New Zealand's Parliamentary Counsel Office ran a six-month, five-partner R&D programme fine-tuning an 8-billion-parameter model on 200 examples of NZ legislation to test AI for drafting explanatory notes, consolidating Acts and answering plain-language legislation queries.
~200 examples of NZ legislative text
Training examples used to fine-tune model
8 billion parameters
Model size
Details
Maturity
Pilot
Promoter
Parliamentary Counsel Office (PCO), New Zealand
Period
2024–2025 (pilot)
Keywords
legislative drafting, generative AI, digital government, parliamentary counsel
Context
New Zealand's Parliamentary Counsel Office (PCO) ran a six-month pilot testing generative AI for legislative drafting work, fine-tuning an 8-billion-parameter open language model on roughly 200 examples of New Zealand legislative text rather than adopting a large third-party model off the shelf. Five technology companies, including Catalyst IT, collaborated on the trial, with models hosted on Catalyst Cloud, New Zealand's sovereign cloud infrastructure, to keep legislative data within national jurisdiction.
Objectives
Test whether generative AI could assist legislative drafters with explanatory notes, prospective consolidations of amended Acts, plain-language drafting, and a prototype chatbot for the New Zealand Legislation website — while keeping sensitive legislative data under national data-sovereignty controls.
Activities
Fine-tuned an 8B-parameter open model on ~200 NZ legislative text examples; tested the model on drafting explanatory notes for bills and generating prospective consolidations of amended Acts; prototyped a legislation-website chatbot; deployed on Catalyst Cloud sovereign infrastructure with five technology-company partners.
Results
PCO published transparent, self-critical results pages with experiment-specific assessments rather than only promotional summaries; the practice scored 53/100 in the catalogue's evaluation (top 57% among 246 scored practices). No quantified accuracy, time-saved, or cost-saved metrics were reported.
Conclusions
The pilot remained an exploratory six-month trial with no production rollout or measured public-value outcomes reported yet; PCO's transparency about limitations is itself notable, but the evidence base is a self-published descriptive account rather than an independent evaluation.
Implementation
Indicative cost
Low (< €50k) — Small-scale six-month R&D pilot: fine-tuning an 8B-parameter open model on ~200 legislative text examples and running it on sovereign cloud infrastructure — low cost relative to large third-party LLM licensing or major IT programmes, though exact budget figures were not published.
Time to results
Short (< 1 year) — Described as a six-month research-and-development pilot in 2024-2025, with no confirmed production deployment date.
Staffing & skills
Parliamentary Counsel Office (PCO) legislative drafters, Five technology company partners, including Catalyst IT
Conditions for success
Fine-tuning a small (8B-parameter) open model on domain-specific legislative text rather than relying on large third-party models
Deploying on sovereign cloud infrastructure (Catalyst Cloud) to keep legislative data within national jurisdiction
Publishing transparent, self-critical experiment-level results rather than only promotional summaries
Common failure modes
Pilot concluded without a production rollout or measured public-value outcomes
Where it fits
Governance type
national legislative drafting office
Scale
single-agency pilot
Income level
high income
Commonly funded by
National / regional programmes
Indicative funding routes for practices of this type — always check each programme's current calls and eligibility rules.
Do you run this practice?
Claim it —
verified implementers get a public contact pathway and can propose corrections.
Data sources
Where this practice's information was retrieved from, and when.
South Korea's National Assembly deployed an on-premises generative-AI platform built by Samsung SDS — combining a custom LLM, retrieval-augmented generation …