There is a particular kind of risk that keeps experienced insurance people up at night. Not the risks they've priced or the risks they've accounted for in their risk registers. The risks they haven't noticed yet.

AI is one of those risks right now. The technology is not inherently dangerous, but the regulatory response to it is finally gathering speed, and for many insurance businesses the gap between what they do and what they're supposed to do is quietly widening.

Over the past year, I've been deep in an AI bubble, for my own businesses, for my personal growth and in conversations with other early AI adopters in the Australian and Lloyd's markets. Here is my real-world take on the obligations in place today, the deadlines approaching, and the decisions insurance businesses need to make now rather than later.

The regulatory landscape nobody has mapped clearly

In Australia, insurers do not yet face a single, AI-specific insurance rulebook. Instead, AI use is pulled into existing privacy, cyber, outsourcing and governance obligations.

The Australian Privacy Act and the Australian Privacy Principles

The Privacy Act 1988 and its 13 Australian Privacy Principles (APPs) apply to APP entities, which include most insurers, intermediaries, and service providers that handle personal information in this context. For insurers and coverholders, this means virtually every piece of customer data you touch.

The most immediately relevant principle is APP 6, which governs secondary use and disclosure of personal information. The core obligation is clear: personal information collected for one purpose cannot be used or disclosed for another purpose without consent or a specific exemption.

When a broker or underwriter drops a client's proposal form into a consumer AI tool, they are disclosing that personal information to a third party, the AI vendor, for a purpose the customer did not consent to.

The OAIC has been explicit about this. Their published guidance states that, as a matter of best practice, organisations should not enter personal information, and particularly sensitive information, into publicly available generative AI tools, due to the significant and complex privacy risks involved. The OAIC even provides a worked example of an insurance company whose employees enter a customer's claims data, including sensitive health information, into a publicly available chatbot, framing it as a clear compliance failure under APP 6.

The financial exposure is real and tiered. The Privacy and Other Legislation Amendment Act 2024, which received Royal Assent on 10 December 2024, introduced a three-tier civil penalty regime. The lower tiers are set in Commonwealth penalty units, and one penalty unit has been $364 for conduct on or after 1 July 2026, up from $330. Administrative failures, such as having a non-compliant privacy policy, attract civil penalties of up to $364,000 (1,000 penalty units) for corporations, and the OAIC can now issue infringement notices of up to $72,800 (200 penalty units) per contravention for listed companies without going to court. Non serious interferences with privacy carry penalties of up to $3.64 million (10,000 penalty units) for corporations. Serious interferences carry penalties of up to $50 million, three times the value of any benefit obtained, or 30% of adjusted turnover, whichever is greater.

The December 2026 ADM deadline

The Automated Decision Making (ADM) transparency requirement will take effect on 10 December 2026, under new APPs 1.7, 1.8 and 1.9 introduced by the Privacy and Other Legislation Amendment Act 2024. These obligations apply to any decision made on or after that date, regardless of when the underlying system was deployed.

APP entities that use automated systems, including AI, to make or significantly influence decisions that could reasonably be expected to materially affect an individual's rights or interests must disclose this in their privacy policies. The required disclosures are specific: the kinds of personal information used, and the kinds of decisions made solely by the system or where the system does something substantially and directly related to the decision. Vague statements about "using technology to improve our services" will not be sufficient.

For insurers, this captures more than you might expect: automated underwriting decisions, AI-assisted claims triage, and algorithmic pricing all potentially trigger disclosure obligations.

The OAIC commenced its first ever compliance sweep in January 2026, targeting approximately 60 organisations across six sectors to assess privacy policy compliance, a clear signal that enforcement is active.

Nine months is not long when you factor in the time required to identify all AI touchpoints, draft compliant policy language, obtain legal review, and publish updates.

Lloyd's MS11 and the CDM requirements

For Lloyd’s businesses, MS11 (Cyber Resilience and Data Management) is an important part of the control environment. It is addressed to managing agents, but it can shape expectations around delegated operations, suppliers and data governance across the wider chain.

MS11 does not contain AI-specific requirements, and Lloyd's has yet to publish dedicated AI guidance. Its general data governance framework still applies directly to AI usage. The relevant provisions are:

An insurance business that has staff using consumer AI tools to process policyholder data, without a documented governance framework and vendor assessment, is likely non-compliant with CDM 1.1, regardless of what the Australian Privacy Act says. The absence of AI-specific Lloyd's guidance creates regulatory flexibility, but it does not create a safe harbour.

APRA CPS 234

For APRA-regulated entities, which include many insurers operating in Australia, CPS 234 on Information Security adds a further layer. While CPS 234 predates the current AI wave, its requirements around information asset classification, third party management, and incident response apply to AI systems and the data they process.

