Regulatory impact assessments (RIAs) are the backbone of evidence-based policy. They're also one of the most resource-intensive analytical tasks in government and regulated industries. AI can make them faster, more thorough, and more consistent, without removing the human judgement that makes them credible.
What RIAs Involve
A regulatory impact assessment analyses the potential effects of a proposed regulation or policy change. It typically covers:
- Problem definition. What issue does the regulation address? What evidence supports the need for intervention?
- Options analysis. What are the alternative approaches? Regulation, self-regulation, education, market-based mechanisms, status quo.
- Impact modelling. What are the expected costs and benefits of each option? Economic impacts, social impacts, environmental impacts, distributional effects.
- Stakeholder analysis. Who is affected? How? What are their likely responses?
- Evidence synthesis. What does existing research, data, and international experience tell us about the likely outcomes?
- Compliance and enforcement. How will the regulation be implemented and enforced? What are the compliance costs?
Each of these components requires extensive research, data analysis, and synthesis. A comprehensive RIA for a significant regulatory change can take months of analyst time.
3-6 months
typical timeline for a comprehensive regulatory impact assessment for a significant policy change
Source: NZ Productivity Commission, Regulatory Quality Report, 2025
Where AI Assists
Evidence Synthesis
The most immediate value: AI can rapidly synthesise large volumes of research, policy documents, and international precedent. Instead of an analyst spending weeks reading and summarising hundreds of documents, AI can process the corpus and produce structured summaries of relevant findings, organised by topic, methodology, and outcome.
This doesn't eliminate the analyst's reading. It prioritises it. The AI surfaces the most relevant evidence, identifies gaps, and highlights contradictions. The analyst focuses their limited time on the highest-value material.
Impact Modelling Support
AI assists with impact modelling by:
- Processing large datasets to identify baseline trends and patterns
- Running sensitivity analyses across multiple parameter ranges
- Identifying analogous regulations in other jurisdictions and their observed outcomes
- Flagging distributional effects that might be missed in aggregate analysis
The models remain human-designed. The AI accelerates the data processing and scenario analysis that feeds those models.
Stakeholder Analysis
AI can analyse stakeholder submissions, consultation responses, and public commentary to:
- Identify themes and patterns across hundreds or thousands of submissions
- Classify stakeholder positions and concerns by type, sector, and sentiment
- Surface minority perspectives that might be lost in volume analysis
- Track how stakeholder positions change across consultation rounds
This is particularly valuable in NZ's consultation processes, which can generate thousands of submissions on significant regulatory changes.
Consistency Checking
AI can check draft RIAs for internal consistency: are the options analysed against the same criteria? Do the cost estimates use consistent assumptions? Are the evidence citations current and relevant? Does the analysis address all required elements?
This quality assurance function catches errors and omissions that human reviewers might miss, especially in lengthy, complex documents.
See the Pattern
Here's what AI-assisted regulatory impact analysis looks like:
Loading demo...
The NZ Context
Government Expectations
NZ's regulatory management system requires RIAs for significant regulatory proposals. The Treasury provides guidance on RIA quality, and the Regulatory Impact Analysis Team (RIAT) assesses RIA quality for Cabinet proposals. AI-assisted RIAs need to meet the same quality standards as manually produced RIAs.
This means the AI's contribution must be transparent. If AI was used to synthesise evidence, that should be noted. If AI-generated analysis was included, it should be identifiable and verifiable. The credibility of the RIA depends on the analyst's judgement, not the AI's output.
Te Tiriti Considerations
Regulatory impact assessments in NZ must consider impacts on Māori. AI systems assisting with RIAs need to:
- Include Māori-specific data sources and perspectives in evidence synthesis
- Identify differential impacts on Māori communities
- Respect Māori data sovereignty principles in data processing
- Avoid algorithmic biases that might underestimate or mischaracterise impacts on Māori
This requires specific configuration and training data. Generic AI systems without NZ-specific tuning will miss these dimensions.
Cross-Jurisdictional Analysis
NZ frequently looks to international experience when assessing regulatory options. AI excels at cross-jurisdictional analysis: scanning regulatory databases across multiple countries, identifying relevant precedents, and summarising observed outcomes. For a small country that can't afford to make every policy mistake itself, this capability is particularly valuable.
Implementation Approach
Start with Evidence Synthesis
The lowest-risk, highest-value starting point. Deploy AI to assist with literature reviews and evidence synthesis for RIAs. This requires:
- Access to relevant research databases and policy repositories
- Configuration for NZ-specific regulatory context
- Quality review processes for AI-generated summaries
- Clear documentation of AI's role in the evidence synthesis
Add Submission Analysis
For RIAs that involve public consultation, add AI-assisted analysis of stakeholder submissions. This requires:
- Natural language processing capability for NZ English (including te reo Māori terms)
- Classification frameworks for stakeholder positions and concerns
- Aggregation and reporting tools for analysis results
- Human review of AI-generated analysis before inclusion in RIAs
Support Impact Modelling
The most advanced application: AI support for impact modelling. This requires:
- Access to relevant datasets (economic, social, environmental)
- Integration with existing modelling tools and frameworks
- Validation processes for AI-generated analysis
- Clear delineation between AI-assisted and human-directed analysis
Governance Requirements
AI-assisted RIAs need governance that preserves the credibility of the assessment:
Transparency. The RIA should clearly indicate where AI was used and how. "Evidence synthesis assisted by AI" is acceptable. Undisclosed AI use undermines trust.
Verification. AI-generated content must be verifiable. Evidence citations must be checked. Statistical claims must be confirmed. Stakeholder analysis must be validated against source submissions.
Accountability. The analyst and the policy team remain accountable for the RIA's content and conclusions. AI is a tool, not a co-author.
Quality standards. AI-assisted RIAs must meet the same quality standards as manual RIAs. AI assistance should improve quality, not provide a shortcut to lower standards.
Regulatory impact assessment is exactly the kind of analytical work where AI adds genuine value. Large volumes of evidence to synthesise. Complex scenarios to model. Thousands of stakeholder submissions to analyse. The work is rigorous, time-consuming, and important. AI makes it faster and more thorough. The analyst's judgement remains the core of credible policy analysis.