Short answer
An AI data inventory lists every source of information an AI system reads, stores or produces, with its owner, sensitivity, whether it holds personal information, where it is processed and stored, and how long it is kept. This template has 15 columns, a four-level sensitivity scale, and the Privacy Act 2020 questions New Zealand organisations need answered, including IPP 3A and offshore processing.
- Inventory information, not just systems: prompts, outputs and logs count too.
- Record where each provider processes and stores data, and for how long.
- Note whether the provider uses data for its own purposes; that changes your privacy position.
- Flag Māori data and any information collected indirectly.
Build the inventory for each AI system before it goes live, and keep it updated as sources change. A spreadsheet is enough. The person who owns the workflow fills in the business columns; the privacy officer checks the privacy columns.
The AI data inventory template
| # | Column | What to record |
|---|---|---|
| 1 | ID | D-01, D-02 and so on |
| 2 | Data source | The specific information: “donor records”, not “the CRM” |
| 3 | System | Where it lives: Xero, SharePoint, CRM, shared inbox |
| 4 | Owner | The person accountable for the information |
| 5 | Example content | A typical record, described without real details |
| 6 | Personal information | None; staff; customers; children; health, financial or other sensitive information |
| 7 | Sensitivity | Public, internal, confidential or restricted (scale below) |
| 8 | AI system using it | Which tool or system reads, stores or produces it |
| 9 | Use | Read; stored by the AI system; produced as output; kept in logs; used for training |
| 10 | How collected | Directly from the person, or from another source (IPP 3 or IPP 3A) |
| 11 | Where processed and stored | Provider and countries, for both processing and storage |
| 12 | Provider’s own use | Does the provider use it for its own purposes, such as training? Yes or no, with the source of the answer |
| 13 | Retention | How long it is kept by you and by the provider, including logs |
| 14 | Access | Who can see it, inside and outside the business |
| 15 | Quality notes | Current, complete, duplicated? Known gaps? |
Sensitivity scale
| Level | Examples | AI rule |
|---|---|---|
| Public | Published web pages, brochures, public standards | Any approved tool |
| Internal | Procedures, templates, de-identified notes | Approved tools on business plans |
| Confidential | Client files, contracts, customer personal information | Only tools approved after a privacy impact assessment |
| Restricted | Health, financial account, identity documents, information about children, Māori data held on behalf of iwi or hapū | Named systems only, with director approval and extra controls |
Privacy questions for each source
- Is collecting this information necessary for the purpose of the AI use (IPP 1)?
- If it was collected from someone other than the person, have we told them, or does an exception apply (IPP 3A, from 1 May 2026)?
- Is it protected with reasonable security, including in prompts and logs (IPP 5)?
- Will we check it is accurate before using it, especially if the AI produced it (IPP 8)?
- Is it kept no longer than needed, including in provider logs (IPP 9)?
- Is the AI use consistent with the purpose it was collected for (IPP 10)?
- Does the provider only process it on our behalf (section 11), or use it for its own purposes (IPP 11, and IPP 12 if offshore)?
- Have we considered Māori perspectives and data sovereignty for any Māori data?
On section 11, the Privacy Commissioner’s guidance says that when a provider only stores or processes information for you, your organisation remains fully responsible, and gives AI training as an example of a provider using information for its own purposes (OPC). The Commissioner also treats inaccurate or made-up information about a person as personal information (OPC), so AI outputs about people belong in the inventory.
For Māori data, MBIE’s business guidance points to the Te Mana Raraunga principles (MBIE), which say Māori data should be stored in Aotearoa New Zealand whenever possible (Te Mana Raraunga).
How to build the inventory
- Start from the workflow map. Every input and output in your workflow map is a candidate row.
- Add what the AI system keeps. Chat histories, generated documents, embeddings in a search index and system logs.
- Get provider answers in writing. Use published terms and data processing agreements, and record the link or document.
- Classify sensitivity. When in doubt, choose the higher level.
- Answer the privacy questions. Do this with your privacy officer, and record the reasoning. Staff guidance is in protecting customer information when using AI.
- Review quarterly and on change. New sources, new tools or changed provider terms trigger an update.
Filled-in example
Illustrative example: a Wellington charity with 2,000 donors plans AI-drafted donor thank-you letters, a version of donor communication drafting.
| ID | Source | Personal information | Sensitivity | How collected | Where processed | Provider’s own use | Retention |
|---|---|---|---|---|---|---|---|
| D-01 | Donor records (name, gift, history) | Donors | Confidential | Directly | API provider, offshore | No, per API terms | Provider deletes API inputs within 30 days (Anthropic) |
| D-02 | Programme impact stories | Beneficiaries, some children | Restricted | Directly, with consent forms | Not sent to AI | n/a | n/a |
| D-03 | Letter templates | None | Internal | Our own | API provider | No | 30 days |
| D-04 | Generated letters and logs | Donors | Confidential | Produced by AI | Our CRM and system logs | n/a | 7 years with donation records; logs 90 days |
The inventory changed the design: impact stories about children (D-02) were kept out of the AI system entirely, and letters quote only approved, de-identified stories.
Common mistakes
- Listing systems instead of information. “Microsoft 365” tells you nothing; “client contracts in the legal folder” does.
- Forgetting outputs and logs. They can hold more personal information than the inputs.
- Assuming retention. Provider retention differs by product and plan; record the actual terms.
- Not recording the source of answers. “Vendor said so” will not survive a client questionnaire.
Next step
Use the inventory to complete a privacy impact assessment and the information rules in your AI use policy. For processing location choices, see data residency for AI systems, and for preparing the data itself, AI data preparation.
Sources
- Privacy Act 2020, New Zealand Legislation, version as at 1 October 2026
- Artificial intelligence and the Information Privacy Principles, Office of the Privacy Commissioner, September 2023
- Working with third-party providers, Office of the Privacy Commissioner, 26 November 2024
- IPP3A: notification requirements for indirect collection, Office of the Privacy Commissioner, accessed 19 October 2026
- Principles of Māori Data Sovereignty, Te Mana Raraunga, October 2018
- Responsible AI Guidance for Businesses, Ministry of Business, Innovation and Employment, July 2025
- How long do you store my organization’s data?, Anthropic, 2 July 2026
Written by the Levy Street team in Wellington. Last updated 19 October 2026.