For coverholders and MGAs that are not themselves APRA-regulated, CPS 234 may still be relevant through contractual flow down from carrier partners. It is worth checking your binder and delegated authority agreements carefully.

The platform problem nobody wants to talk about

The consumer AI tools most insurance professionals use were not designed for regulated industry use. They offer no Australian data residency, no enterprise data processing agreements, and in several cases, have default settings that involve using submitted content to train future model versions.

This is a documented risk, not a hypothetical one:

ChatGPT Free and Plus: Data used for training by default, no data processing agreement available, no data residency controls, no administrative oversight for organisations.

Claude Pro and Max: Since a consumer terms change in late 2025, chats and coding sessions can be used for model training if the user allows it. The choice is a single privacy setting on each individual account, which each user controls themselves. There are no organisational controls at these tiers.

The enterprise tiers of both platforms address most of these issues, but at a higher cost and with ongoing management overhead. The more interesting question for many organisations is whether there is a compliant path that does not require expensive enterprise contracts.

For organisations already running Google Workspace, the answer is often yes. Gemini has been included in Workspace Business and Enterprise plans since January 2025, with no add-on to buy, though Google raised plan prices when it bundled it in. Google commits not to use customer data to train models without permission, extends existing Workspace DLP controls to Gemini output, and lists Gemini as certified under ISO/IEC 42001. The gap is residency: Gemini processing can be confined to the US or EU, not Australia, and only on Enterprise Plus or with the Assured Controls add-on. This is not a product recommendation. It is an observation about where the compliance bar currently sits and which platforms clear it without additional investment.

What good governance actually looks like

Classify data before choosing tools. Not all insurance data carries the same risk. Internal communications, market research, and publicly available information can be processed by a wider range of tools than sensitive PII or claims data. A simple data classification framework (public, internal, confidential, sensitive) that maps to approved tool usage is the foundation of everything else.

Make the compliant path the easy path. The OAIC's own case study illustrates the failure mode: a car insurer's policy banned personal information in public AI tools, yet an employee uploaded a customer's hardship application to ChatGPT anyway. The fix the OAIC describes pairs training with technical controls that block the upload, rather than relying on the policy alone. Governance that creates friction will be worked around.

Document the decisions. When an auditor or OAIC investigator asks what controls you have in place, the answer is not a verbal description of what you think your staff are doing. It is a policy document, a training record, a vendor assessment, and a review log. The documentation is the evidence.

Assign director level accountability. Under Lloyd's CDM 2.2, this is a requirement. In practice, it also changes behaviour. When a named director is responsible for AI governance outcomes, the governance work gets done.

What to do now

If you are reading this and recognising your own organisation, the practical starting point is not a technology project. It is an audit.

Map where AI is being used across your business: the approved tools, and what staff are actually doing. Survey your staff. You will likely find things that surprise you.

Then work backwards: which of those use cases involve personal information? Which platforms are those use cases running on? Do those platforms have appropriate data processing agreements, residency controls, and training opt-outs?

The answers will tell you where your exposure is. From there, the path forward is straightforward, if not quick.

The December 2026 deadline is confirmed in law, and the OAIC is already enforcing. Businesses that treat this as a compliance checkbox will do the minimum required. Businesses that treat it as a governance problem will build controls that last.

The gap between those two groups will become more visible over the next nine months.

Sources

Claim Source
Three-tier civil penalty structure, Royal Assent 10 December 2024 Johnson Winter Slattery: Changes to Australian privacy laws now in force
Commonwealth penalty unit of $364 from 1 July 2026 (Crimes Act 1914 s 4AA) Crimes (Amount of a Penalty Unit) Instrument 2026, F2026N00424
ADM transparency obligations, APP 1.7–1.9, December 2026 OAIC APP 1 Guidance
OAIC compliance sweep, January 2026, ~60 organisations Landers & Rogers 2026 Privacy Update
OAIC guidance on AI tools and APP 6, insurance scenario OAIC Guidance on Commercially Available AI
OAIC GenAI workplace case study (CarCover) OAIC blog: GenAI tools in the workplace
Lloyd's MS11, CDM 1.1, 1.5 and 2.2 Lloyd's Minimum Standards MS11 – Cyber Resilience and Data Management (January 2021)
APRA CPS 234 Information Security APRA CPS 234
Claude consumer training setting Anthropic: Updates to Consumer Terms; Anthropic Privacy Center: model improvement settings
Gemini included in Workspace plans Google Workspace blog, January 2025
Gemini no-training commitment, DLP, ISO/IEC 42001 Generative AI in Google Workspace Privacy Hub
Gemini data regions (US/EU only) Workspace Updates: data regions support for Gemini