By The HelloHQ team

Using AI on Family Finances Without Handing Over the Data

What happens to a bank statement or a trust deed after it is pasted into an AI assistant, how personal and business accounts differ, what Singapore's rules require of a family office, and how to decide what goes where.

AI assistants are now good at reading statements, summarising trust deeds and checking spreadsheets. Each of those tasks means sending the document to a company that runs the model. This guide sets out what the main providers do with that data, how long they keep it, and what Singapore's rules expect of a family office that uses them.

Where a prompt goes

Text and files typed into an AI chat are sent to the provider's servers, processed there, and stored. What happens next depends on the account type and its settings.

What happens to a chat after it is sentA person pastes a document into an AI chat. The provider processes and stores the chat. In the usual case, the chat is deleted within about 30 days of the user deleting it. It can be kept longer if it is used for training, read by human reviewers, flagged by safety systems, or held under a court order.A document is pasted into a chatSent to the provider's serversThe provider processes and stores itKept in chat history until deletedUsual caseDeletedUsually within 30 days of theuser deleting itExceptionsKept longerTraining, human review,safety flags or a court order
Typical deletion windows from provider policy pages[1–3]. Settings and account types change the outcome.

Personal and business accounts are different products

The same assistant comes in personal and business versions, and the data terms differ. Personal accounts at OpenAI and Google are used for model training unless the user turns it off[3, 4]. Business accounts and paid developer access (the API) at OpenAI, Anthropic, Google and Microsoft are not used for training by default[5–8].

Three ways to use an AI modelThree columns. Personal account: may be used for training unless switched off, deleted chats purged in about 30 days, some providers use human reviewers. Business account or paid API: not used for training by default, admin-set retention, contract terms, zero retention available on some APIs by agreement. Own hardware: an open-weight model runs on the family's or office's machine and nothing is sent to a provider.Personal accountTraining on bydefault at OpenAIand Google; achoice at AnthropicDeleted chatspurged in about 30daysHuman review atsome providersBusiness account orpaid APINo training bydefaultRetention set bythe administratoror the contractZero retention onsome APIs, byagreementOwn hardwareOpen-weight modelon the family's oroffice's machineNothing sent to aproviderSmaller models,less capable
Defaults as stated by the providers[5–8].
Service Trains on chats by default After a chat is deleted Human review
ChatGPT Free, Plus, Pro Yes, can be turned off Removed within 30 days Not stated
ChatGPT Business, Enterprise No Removed within 30 days Not stated
Claude Free, Pro, Max The user chooses[9] Removed within 30 days; up to 5 years if training is on Flagged chats only
Claude Team, Enterprise No Removed within 30 days Flagged chats only
Gemini app, personal account Yes, while Keep Activity is on 72 hours if activity is off; reviewed chats up to 3 years Yes, a subset of chats
Gemini in Google Workspace No Admin choice: 3, 18 or 36 months No
Microsoft Copilot, work account No Under the organisation's retention policy No

Microsoft released a new consumer Copilot app on 18 August 2026. Its pages say prompts, responses and files are not used to train foundation models, but they do not state a retention period[10]. The older app kept conversations for 18 months[11].

How long a chat can be kept

Deleting a chat removes it from the history screen at once. Copies in the provider's systems follow their own schedule, and some are kept well beyond 30 days.

Longest stated retention, in daysFive bars. Gemini app with activity off: 3 days. OpenAI or Anthropic chat after deletion: 30 days. Anthropic chat flagged for a policy breach: 730 days. Gemini chat read by a human reviewer: 1,095 days. Anthropic consumer chat with training switched on: 1,825 days.Gemini, activity off3OpenAI or Anthropic, deleted30Anthropic, flagged chat730Gemini, human-reviewed1,095Anthropic, training on1,825
Google keeps reviewed Gemini chats separately from the account, and deleting activity does not remove them. Anthropic keeps de-identified data for training[2, 3].

Google's own advice for the Gemini app is not to enter confidential information that you would not want a reviewer to see[3].

Deletion can be overridden

Every provider's retention terms allow it to keep data longer when the law requires it. In 2025 a US court did require it.

The court order to keep ChatGPT chatsOn 13 May 2025, in the New York Times copyright case, a US court directed OpenAI to keep chat output that would otherwise be deleted. It covered ChatGPT Free, Plus, Pro and Team and API use without zero retention; Enterprise, Edu and zero-retention API use were excluded. On 26 September 2025 the obligation ended for new chats. Chats kept during the order remain held, except those from the EEA, Switzerland and the UK.13 May 2025Court directs OpenAI to keep deleted chatsFree, Plus, Pro, Team and most API use.Enterprise, Edu and zero-retention APIexcluded.26 Sep 2025Obligation ends for new chatsDeleted chats are again removed within 30days.Since thenChats from the order period stay heldExcept those from the EEA, Switzerland andthe UK.
From the court order, the October 2025 stipulation and OpenAI's statement[12–14].

For four and a half months, deleting a chat on a personal ChatGPT account did not delete it. ChatGPT Enterprise and Edu accounts were outside the order.

Shared links are public

Most assistants can create a link to a conversation. On a personal ChatGPT account, anyone with the link can read it[15]. In mid-2025 an opt-in setting let shared ChatGPT chats appear in search results; OpenAI removed the setting at the end of July 2025. In August 2025 Forbes reported that more than 370,000 shared Grok conversations had been indexed by Google[16]. On ChatGPT business workspaces, shared links open only for members of the same workspace[15].

What Singapore's rules expect

The Personal Data Protection Act (PDPA) applies to organisations, not to individuals acting in a personal or domestic capacity[17]. A family member asking an AI about their own affairs is generally outside it. A family office is an organisation, and its staff act for it.

PDPA obligations that apply when an office uses an AI providerThree columns. Protection: make reasonable security arrangements to prevent unauthorised access or disclosure. Transfer limitation: personal data sent overseas must get protection comparable to the PDPA. Retention limitation: stop keeping personal data once it is no longer needed, which includes copies a provider keeps.ProtectionReasonable securityarrangementsagainstunauthorised accessor disclosureTransfer limitationData sent overseasmust get protectioncomparable to thePDPARetentionlimitationStop keeping dataonce it is nolonger needed,including aprovider's copies
Applies to organisations, including family offices[18].

Pasting a family member's or a beneficiary's details into an AI tool is a disclosure of their personal data to the provider, usually to servers outside Singapore. The Personal Data Protection Commission's 2024 guidelines on AI systems encourage organisations to minimise personal data and to pseudonymise or anonymise it where they can[19]. Those guidelines were written for recommendation and decision systems, not chat assistants, but the same practice applies.

The Monetary Authority of Singapore (MAS) issued Guidelines on AI Risk Management on 7 October 2026[20]. They take effect in October 2027 and apply to financial institutions, including the private banks and wealth managers a family uses. They name leakage of confidential or customer data through third-party generative AI services as a risk, and give a ban on entering client information into public AI tools as an example policy[21]. A single-family office that is exempt from licensing is not directly bound, but the guidelines show what its banks will be asked about.

Sorting data before it is pasted

One practical approach is to decide, by type of data, which of the three options it may go to.

One way to sort data by where it may goThree columns. Any account: public information, general questions, and figures with names and account numbers removed. Business account with no-training terms: working drafts, de-identified portfolio data, internal analysis. Own hardware, or not at all: identity documents, account numbers, statements with names, other people's personal data, passwords and access codes.Any accountPublic informationGeneral questionsFigures with namesand account numbersremovedBusiness account,no trainingWorking draftsDe-identifiedportfolio dataInternal analysisOwn hardware, ornot at allIdentity documentsand account numbersStatements anddeeds with namesOther people'spersonal dataPasswords andaccess codes
An example policy, not a standard. Adjust to the family's own risk appetite and contracts.

Open-weight models can be downloaded and run without sending anything to a provider. Examples include OpenAI's gpt-oss[22], Alibaba's Qwen3[23] and Google's Gemma 4[24], all under the Apache 2.0 licence, and Meta's Llama 4 under its own community licence[25]. OpenAI states that its smaller gpt-oss model runs in 16 GB of memory[22]. Models of that size are less capable than the hosted ones, so the trade-off is accuracy against control.

Practical steps

  1. List which AI tools family members and staff already use, and on which account type.
  2. Move office work to business accounts, and check that training is off and retention is set.
  3. In personal accounts, turn off training and use temporary chats for anything sensitive.
  4. Remove names, account numbers and identity numbers before pasting a document.
  5. Do not share chat links that contain financial details.
  6. Write down which data may go where, and tell staff.
  7. For documents that should not leave the office, run a model on the office's own hardware.

What we could not confirm

Provider terms change often; the figures here are from policy pages read on 12 October 2026. OpenAI's consumer help pages do not say whether staff review personal chats. We could not confirm whether Anthropic's training setting starts switched on for new accounts, or the retention period and human review terms of Microsoft's new consumer Copilot app. Google states only that paid Gemini API logs are kept for "a limited period of time". We relied on the PDPC's summary of the PDPA obligations rather than the statutory text, and the PDP Regulations were amended in 2026. We did not confirm whether the court endorsed the October 2025 stipulation in the OpenAI case. OpenAI announced the removal of the search setting in a post on X, which we read through news reports. The Grok figure comes from a news report, not from xAI. This is education, not legal advice.

References

  1. OpenAI. “Chat and file retention policies in ChatGPT.” OpenAI Help Center. help.openai.com/en/articles/8983778-chat-and-file-retention-policies-in-chatgpt. Accessed 12 October 2026. ↩︎
  2. Anthropic. “How long do you store my data?” Anthropic Privacy Center. privacy.claude.com/en/articles/10023548-how-long-do-you-store-my-data. Accessed 12 October 2026. ↩︎
  3. Google. “Gemini Apps Privacy Hub.” Gemini Apps Help, updated 24 September 2026. support.google.com/gemini/answer/13594961?hl=en. Accessed 12 October 2026. ↩︎
  4. OpenAI. “How your data is used to improve model performance.” OpenAI Help Center. help.openai.com/en/articles/5722486-how-your-data-is-used-to-improve-model-performance. Accessed 12 October 2026. ↩︎
  5. OpenAI. Enterprise privacy at OpenAI. Updated 8 January 2026. openai.com/enterprise-privacy/. Accessed 12 October 2026. ↩︎
  6. Anthropic. “How long do you store my organization's data?” Anthropic Privacy Center. privacy.claude.com/en/articles/7996866-how-long-do-you-store-my-organization-s-data. Accessed 12 October 2026. ↩︎
  7. Google. “Generative AI in Google Workspace Privacy Hub.” Google Workspace Admin Help, updated 6 October 2026. knowledge.workspace.google.com/admin/gemini/generative-ai-in-google-workspace-privacy-hub?hl=en. Accessed 12 October 2026. ↩︎
  8. Microsoft. “Data, privacy, and security for Microsoft Copilot.” Microsoft Learn, updated 8 October 2026. learn.microsoft.com/en-us/copilot/microsoft-365/microsoft-365-copilot-privacy. Accessed 12 October 2026. ↩︎
  9. Anthropic. Updates to consumer terms and privacy policy. 28 August 2025. www.anthropic.com/news/updates-to-our-consumer-terms. