) and by Greek Law 4624/2019; specific cross-referencing operations require an articulated legal basis and, typically, a Data Protection Impact Assessment [24]. Public communication of these guarantees should be part of the cadastral modernisation programme, not an afterthought.

This paper has been, in the main, a documentary synthesis. Its empirical base is the public record (primary institutional sources, the implementation partner’s case study, the peer-reviewed literature where it exists, and Greek and international press), read against the European literature on AI in cadastres and on GNSS for rail. Its new contribution is interpretive rather than evidentiary. We position the Greek deployment within the European cohort, map the case explicitly onto the nine FELA pathways, and set out a typology of what does and does not travel from the Greek institutional configuration to other countries. Several lines of follow-up work would put that analysis on firmer ground, and we list them below in the hope that we or others will pursue them.

The most pressing gap is on the substantive accuracy of the AI legal-review pipeline. As noted above, no public source has yet disclosed the false-positive or false-negative rates of the system, the accuracy of party-name extraction across diacritical and transliteration variants, or its performance on handwritten or pre-1980 deeds. An independent evaluation (ideally by an external auditor with access, under a non-disclosure agreement, to a stratified random sample of pre- and post-AI cases with ground-truth annotations from senior cadastre directors) would do more to ground the cost-reduction figures than any amount of additional documentary work. Equally useful would be a structured set of interviews with named decision-makers: the Deputy Minister of Infrastructure and Transport, the leadership of the Hellenic Cadastre, engineers at HPE and MainSys S.A. on the DAEDALUS build, and operational staff at OSE on how the HEPOS-derived position feeds are used in practice. Such interviews would also clarify decisions for which the public record only offers outcomes: why the cadastral CORS network was chosen as the source of corrections rather than a dedicated railway reference network, and how the institutional negotiation between the cadastre and OSE was conducted on what appears to have been a very short timeline.

The Greek case also invites at least two theoretical extensions. One is a typology of post-crisis institutional reuse of geospatial infrastructure: when does a country, in the wake of a disaster, repurpose existing public infrastructure across silos instead of building new infrastructure in each one, and what political, financial and administrative conditions make that choice possible? Section 5 begins this comparison against Italy (Genoa, 2018), Spain (Santiago de Compostela, 2013) and the United Kingdom (Grenfell, 2017); a systematic, coded multi-case treatment remains to be done. The other extension is a stress-test of the FELA itself. The nine pathways were endorsed in 2020 and have not, to our knowledge, been systematically tested against a single real-world case of cross-sectoral reuse at national scale. The Greek experience can be read as either confirming or under-specifying the existing pathways, and either reading would justify a methodological paper on the framework.

On the empirical side, the comparative discussion in Section 3 stops at the descriptive level. A natural extension is a coded comparative dataset across the ten or so European NMCAs currently deploying AI in production or in serious pilot, capturing for each agency the technical architecture (classical NLP versus generative LLMs versus computer vision); the governance model (Algorithm Register entry; published DPIA; statutory authority for human override); the disclosed performance metrics; and the institutional independence of the implementation partner. A dataset of this kind would allow a quantitative answer to the question that the present paper can only gesture at: what predicts whether AI in a national NMCA delivers public-value gains commensurate with its costs [21,68].

8. Policy Recommendations

From the analysis above we draw seven recommendations for governments weighing AI-enabled modernisation of land administration.

R1. Treat the national CORS network as horizontal critical infrastructure. A modern multi-constellation CORS network, originally justified by the cadastre, is also the natural backbone of precise navigation for transport, agriculture, autonomous mobility and disaster monitoring. Plan for, document and budget for these reuses from the outset.

R2. Sequence digitisation before automation. AI cannot review what has not been digitised. The Greek experience confirms that the rate-limiting step in cadastral modernisation is the digitisation and structuring of legacy archives. Begin the bulk-digitisation programme as early as possible, in parallel with the design of the AI pipeline.

R3. Apply AI to high-volume, well-codified review tasks first. The deeds-review use case in Greece succeeded because the rule set is large but stable, the input is reasonably structured (notarial deeds), and human discretion was preserved at the decision point. Generalise this design pattern; do not extend AI into discretionary or judgement-intensive areas without further safeguards.

R4. Mandate publication of an independent accuracy audit. Cost reductions must be reported alongside an independent evaluation of the AI system’s accuracy versus the human baseline. A coordinating international body could usefully maintain a comparable methodology and template.

