Responsible AI for small NZ teams: a five step workflow

In 2025 the New Zealand government published two documents that matter to anyone shipping AI features: a national AI strategy, and Responsible AI Guidance for Businesses from the Ministry of Business, Innovation and Employment (MBIE). Both are written in plain English, and both point toward the same habit: start small, measure, and keep people informed. This post turns that habit into a workflow a small team can follow this week.

What the guidance encourages

MBIE's guidance is aimed at businesses of every size and is voluntary. It encourages organisations to adopt AI in ways that are safe, transparent and useful, and it leans on existing New Zealand law such as the Privacy Act 2020 and the Fair Trading Act. For a product team, that translates into four practical habits:

1. Pick a specific job for the AI to do.
2. Test it on your own examples before launch.
3. Tell users when AI is involved.
4. Keep a person able to step in.

Step 1: choose one narrow job

"Add AI to our app" is too wide to test. "Draft a reply to a booking enquiry in our brand voice" is a job you can score. Write the job as one sentence and list three to five inputs you would expect, including a few awkward ones such as an empty message or a message in te reo Māori.

Step 2: build a small test set

Collect 20 to 50 real examples from your own business, with personal details removed. Next to each, write what a good answer contains. This becomes your eval set. The evals pages on app.nz explain the approach of combining several kinds of evidence, and the same idea works at home: objective checks where possible, human review where judgement matters.

A simple scoring script is enough to start:

def score(answer: str, must_include: list[str]) -> float:
    hits = sum(1 for phrase in must_include if phrase.lower() in answer.lower())
    return hits / len(must_include)

Run it each time you change the prompt or the model, and keep the results in your repository next to the code.

Step 3: compare models on your examples

Public leaderboards give a useful first shortlist, and your own test set gives the final answer. The model directory lists the models available through the app.nz gateway, and because the endpoint follows the OpenAI format, switching between them is a one line change to the model field. Try two or three candidates, record quality and cost per request, and choose the cheapest one that meets your bar.

Step 4: tell people, and keep a human in the loop

Add a short line wherever AI output appears, such as "Drafted with AI, reviewed by our team". For decisions that affect someone's money, health or job, route the draft to a person for approval. This protects customers and also gives you a steady stream of corrections to add to your test set.

Step 5: review monthly

Models and prompts change, and so do your customers' questions. Put a recurring thirty minute review in the calendar: rerun the test set, read a sample of real outputs, and note anything to improve. Small and regular beats large and rare.

Where to begin

Pick the narrowest useful job in your product, write ten examples, and see how a model handles them. You can try prompts straight away in the gateway playground and keep your prompt and test files in a repository hosted on app.nz. The official sources are the MBIE guidance and the national AI strategy on the MBIE and government websites, so read them directly for the full detail.