Short answer
AI procurement questions are what a buyer asks an AI supplier in writing before signing, so the contract records how data is used, where it is processed, how accuracy is shown, what it costs at scale and how you leave. This list has 30 questions in six groups for New Zealand buyers, with what a good answer looks like, contract clauses to request and red flags.
- Get answers in writing and attach them to the contract.
- Storage location and processing location are different questions.
- Ask for breach notice fast enough to meet Privacy Act obligations.
- Ask how much notice you get before a model is changed or retired.
Use these questions once you have a shortlist, alongside the AI vendor evaluation checklist that scores the answers. Send them in one document, ask for written replies, and treat “we take privacy seriously” as no answer at all. Delete questions that do not apply to a simple subscription tool; keep all of them for a supplier building or running a system for you.
The procurement questions
A. Data use and privacy
| # | Question | A good answer |
|---|---|---|
| 1 | On the plan we would buy, are our inputs, outputs or files used to train or improve any model? | No, by default, stated in the terms for that plan |
| 2 | Do you process our information only on our behalf, or also for your own purposes? | Only on our behalf, in a data processing agreement |
| 3 | How long do you keep prompts, outputs, files and logs? Can we change that? | Specific periods; admin controls; deletion on request |
| 4 | Which staff and sub-processors can access our data, and why? | A named list; access limited and logged |
| 5 | How will you help us answer a person’s request to see or correct their information? | Search and export tools, with a stated turnaround |
Question 2 matters under New Zealand law. A provider that only stores or processes information for you is treated as your agent, and you remain responsible; one that uses it for its own purposes, such as training, is treated as holding it too (OPC).
B. Location
| # | Question | A good answer |
|---|---|---|
| 6 | In which countries is our data stored? | A list, including backups |
| 7 | In which countries is it processed by the model? | A list, separate from storage |
| 8 | Can we restrict processing location, and at what cost? | Options and prices stated |
| 9 | Which model providers do you use, and will you tell us before changing them? | Named providers; written notice |
| 10 | Which sub-processors are offshore? | A list with countries |
Storage and processing differ: OpenAI’s Australian data residency stores data in Australia but does not process it there (OpenAI).
C. Security
| # | Question | A good answer |
|---|---|---|
| 11 | Do you support single sign-on, multi-factor authentication and immediate user removal? | Yes to all three |
| 12 | Which security standards do you follow, and can we see the evidence? | A certificate or audit report |
| 13 | How do you test for prompt injection and leakage between customers? | A described, regular testing process |
| 14 | What audit logs can we see and export? | User, action and time, exportable |
| 15 | How fast will you tell us about a breach affecting our data? | A fixed number of hours, with details we need |
For question 15, you must notify the Privacy Commissioner of a serious-harm breach as soon as practicable, and the Commissioner asks for this ideally within 72 hours (OPC). Your supplier’s notice period has to leave you time to assess and notify. For a deeper check of a supplier’s security answers, see our AI security reviews.
D. Accuracy and evidence
| # | Question | A good answer |
|---|---|---|
| 16 | Can we test on our own examples before signing? | Yes, with a trial or paid pilot |
| 17 | What accuracy have you measured on work like ours, and how? | Method, sample size and error types |
| 18 | How does the system show sources so we can check outputs? | Citations or links to source text |
| 19 | What happens when the system is unsure? | It flags or escalates, rather than guessing |
| 20 | What evidence supports your marketing claims? | Test results you can share |
The Commerce Commission says claims must be substantiated when they are made (Commerce Commission).
E. Cost
| # | Question | A good answer |
|---|---|---|
| 21 | What will we pay in year one and year two at our expected volume? | A worked figure, not a starting price |
| 22 | Is pricing per seat, per use, per outcome or in credits? What happens at the limit? | Clear units; overage rates stated |
| 23 | Are prices in NZD or USD, and is GST included? | Both stated |
| 24 | What minimum seats, terms or commitments apply? | Stated in the order form |
| 25 | How much notice do you give of price changes? | At least one billing cycle, in writing |
F. Change, support and exit
| # | Question | A good answer |
|---|---|---|
| 26 | How much notice do you give before changing or retiring a model? | Weeks or months, with time to re-test |
| 27 | What support hours apply in New Zealand time? | Hours and response times stated |
| 28 | Can we export our data, prompts, configuration and logs? | Yes, in standard formats |
| 29 | What happens to our data when we cancel? | Deleted within a stated period, with confirmation |
| 30 | What happens if you are acquired or close? | Notice, export window and contract continuity |
Model changes are routine: Anthropic notified developers on 30 September 2026 that Claude Sonnet 4.5 retires from its API on 30 November 2026 (Anthropic). A supplier built on a retiring model has to switch, and you need time to re-test.
Contract clauses to request
The joint small-business guide from NCSC New Zealand and Australian partners says business contracts should include clear clauses on data protection, breach notification and service uptime (NCSC). Ask for:
- No use of our data for training or product improvement.
- Processing only on our documented instructions.
- Data locations and a sub-processor list, with notice of changes.
- Breach notification within a fixed number of hours.
- Service levels for availability and support.
- Notice before model, feature or price changes.
- Data export and deletion on exit, with written confirmation.
How to use the questions
- Pick the questions that apply. All 30 for a supplier running a system for you; fewer for a subscription tool.
- Send them in one document. Ask for answers in writing by a date.
- Score the answers. Use the scoring sheet in the vendor evaluation checklist.
- Attach answers to the contract. Promises that matter should become clauses.
Filled-in example
Illustrative example: an Auckland accounting firm compares two document-extraction suppliers’ written answers.
| Question | Supplier 1 | Supplier 2 | Follow-up |
|---|---|---|---|
| 1. Training | “No; see clause 4.2 of our data processing agreement” | “We take privacy seriously” | Supplier 2 must answer for the specific plan |
| 7. Processing | “United States and Australia; list attached” | “Your data stays in Australia” | Ask Supplier 2 where the model runs, not just storage |
| 15. Breach notice | “Within 24 hours of confirmation” | “Promptly” | Ask Supplier 2 for a fixed number of hours |
| 26. Model changes | “60 days’ notice; we re-run your test set” | Not answered | Supplier 2 is a risk until answered |
Supplier 1’s answers became contract clauses. Supplier 2 was given a week to answer in writing.
Red flags and common mistakes
- Asking in a meeting. Answers given on a sales call are hard to rely on later; get them in writing.
- General policies instead of answers. A link to a privacy policy does not tell you what applies to the plan you would buy.
- Storage dressed up as processing. “Your data stays in Australia” may only mean storage.
- Accuracy without method. A percentage with no sample size or test description is marketing.
- No exit. No export option, or export only on request at a fee, means lock-in from day one.
Next step
Turn the questions that matter most into written requirements with writing AI vendor requirements, and compare subscription prices on AI tools pricing for New Zealand businesses. For location choices, see data residency for AI systems.
Sources
- Artificial intelligence for small business, ASD’s ACSC with NCSC New Zealand and COSBOA, 14 January 2026
- Working with third-party providers, Office of the Privacy Commissioner, 26 November 2024
- Making accurate claims, Commerce Commission, accessed 19 October 2026
- Data controls in the OpenAI platform, OpenAI, accessed 19 October 2026
- Model deprecations, Anthropic, accessed 19 October 2026
- Notify us of a privacy breach, Office of the Privacy Commissioner, accessed 19 October 2026
Written by the Levy Street team in Wellington. Last updated 19 October 2026.