Filter articles
NZ Parliamentary Counsel Office explores legislative drafting with AI
22 Jul 2025, 3:06 PMLearn how AI strategists and engineers at Catalyst supported the NZ Parliamentary Counsel Office in their research and development programme exploring the use of AI.
Background
The Parliamentary Counsel Office (PCO) drafts all of New Zealand's legislation. It recognised that artificial intelligence (AI) might be able to support staff to speed up legislative drafting and improve the experience for people using New Zealand's legislation website.
PCO launched a six-month research and development programme, partnering with five different companies to explore five AI proof-of-concepts across their operations. The kaupapa of the project focused on research and development. The PCO team wanted to understand the potential benefits, risks, challenges, and opportunities of AI before making any implementation decisions.
The challenge
Within this broader initiative, PCO had identified a potential opportunity: writing clause-by-clause explanatory notes for new bills. These notes provide legislators with an "at a glance" summary of each clause's meaning and proposed changes. Currently, drafters manually summarise every legislative change, can be a time-intensive process requiring deep legal expertise.
The use case for this project was: “As a drafter, I want to have the first draft of an explanatory note generated for me. It should be accurate and relevant in its summary of what the provisions of a Bill or the Bill itself means.”
Catalyst explored a solution that could:
- Produce human-quality first drafts meeting their strict guidelines.
- Generate drafts accurate enough for legal experts to review efficiently.
- Achieve data sovereignty (meaning no data is sent to overseas AI providers or subject to foreign data retention laws) due to handling sensitive draft legislation.
- Integrate with existing workflows without forcing operational changes.
R&D approach to identifying the right AI solution for PCO
Our team of AI strategists and engineers started by mapping PCO's internal processes to understand exactly how their drafters worked. This foundation enabled us to design a proof of concept solution that fitted their workflow versus forcing them to change how they operated.
We then worked with PCO to run three targeted experiments that would answer key questions:
Can AI produce quality first drafts for human review?
Initial attempts at getting AI to review entire bills showed room for improvement, with output still requiring significant human review. Giving the AI clearer instructions and guidelines worked much better for simpler amendments, but larger, more complex bills still needed work.
The breakthrough came when we broke down the task. Instead of asking AI to handle entire bills, we processed each clause individually. This approach removed the limitations and enabled PCO to effectively process any type of legislation.
What’s the best AI model and approach?
We tested AI models of various sizes (from smaller 8B to larger 70B versions), and determined the trade-offs between accuracy and cost.
Larger models gave better responses but also raised an important concern: system bias. This is where people might trust the AI output too much. To address this, we built in ways for drafters to compare multiple options and maintain their critical judgement when reviewing drafts.
To find the right balance for PCO, we moved to fine-tuning a small 8B model, which was trained on 200 examples of New Zealand legislation. Fine-tuning is the process of training a large language model (LLM) on data relevant to a specific task or field of expertise. This helped PCO optimise the outputs while opting for a small LLM.
The solution
Catalyst’s AI experts built an AI proof-of-concept for PCO that handles the complex structure of legal documents automatically, organising content to provide the right context for AI processing. The solution also addressed concerns about over-reliance on AI by creating pathways for staff to compare multiple options and maintain critical judgement throughout the process.
The proof of concept solution was deployed on Catalyst Cloud, NZ's sovereign cloud infrastructure with AI computing power hosted solely in Aotearoa.
The results
The structured approach delivered measurable improvements at each stage:
- Initial approach: Submitting entire bills produced output that didn’t streamline workflows enough.
- Structured prompting: Specific guidelines and guardrails showed significant improvements on simpler amendments.
- Prompt-chunking: Breaking down tasks to handle each clause separately removed limitations and enabled the processing of any legislation type.
The proof of concept demonstrated that AI could produce meaningful first drafts that PCO staff could efficiently review and refine, which might enable expert drafters to focus on deeper legal analysis rather than repetitive summarisation.
The Parliamentary Council Office's principled R&D approach sets the standard for innovation excellence. Its collaboration with New Zealand's local tech sector demonstrates how to drive meaningful technological progress in Aotearoa for Aotearoa.
PCO's thorough due diligence, its approach to transparent innovation, and commitment to open-sourcing research sets a benchmark we advocate for being more widely adopted.
We demonstrated how Drafters might:
- Upload legislative bills and generate notes for individual clauses or whole sections.
- Compare AI output with the official text and export results for review.
- Feel confident that any sensitive data, including draft legislation, remains hosted in NZ only.
- Ensure human-in-the-loop and critical judgement is maintained while also automating repetitive tasks.
The model-agnostic approach means PCO isn't dependent on any single AI provider, giving them freedom to choose the best options as technology changes while maintaining digital sovereignty and workflow flexibility.
Considering how AI can help your organisation?
Catalyst is a leader in open source technology, focusing on helping clients design and build solutions that achieve their goals, remove vendor lock-in and licensing fees, and provide more freedom to innovate. Our AI strategists and engineers can help you create better AI solutions for your organisation, including:
Content analysis & processing
Building AI solutions that ingest, evaluate and reason through highly specialised structured documents (like XML). Applications include automated moderation, compliance checking, evaluation, quality assurance and content tagging.
Secure & sovereign infrastructure
Catalyst’s AI solutions are ideal for customers prioritising digital independence: improving data security, removing vendor lock-in and licensing fees, using NZ-built open-source models and infrastructure to better protect their data.
Custom AI solutions
End-to-end development including AI consulting, requirements analysis, mapping internal processes, prompt engineering, complex data parsing for model fine-tuning, and front-end UI development.
Process automation
Tailored AI-augmented workflows to streamline specialised business processes, reduce manual effort and unblock critical operations.
Related projects
The curious users ultimate guide to asking better AI questions
This guide covers seven essential questions to ask when using GenAI, though most considerations apply to other AI systems as well.
The basics of artificial intelligence (AI)
If you're new to artificial intelligence (AI), or want to revisit some of the basics about this technology, this article is for you.