Artificial Intelligence in Cadstral Practice: The Servitut Engine
Authors
Thernøe, Magnus Buur ; Binderup, Johan
Term
4. semester
Publication year
2026
Submitted on
2026-06-04
Pages
82
Abstract
In Danish cadastral practice, preparing a servitude statement requires a land surveyor to read large volumes of registered acts to identify burdens on a property, interpret them, and advise on how to handle each. This thesis examines to what extent, and for which parts of that task, current artificial intelligence can assist while maintaining factual grounding, traceability, and professional oversight. It develops and evaluates a prototype, the Servitut Engine, built on large language models with retrieval-augmented generation. Scanned acts pass through a staged pipeline—OCR, text chunking, keyword pre-screening, structured extraction, and report generation—with outputs exposed at each step so every report entry is traceable to its source. The prototype was benchmarked against professional statements in four real cases and refined across three iterations. After refinement, the engine closely matched experts on tasks that depend on reading and structuring source material but performed weaker on judgment-heavy tasks such as handling recommendations and classification of a servitude’s scope. The thesis concludes that this kind of AI should not replace the surveyor, because errors are too costly, but it is well suited as decision support to validate drafts and surface servitudes buried in long documents, under responsible professional oversight.
I dansk matrikelpraksis kræver en servitutredegørelse, at en landinspektør manuelt gennemgår mange tinglyste dokumenter for at identificere, fortolke og rådgive om ejendommens byrder. Dette speciale undersøger, i hvilket omfang og for hvilke delopgaver nutidens kunstige intelligens kan understøtte arbejdet uden at gå på kompromis med faktuel forankring, sporbarhed og fagligt tilsyn. Til formålet udvikles og evalueres en prototype, Servitut Engine, der bygger på large language models og retrieval-augmented generation. Scannede dokumenter behandles gennem en trinvis pipeline med OCR, tekstopdeling, nøgleordsforfiltrering, struktureret udtræk og rapportgenerering, hvor hver fase gør sine resultater synlige, så hver rapportpost kan spores til den oprindelige kilde. Prototypen blev sammenholdt med professionelle servitutredegørelser i fire reelle sager og løbende forbedret over tre iterationer. Efter forbedringerne viste motoren høj overensstemmelse på de dele, der handler om at læse og strukturere kildematerialet, men svagere resultater på vurderingsprægede dele som håndteringsanbefalinger og klassifikation af servitutters rækkevidde. Specialet konkluderer, at denne type AI ikke bør erstatte landinspektøren, fordi fejl er for dyre, men at den er velegnet som beslutningsstøtte til at kvalitetssikre udkast og fremhæve servitutter, der er gemt i lange dokumenter, under ansvarligt fagligt tilsyn.
[This apstract has been generated with the help of AI directly from the project full text]