An AI security review for an Australian Financial Services Licence (AFSL) holder is a set of seven questions about data ownership, where data is processed and stored, model training, access, human sign-off, audit trails and licence obligations. A funds manager that can answer all seven in writing has what a board and a compliance lead need before AI goes near investor or credit information.
In one week, two funds managers sent us almost the same list of questions. One came from a board, one from a compliance lead. Both hold an AFSL, both want AI working across their teams, and both needed to show that it wouldn't create a problem they'd have to explain to ASIC.
The questions were good. Here are the seven that came up, with how we answer them. If you hold a licence and you're weighing up AI, these are the questions to ask whoever you're working with.
1. Who owns our data, prompts and outputs?
You should, outright. That covers documents, prompts, outputs and everything an AI system learns from them. Ask where every copy lives, including logs, backups, vendor retention and any copies kept for abuse monitoring, and get the answer in writing.
Then ask about the system itself. Connectors, workflows and dashboards built for you should be yours to change and reuse. If the core product is licensed, you should be able to run it without the vendor. Lock-in becomes a compliance problem as well as a commercial one.
2. Where is data processed and stored, and who else touches it?
This question matters most, and it has the most options. Using Claude as the example, here is what each provider says about its own service:
- A standard Claude business plan (Team or Enterprise). Anthropic says it does not train on inputs or outputs from its commercial products by default, and admins control access. Anthropic says data is stored in the US, and that by default it may route customer traffic to select countries in the US, Europe, Asia and Australia unless otherwise agreed. That suits everyday work that doesn't touch investor or credit data.
- Claude through AWS Bedrock. It sits under your existing AWS agreement. AWS says model providers have no access to Bedrock logs or to customer prompts and completions. If you call an Australian cross-Region inference profile, AWS says Bedrock routes requests between Sydney and Melbourne and processes them within Australia. A global profile can route to any supported AWS Region worldwide, so the profile you choose decides the answer. Retention depends on the model. For Claude Fable 5 and 5.1, AWS says all traffic is retained for up to 30 days for abuse detection, that flagged traffic may be reviewed by a person at AWS, and that the retained copy sits in the Region that processed the request. As of 6 November 2026, AWS lists Fable 5 and 5.1 on its US and global profiles only, so from Australia they run through the global profile. Check that the model you want is offered on the Australian profile.
- The Claude API with zero data retention. Anthropic approves this per organisation, and it applies to eligible APIs. Anthropic says it still keeps user safety classifier results, and that for Covered Models it requires limited retention and review (see question 3 for the conditions). Anthropic also documents options for choosing a routing location, so ask which apply to your plan before you tell your board where processing happens.
- Microsoft 365 Copilot. Microsoft says the content of your prompts and responses is stored at rest in your tenant's geography, and Australia is a listed location for tenants that signed up in Australia. Where the model processes a request is a separate question. Microsoft's own update says local processing for Copilot is expected to become available in Australia by the end of 2026, so ask Microsoft to confirm what applies to your tenant today.
- An anonymisation layer on top of any of these strips investor names, deal names and personal details before anything reaches the model.
A common setup is a standard business plan for everyday staff use, with a stricter option for anything that touches investor or credit data. If keeping processing in Australia matters most to your board, Bedrock with an Australian inference profile is where we would start, after checking that the exact model is on the Australian profile and what its retention rules are.
3. Is our data used to train models?
On Anthropic's business products, not by default. There's one catch worth knowing: if someone submits feedback with a thumbs up or down, Anthropic can store the whole conversation and use it for training. Owners of Team and Enterprise plans can turn the feedback button off for the whole organisation, so switch it off.
Ask about retention too. Anthropic says its commercial products keep chats so people can continue them. Deleted chats leave the history straight away and are removed from Anthropic's back end within 30 days. Chats flagged by safety systems can be kept for up to 2 years, and feedback conversations for 5 years. Anthropic's retention page also says that for Covered Models it requires limited data retention and review as part of its safety work. Anthropic's page on Covered Models limits that change to organisations using zero data retention, including Claude through AWS Bedrock with zero data retention, and says Claude chat in Claude for Enterprise keeps its standard retention. It also says no Anthropic personnel can read retained conversations by default, and that human review can happen only through a controlled, logged access path, for example when its safety systems flag content. It does not address Team plans, so ask Anthropic whether a model you plan to use is a Covered Model and which conditions apply. Anthropic documents custom retention controls for Enterprise plans separately.