R5. Default to open data, governed by GDPR-equivalent safeguards. The deterrent effect of transparency on corruption and on conspiratorial narratives is much higher when independent third parties (academia, journalists, civil society) can verify the data themselves. Greece’s INSPIRE Geoportal, archive.ktimatologio.gr and railway.gov.gr provide concrete templates.

R6. Plan for vendor independence. Treat reliance on a single hyperscaler or single LLM provider as a known risk; require interoperability, contractual data portability, and ongoing benchmarking against open-weight alternatives running on national HPC.

R7. Embed AI deployment into FELA reporting. AI deployment, accuracy audits and cross-sectoral data-reuse cases should be included as standard reporting items under FELA pathways 5 (Innovation) and 9 (Communication and engagement). This would let international comparison and learning accelerate.

9. Conclusions

The cadastre’s Azure OpenAI deployment is one of the more advanced documented production uses of generative AI for administrative work in a national land registry. But Greece is part of a recognisable European cohort (Kadaster, Lantmäteriet, Estonia’s Land Board, IGN, BKG, Ordnance Survey, ICGC), and the operational priority claim should be read narrowly. Three things mark it out: the use of state-of-the-art, general-purpose large language models for the substantive review step, the public per-transaction economics, and the unusually severe legacy archive on which all of this has been thrown. The reported gains are material: review time has dropped from about thirty minutes on average (one to two hours for complex cases) to under ten minutes, the vendor-reported per-act cost from €15.00 to €0.11, and roughly 310,000 acts have been processed.

The part of the story that we think matters most for policy is the cross-sectoral one. HEPOS, which was built and operated to underpin the cadastre, is now also the positioning backbone of a national real-time train tracking platform put together in response to a rail disaster. The technology is not novel; satellite-based train positioning has been a mature European programme for years, with pilots running in Italy, Switzerland, Germany and France, and public real-time train tracking is offered by national operators across the continent. What is specific in Greece is the institutional assembly: the cadastral agency’s CORS network as the positioning source, that data embedded in a public-facing platform, not confined to internal signalling, and the whole arrangement deployed as the principal short-term response to a national safety crisis ahead of a still-incomplete ETCS rollout. This is, more or less, what the IGIF means when it asks governments to treat geospatial information as horizontal national infrastructure rather than as the property of one ministry.

In its main lines the Greek model is replicable, but only under the conditions set out in Table 7. The technical components are well-standardised (RTCM, NTRIP, RINEX, GNSS, INS, ETRS89 [56]). A unified parcel-based cadastre with a national CORS network is a widespread institutional pattern. The policy framework (FELA, IGIF, VGGT, the SDGs) is already global. What is hard to replicate (and what the Greek case really teaches) is the political, financial and institutional alignment that let a country reuse infrastructure across silos instead of building new infrastructure in each silo and do it quickly, in the wake of a crisis, when the temptation to do nothing, or to build something elaborate from scratch, would have been just as strong. That this disposition took hold in a country only recently out of a decade-long fiscal emergency is probably no accident. When a state cannot afford to procure its way to modernity, reusing and recombining the assets it already owns stops being a clever improvisation and becomes the rational default. That is the part we think other countries should take away.

Table 7. Transferability of the Greek model.

Element	Transferable?	Condition/Note
Technical components (multi-constellation CORS + Network-RTK; RTCM/NTRIP/RINEX; GNSS + INS + map-matching; generative-LLM deeds-review pattern; parcel-based digital cadastre)	High	Standardised and vendor-neutral; the principal portable layer
Institutional and legal prerequisites (unified statutory cadastral agency with data-custody mandate; legal authority to reuse CORS cross-sectorally; preserved human statutory decision authority)	Conditional	Present in many jurisdictions but not all; determines whether reuse is legally possible
Data-governance requirements (DPIA and articulated legal basis for administrative cross-referencing; disclosure of AI deployment type/region; independent accuracy audit)	Required	Not yet fully met even in Greece; a precondition, not an outcome
Financial conditions (access to EU co-financing/RRP-type funds; low per-transaction economics; cost-recovery pricing)	Context-dependent	Hinges on access to EU or equivalent instruments
Context-specific elements (post-Tempi urgency; fiscal-crisis reuse-first disposition; pre-existing 98-station national CORS; the specific cadastre–OSE negotiation)	Low/none	Explains the speed of the Greek case, not a replicable design