Independent, vendor-neutral help with government AI projects: deciding where AI is worth using, choosing a platform, writing the requirements, and testing the answers before people rely on them. I advise, specify and test; your team or your vendor builds.
- For
- New Zealand government agencies, councils and Crown entities. Private-sector and non-profit organisations are welcome too.
- You get
- A use-case review, platform evaluation, requirements or answer testing, and the methods to run it again
- Typical time
- Four to six weeks for a single piece of work, or alongside your team for longer
- Price
- A fixed quote agreed before the work starts, or a day rate for longer help
Where I can help
Answer test report
Council assistant, release 6
Summary 104 of 120 answers were correct (87%), up from 77% at release 5. Three were wrong, one more than the agreed limit of two, so release 6 should wait until they're fixed.
| Result | Answers | What it means |
|---|---|---|
| Correct | 104 | Matches the council's pages and links to the right one |
| Incomplete | 11 | Right, but leaves out a condition that matters |
| Wrong | 3 | Contradicts the council's pages |
| Declined | 2 | Said it didn't know, and the pages don't cover it |
Wrong answers and who fixes them
| Question | What went wrong | Who fixes it |
|---|---|---|
| My bin wasn't collected. What now? | Reversed the 3pm rule for missed collections | Content team |
| How much is a resident parking permit? | Quoted last year's fee from an archived page | Search index |
| Do I need a consent for a deck? | Said no, leaving out the height condition | Prompt |
What I do
- Use cases A short list of places to use AI, ranked by how much they help the people using the service and how much a wrong answer would cost. Sometimes the answer is to fix the content first.
- Choosing a platform or vendor A vendor-neutral comparison on your own content and real test questions, so the decision rests on evidence, not demos.
- Requirements What the feature should answer, what it shouldn't, and how you'll know it works, written so it can be built in-house or bought.
- Testing the answers Real questions, judged the way people would judge them, with each answer checked against its source. It's the method I used at MBIE when checking AI Overviews for accuracy.
- From pilot to everyday use Deciding what good looks like, getting the content and structure ready for AI to answer from, and making sure the people who'll support it are part of the decision. That includes who checks the answers and how often, what happens when one is wrong, and what to say publicly about how the tool is used.
- AI search on your website Search roadmaps, requirements for AI-assisted search, and platform evaluation.
- Staff use of AI Where tools like Copilot, ChatGPT or Claude help staff, guidance people can follow, and a way to check what the tools produce before anyone relies on it.
Each area in more detail, with the work behind it:
What you get
- Use-case review. A ranked shortlist, with the benefit, the risk of a wrong answer and what it would take for each.
- Platform evaluation. The test questions, the scored results for each platform and a recommendation you can defend in procurement.
- Requirements. A document your team or vendor can build or buy against.
- Answer testing. The test set, the results, the content fixes behind wrong answers, and a method your team can run again.
What I need from you
- One person who owns the project and can make decisions.
- Real questions people ask, from search logs, Search Console or the contact centre.
- Access to the content and, where relevant, to the tool or the vendor.
- A short session with the people who will support the tool after launch.
How I work
- Vendor-neutral when comparing tools and platforms.
- Decisions backed by test questions and evidence, not demos.
- Plain English that leadership teams can act on.
- Methods and documentation that stay with your team.
- Privacy, accessibility, te ao Māori and public trust from the start, not as a final check, in line with the Public Service AI Framework.
How it works
- A short call about the project and where it's stuck.
- A written quote: a fixed scope and price for a single piece of work, which typically takes four to six weeks, or a day rate for longer help.
- I work alongside your team, your vendor and the people who'll support it.
- You keep the methods, test questions and documents.
Common questions
Are you tied to any AI vendor?
No. I compare platforms vendor-neutrally, using your own content and the questions your users really ask, not the vendor's demo.
Do you build the AI?
No. I help you decide what it should do, write the requirements, choose the platform and test the answers; your team or your vendor builds it. I know how these tools are built, so I can work closely with the people building them.
How does this fit New Zealand government expectations on AI?
The work aligns with the Public Service AI Framework and the GCDO's guidance on generative AI: human oversight, being open about where AI is used, and checking answers against trusted sources. Privacy, accessibility, te ao Māori considerations and public trust are part of the work from the start, and it fits alongside your privacy impact assessment and any Māori data considerations. Councils aren't bound by the central guidance, but it's a sound reference and I work with it.
Do you use AI in your own work?
Yes, for drafting and analysis. Every finding, figure and source is checked by me before it reaches you, and I'll tell you where AI tools were used.
Can you work with us for longer than one piece of work?
Yes. I can work alongside your team for a project or longer, on a day rate or a fixed price.
How does procurement work?
We start with a short call about what you need. I then send a written quote with the scope and a fixed price, agreed before any work starts. A single piece of work is sized to fit within most agencies' single-quote limits, and your procurement team confirms the purchasing route and any supplier requirements. Longer work alongside your team is quoted separately.
How is this related to the findability audit?
AI answers are only as good as the content behind them. The audit checks what public AI tools already say about you; this work is for AI your organisation runs itself.
Talk about your AI project
Send a few lines through the contact form or call (027) 450 5720.