Data residency for AI systems.

What data residency means for an AI system

Data residency for an AI system is three questions, not one: where data is stored at rest, where the model processes it during inference, and where the side copies go (logs, caches, embeddings, backups, support access). A provider can answer the first with “Australia” and the second with “anywhere”, and both can be true of the same request.

Residency is location; sovereignty is whose laws and control apply, and an overseas-headquartered provider can be subject to its home country’s laws while your data sits in Auckland. Māori data adds a question of governance. Te Mana Raraunga’s Māori Data Sovereignty Principles state that, whenever possible, Māori data shall be stored in Aotearoa New Zealand.

What New Zealand law requires

The Privacy Act 2020 regulates disclosure, security and use, not storage location. Three provisions decide most AI architecture questions.

• Section 11: providers acting on your behalf. If a provider holds personal information only to store or process it for you, the information is treated as held by you, and section 11(4) says it does not matter that the provider is outside New Zealand. Sending data to that provider is not a disclosure. If the provider also uses the information for its own purposes, it holds it as well.
• IPP12: disclosure outside New Zealand. IPP12 allows disclosure to a foreign person or entity only on specific grounds: informed authorisation by the individual; the recipient is subject to the Act; comparable safeguards through its laws, a prescribed scheme or a prescribed country; or an obligation, such as a contract, to protect it comparably.
• IPP5: security. You must protect information with safeguards that are reasonable in the circumstances, wherever it is processed.

The Privacy Commissioner’s guidance on AI and the information privacy principles applies this directly: using offshore technology providers to store or process data is not treated as a disclosure under IPP12, so long as they are not using it for their own purposes. Whether IPP12 applies can therefore turn on the provider’s terms, for example whether your prompts are used to train its models. The Commissioner’s guidance on working with third-party providers adds the contract terms to look for: use only for your services, breach notification, retention and deletion, and consultation where there are Māori data sovereignty implications.

This is general information, not legal advice; contracts, sector regulators and public sector policies can impose onshore requirements the Act does not. Our lesson on protecting customer information when using AI covers the everyday practice.

Cloud regions in and near New Zealand

As of 19 October 2026, two of the three large cloud providers run regions in Auckland.

Microsoft 365 also treats New Zealand as a local region geography eligible for Advanced Data Residency, which covers storage of Microsoft 365 Copilot and Copilot Chat content. Storage is not processing: Microsoft’s in-country processing plan for Microsoft 365 Copilot, as updated in April 2026, lists Australia, India, the United Arab Emirates, the United Kingdom and the United States by the end of 2026, and does not list New Zealand.

Where each AI provider runs inference

A cloud region in Auckland does not mean model inference in Auckland. As listed on 19 October 2026:

Inference guaranteed to stay in New Zealand currently means self-hosted open-weight models, covered in running open models for business. Regional model lists change with every release, so treat region checks as part of AI model migration.

Pin the region and verify where inference ran

Configuration states your intent; logs prove it. Pin regional endpoints and profiles in code, block the alternatives with policy, and check where requests were actually processed.

import boto3

ALLOWED_INFERENCE_REGIONS = {"ap-southeast-6", "ap-southeast-2", "ap-southeast-4"}  # Auckland, Sydney, Melbourne

bedrock = boto3.client("bedrock-runtime", region_name="ap-southeast-6")
response = bedrock.converse(
    modelId="au.anthropic.claude-sonnet-4-6",   # geographic profile, not the global one
    messages=[{"role": "user", "content": [{"text": "Summarise this tenancy notice: ..."}]}],
)

def out_of_region(cloudtrail_events: list[dict]) -> list[dict]:
    """Bedrock records the processing region in additionalEventData.inferenceRegion."""
    return [e for e in cloudtrail_events
            if e.get("eventSource") == "bedrock.amazonaws.com"
            and e.get("additionalEventData", {}).get("inferenceRegion") not in ALLOWED_INFERENCE_REGIONS]

Other providers expose similar evidence: on Claude 4.6 and later models, Anthropic’s API returns the inference geography in each response’s usage object. Run the check daily and alert on any result. AWS’s announcement also shows service control policies that confine Bedrock to the three Australia and New Zealand regions.

Map every data flow, not just the model call

Most residency surprises come from the side flows. List each one with its location and retention.

Our AI data inventory template has columns for each, and building private AI search covers the permission questions retrieval adds.

Choose an architecture by data class

Where residency plans go wrong

Recording these decisions is part of AI governance work, and vendor questions to ask before signing are in the AI vendor evaluation checklist.

• Every data flow is listed: inference, storage, embeddings, provider logs, your traces, tools, support access and datasets
• Each provider’s terms confirm whether it uses your data for its own purposes, such as training
• The section 11 or IPP12 position is recorded for each provider in a privacy impact assessment
• Regional endpoints, profiles or deployment types are pinned in code, and global options are blocked by policy
• Processing location is verified from provider logs or response fields, with alerts on exceptions
• Embeddings, vector backups and traces are hosted and protected like the source documents
• Contracts cover use limits, breach notification, retention, deletion and sub-processors
• Māori data and health information have an explicit, reviewed decision
• Provider residency pages are re-checked on a schedule, with the date recorded

Written by the Levy Street team in Wellington. Last updated 19 October 2026.