CuttingEdgeAI - Agentic AI

I arbeidet med å innføre den nye grunnrenteskatten på landbasert vindkraft skal prosjektet blant annet støtte “likebehandling i konsern”. Bakgrunnen til dette er følgende:

• Komplekse lover og reguleringer
• Prinsipp om likebehandling-må ta utgangspunkt i tidligere vurderinger
• Mange kilder å hente informasjon fra
• Tidkrevende å lete i tidligere saker

Et tverrfaglig team ble derfor satt sammen for gjennomføre en Proof-of-concept (POC) for å teste ut følgende hypotese:

‘Kunstig Intelligens kan brukes som hjelp i komplekse saker i skatteforvaltningen’. 

Vi ønsket i denne POCen å finne ut om KI verktøy kan:

• korte ned saksbehandlingstiden
• hjelpe til med å finne relaterte saker som overlapper i problemstilling
• forenkle arbeidsprosessen
• bidra til informasjonsdeling og et løpende oppdatert kunnskapsgrunnlag
• bygge forståelse av hvordan KI kan brukes i lignende problemstillinger

I POC’en ble det etablert en Grunnrenteskattagent: Hvor et sett av kunnskapsagenter som er spesialisert etter kilde og rolle. I stedet for én universalagent sender vi oppgaven ut til flere agenter som i hovedsak skiller seg i instruksjoner og kildetilfang. Denne tilsynelatende lille variasjonen gir store utslag i hva som oppdages, hvordan tekst tolkes i lys av lovhierarkiet (lov til forskrift til forarbeider til etatspraksis/rundskriv) og hvilke funn som løftes frem for saksbehandler.

Alle agentene er kunnskapsagenter: de gjør ikke eksterne handlinger, men forbedrer kvaliteten på beslutningsgrunnlaget. Informasjon flyter mellom agentene gjennom strukturerte mellomprodukter (sitater, vurderinger, åpne spørsmål), noe som gir mer treffsikre og etterprøvbare svar. I praksis ser vi raskere treff på riktig hjemmel, bedre sporbarhet til kilder og mer konsistent bruk av praksis som er viktige forutsetninger for likebehandling.

Neste steg er å bygge inn varslingsmekanismer: dersom lovtekst endres eller etatspraksis oppdateres uten at saksbehandlernes håndbøker er ajour, skal agentene automatisk varsle et redaksjonsteam. Slik reduserer vi misforståelser og merarbeid for både brukere og saksbehandlere. Samlet sett markerer Grunnrenteskattagenten et første, konkret skritt mot mer kunnskapsdrevne og konsistente tjenester i Skatteetaten.

Municipal planning processes in Norway are characterized by complexity, lengthy processing times, fragmented information systems, and limited citizen participation. A typical zoning plan requires on average 856 days from initiation to approval, and total planning-related costs exceed NOK 2.3 billion annually. These inefficiencies reduce trust in public administration, limit participation, and slow the implementation of critical infrastructure. The PlanKI project aims to address these challenges by introducing an AI-augmented planning framework that integrates advanced geospatial analysis, legal reasoning, and agentic AI systems to support decision-making and stakeholder interaction.

At the core of PlanKI is a new class of multimodal geospatial models capable of integrating and reasoning over heterogeneous planning data, including textual regulations, geospatial vector and raster data, zoning plans, environmental information, and temporal records. Building on state-of-the-art GeoAI research such as AlphaEarth and PlanBERT, the project develops regulation-aware embeddings specifically adapted to Norwegian planning law. By embedding both legal and spatial semantics in a unified vector space, PlanKI enables advanced knowledge retrieval, compliance checking, and scenario simulation that go beyond the capabilities of traditional GIS systems.

Technically, PlanKI focuses on three main AI components: (1) multimodal embeddings that fuse geospatial layers, imagery, and legal text while preserving spatial and regulatory semantics, (2) reasoning modules that combine these embeddings with rule-based engines and knowledge graphs to produce transparent and traceable assessments, and (3) agentic AI interfaces that allow autonomous and interactive task execution within municipal planning workflows. These agentic components, implemented through Model Context Protocols, can retrieve relevant data, conduct preliminary compliance checks, and generate scenario analyses in response to user queries. They are designed to act as intelligent assistants for planners, elected officials, and citizens, providing explainable outputs, interactive visualizations, and adaptive reasoning.

The system will be piloted in Lindesnes municipality to assess its impact on processing time, planning quality, participation, and trust, and subsequently scaled through national infrastructures such as ePlanSak and FellesTjenestePlan. Scientifically, the project advances regulation-aware GeoAI, multimodal geospatial embeddings, and interactive agent frameworks for public decision-making. Practically, it enables more transparent, efficient, and participatory planning processes, contributing to the modernization of public-sector spatial governance in Norway.

Agentic AI is making headlines, but most business processes aren’t built for it. In industries like insurance, workflows are highly structured: if-this-then-that logic and deterministic rails still rule. Dropping a fully autonomous agent into that environment? It rarely works. What we see in production is a hybrid approach—agents embedded in flows, not replacing them. They handle the messy parts: classifying multi-page documents, extracting key fields, grounding answers—while the process engine enforces compliance and predictability.

The real challenge isn’t the technology; it’s the process. Many organizations still design for linear automation, which makes agentic orchestration hard. We’ll share what works: start with the process, keep deterministic rails for trust, and add human-in-the-loop where risk matters.

And yes, customers ask: “When will you have GPT 5?” as if the model name solves everything. Our answer: success isn’t about chasing the latest logo—it’s about choosing the right model for the right task, combining it with retrieval, policy layers, and benchmarking for cost and accuracy. Bigger isn’t always better. Integrated is.