4. Who can see what?
The AI should see what each person can already see, and nothing more. Sign-in should go through your existing identity system (Entra, for most), so disabling someone's account cuts their access straight away. Connectors to SharePoint and Outlook should act as the signed-in person, read-only by default.
Decide what goes in at all. One rule we like is to put foundational documents and status in the shared brain, with a link back to the source. Keep confidential deal and investor data in the systems built for it.
5. Does a person check before anything goes out?
A person should sign off before any AI output reaches an investor, a lender or a credit decision. AI drafts; people send. For agents, write down before go-live what (if anything) may go out without a human.
6. Can we show how an answer was reached?
You need an audit trail of who asked, when, what was asked and what came back. Check where those logs live (yours, ideally), how long they're kept and who can see them. On AWS, CloudTrail logs every cross-Region inference request in your source Region and records where it was processed. On Claude, Anthropic's Compliance API includes audit log events and chat data, and its help page lists it for Enterprise plan organisations and Claude Platform customers. That page does not mention Team, so confirm what your plan includes before you rely on it.
7. What about our AFSL obligations?
Your licence obligations stay with you when you use AI or an outside provider. ASIC's report REP 798 says the financial services framework is technology neutral, and that licensees remain responsible for outsourced functions. A good partner makes the obligations easy to meet:
- General obligations (s912A). REP 798 says licensees should have measures for complying with their general obligations, and that these should be documented, implemented, monitored and regularly reviewed. In practice: put each AI use on your risk register with its controls, and make sure nothing critical depends on the vendor to keep running.
- Cyber resilience. In ASIC v RI Advice ([2022] FCA 496), the Federal Court found that RI Advice did not have adequate cybersecurity documentation and controls across its authorised representative network. The Court accepted that cyber risk cannot be reduced to zero, but said it can be materially reduced. ASIC says it is imperative for licensees to have adequate cybersecurity systems in place. The controls we put in place for clients are our practice, not a legal standard: 2FA everywhere, least-privilege access and a separate credential vault for each client.
- Outsourcing oversight. REP 798 says licensees should have measures to choose suitable service providers, monitor their performance and deal appropriately with their actions. We would also ask any provider for full access, logs, documentation and a clean exit.
- Privacy Act. APP 8.1 requires reasonable steps to make sure an overseas recipient doesn't breach the Australian Privacy Principles, and you stay accountable for what the recipient does. The OAIC says storing data with an overseas cloud provider may count as a "use" rather than a disclosure where a binding contract limits what the provider can do, subcontractors are bound the same way and you keep effective control. Where your model's processing and storage sit, from question 2, feeds straight into this.
- Automated decisions (from 10 December 2026). Under APP 1.7, the duty applies where you have arranged for a computer program to make, or do something substantially and directly related to making, a decision, and the decision could reasonably be expected to significantly affect a person's rights or interests, and the program uses personal information. If all three apply, your privacy policy must say what kinds of personal information the program uses and what kinds of decisions it makes or informs. Keep a record of where AI is used so you can check each use against that test.
- ASIC's AI governance report (REP 798, "Beware the gap", October 2024). ASIC reviewed 23 licensees and found governance can lag AI adoption. Its reflection questions ask whether you have an AI inventory and who is accountable for AI use and outcomes. A register of AI use cases with a named owner is the simplest way to show you have both.
What about SOC 2 and ISO 27001?
Someone always asks. Those certifications audit how a vendor runs its own systems, so they matter most where the vendor hosts your data. If the AI runs in your own environment, under your controls, your own controls carry more of the weight. You should still ask each provider for its current certifications and reports for the parts it runs (AWS, Anthropic, Microsoft). A vendor that will fill in your security questionnaire is worth more than a badge.
Related reading

Written by
Finlay EkinsCo-founder
Finlay Ekins is the co-founder of Hourglass AI, where he leads the technical side of the build: the agents, the integrations, and the engineering judgement about what will hold up in production. He studied mechatronics engineering in Melbourne and left it for AI when it became clear the intelligence, not the hardware, was the interesting part. He started the business that became Hourglass on his own, made $700 in its first month, and spent ten months of failing and pivoting before it worked. He writes a weekly newsletter on AI from Australia at finlayekins.com, 14 essays since June 2026, and his standing advice to clients is to try ChatGPT properly for a month before commissioning anything custom.
More from FinlayCommon questions
Does using an AI provider move my AFSL obligations to the provider?
No. ASIC's report REP 798 says licensees remain responsible for outsourced functions, and should have measures to choose suitable providers, monitor their performance and deal appropriately with their actions. A good provider makes your obligations easier to meet, but they stay with you.
Is my data used to train Claude if we are on a business plan?
Anthropic says that by default it does not use inputs or outputs from its commercial products to train its models. The exception is feedback: if someone submits a thumbs up or down, the related conversation can be stored and used for training. Owners of Team and Enterprise plans can turn the feedback button off for their organisation.
Can Claude on AWS Bedrock keep all processing in Australia?
AWS says a geographic cross-Region inference profile keeps processing inside the geography, and for Australia it routes requests between Sydney and Melbourne. A global profile can route anywhere in the world. Retention depends on the model, so check the specific model's retention rules before you promise your board anything.
When do the new privacy policy rules for automated decisions start?
On 10 December 2026. From then, an organisation covered by the Privacy Act must describe in its privacy policy the kinds of personal information used, and the kinds of decisions made by computer programs, or where a program does something substantially and directly related to making the decision, where those decisions could reasonably be expected to significantly affect people's rights or interests. Not every AI use triggers it, so keep a record of where you use AI and check each use.
References
- 1. ASIC: 22-104MR Court finds RI Advice failed to adequately manage cybersecurity risks: ASIC v RI Advice Group [2022] FCA 496: the Court's finding on cybersecurity documentation and controls. Read 16 November 2026
- 2. ASIC: REP 798, Beware the gap: governance arrangements in the face of AI innovation: Published 29 October 2024. Technology-neutral obligations, outsourced functions, the review of 23 licensees and AI inventories. Read 16 November 2026
- 3. OAIC: New resources on transparency for use of AI and automated decision-making: APP 1.7 to 1.9, from 10 December 2026. Read 16 November 2026
- 4. OAIC: APP guidelines, chapter 8: cross-border disclosure of personal information: APP 8.1 reasonable steps, accountability under s 16C, and when overseas cloud storage is a use. Read 16 November 2026
- 5. AWS: Amazon Bedrock cross-Region inference for Claude in Japan and Australia: Published 31 October 2025. The Australian profile routes between Sydney and Melbourne. Read 16 November 2026
- 6. AWS: Route model inference requests across AWS Regions with cross-Region inference: Global profiles, and CloudTrail logging of cross-Region requests in the source Region. Read 16 November 2026
- 7. AWS: Data protection in Amazon Bedrock: Model providers have no access to Bedrock logs or to customer prompts and completions. Read 16 November 2026
- 8. AWS: Data retention in Amazon Bedrock: How retention is set per Region and depends on the model. Read 16 November 2026
- 9. AWS: Amazon Bedrock abuse detection: Fable 5 and 5.1 traffic kept for up to 30 days, possible human review at AWS, stored in the processing Region. Read 16 November 2026
- 10. AWS: Claude Fable 5.1 model card (inference profiles): Fable 5.1 listed on the US and global inference profiles only. Read 16 November 2026
- 11. Anthropic Privacy Center: Where are your servers located?: Default traffic routing to the US, Europe, Asia and Australia, with data stored in the US. Read 16 November 2026
- 12. Anthropic Privacy Center: Is my data used for model training?: No training on commercial inputs or outputs by default, and the thumbs feedback exception. Read 16 November 2026
- 13. Anthropic Privacy Center: How long do you store my organisation's data?: Deletion within 30 days, flagged chats up to 2 years, feedback 5 years, and Covered Models. Read 16 November 2026
- 14. Anthropic Privacy Center: Data retention practices for Covered Models: Scope and review conditions. Read 16 November 2026
- 15. Claude Help Center: Manage user feedback settings on Team and Enterprise plans: Owners can turn thumbs feedback off for the whole organisation. Read 16 November 2026
- 16. Claude Help Center: Access the Compliance API: Which plans the Compliance API is available on. Read 16 November 2026
- 17. Microsoft Learn: Data residency for Microsoft 365 Copilot offerings: Copilot prompts and responses stored at rest in the tenant's geography, with Australia listed. Read 16 November 2026
- 18. Microsoft: In-country data processing for Microsoft 365 Copilot in 15 countries: Local processing for Copilot in Australia expected by the end of 2026. Read 16 November 2026









