In April 2026, the Minister of Justice asked the New Zealand Law Commission to undertake a review of legal issues related to the use of automated decision-making by government. The announcement was measured in tone. The implications are not.

New Zealand currently has no overarching law, standards, or guidance specifically addressing how government agencies should use automated decision-making in a legally compliant and consistent manner. The Law Commission President, Dr Mark Hickford, framed the absence plainly: the lack of a clear framework creates risks of poor practice and Crown liability, and an unnecessary risk aversion that can discourage responsible innovation.

That is a government body formally acknowledging that AI-driven decision-making has outpaced the legal architecture designed to govern it.

Employers should read that acknowledgement carefully. Because the gap does not belong only to government.

The review and what it signals

The Law Commission will begin this project around the middle of 2026. Public consultation is expected later this year, with a final report anticipated in 2027. If the Commission concludes that law reform is necessary — and the framing of its announcement suggests it is already thinking in that direction — legislative proposals will follow.

The review covers automated decision-making by government. But the legal issues it will examine — accountability, procedural fairness, transparency, judicial review, and the proper limits of algorithmic authority — are not confined to the public sector. They are the same issues playing out right now inside private organisations.

AI tools are already being used in hiring, performance management, disciplinary triage, restructuring analysis, and workplace investigations. In most organisations, they are being used without a governance framework, without documentation standards, and without a considered view of what happens when those decisions are challenged.

When the state admits its own framework is inadequate, it is signalling something broader: that AI-assisted decision-making, across all sectors, is ahead of the law.

Why this matters to employers, not just government

The Employment Relations Act 2000 requires employers to act as a fair and reasonable employer would in the circumstances. The Pathfinder test under section 103A asks whether the employer's actions — including the investigation, the decision-making process, and the outcome — were substantively and procedurally justified.

That test does not become more forgiving when an algorithm is involved. It becomes more demanding.

Consider what an AI-assisted employment decision actually looks like in practice. A recruitment screening tool filters candidates by scoring résumés against criteria it has learned from historical hiring data. A performance management system flags patterns in productivity metrics and generates a recommendation. An AI tool drafts the letter notifying an employee of a disciplinary outcome based on investigation notes entered by a manager.

In each case, a human is nominally in the loop. But if that human cannot explain what the AI did, cannot produce the inputs it worked from, and cannot demonstrate that they exercised genuine independent judgment before acting, the procedural fairness analysis becomes very difficult to satisfy.

The Employment Relations Authority does not yet have a body of case law specifically addressing AI-assisted employment decisions. What it does have is a well-developed jurisprudence on process, documentation, and the reconstruction of decision-making. Those principles apply fully — and they raise serious questions about how most organisations are currently operating.

The ERA does not ask whether AI was involved. It asks whether the decision was fair and reasonable. If you cannot reconstruct what the AI did and why a human accepted it, you may not be able to answer that question.

The five questions an ERA member would ask

When a grievance reaches the Authority involving an AI-assisted employment decision, the inquiry will likely come back to a small number of questions. They are worth considering now, before the decision is made.

Five questions the ERA is likely to ask
1
What did the AI actually do?
Not what the tool is marketed to do — what did it specifically do in this case. What inputs did it receive. What output did it produce. What version of the model was running at the time.
2
Who reviewed it, and how?
Was there a human reviewer, or was the AI output accepted unchecked. If there was a reviewer, what did they actually assess. Did they have the information needed to make a genuine independent judgment.
3
Was the employee told?
Good faith under section 4 of the ERA requires active and constructive engagement. An employee who is not informed that AI was used in a process that affects their employment has been denied something material.
4
Can the decision trail be reconstructed?
Employment records must be retained for seven years under section 130 of the ERA. If the AI tool generates outputs that are not retained, or if the human reviewer does not document their analysis, the trail is broken at the point it matters most.
5
Was the AI output accurate?
AI tools can produce outputs that are confidently wrong. If a dismissal, a performance assessment, or a restructuring recommendation was based on AI output that contained an error, and that error was not caught before the decision was made, the employer carries that consequence.

Recruitment, investigations, and the highest-risk domain

Not all AI uses carry equal risk. Based on the Lex Praxis AI Employment Risk Audit methodology, disciplinary processes represent the highest-risk domain — followed closely by performance management and workplace investigations.

The reason is the interaction between AI capability and procedural consequence. Disciplinary decisions carry significant exposure: reputational damage, economic loss, and the prospect of proceedings before the Authority. They also require careful, documented process. AI tools operating in this space without oversight create an almost certain audit trail failure.

Recruitment sits in a different risk category, but with its own exposure. The Human Rights Act 1993 prohibits discrimination in employment on specified grounds. A recruitment AI trained on historical data can embed and perpetuate discriminatory patterns without any individual intending that outcome. The MBIE Responsible AI Guidance for Businesses (July 2025) documents exactly this scenario — a CV screening tool disadvantaging female candidates, generating Human Rights Act liability. That scenario is not hypothetical. It is the government's own illustration of where AI-assisted hiring can lead.

The Law Commission review will eventually produce a framework. That framework will reflect principles employers are already subject to. The question is whether those employers are governing their AI use before the framework arrives, or after the first Authority determination requires them to.

What good governance looks like now

The Law Commission's review will take time. The final report is expected in 2027. Legislative change, if any follows, will take longer still. None of that is a reason to wait.

Employers using AI in employment decisions can act now. The practical foundation is not complicated.

These are not burdensome obligations. They are the minimum conditions for being able to defend a decision later. The employers who build these systems now will face the Law Commission's eventual framework from a position of readiness. The employers who do not will face it — or an Authority determination — from a position of exposure.

A closing observation

The Law Commission review is not a warning that the law is about to change. It is a signal that the law is already struggling to keep up with what is already happening. The gap between AI capability and legal governance is a present reality, not a future risk.

Employers who understand that now are better positioned than those who discover it later.

— · —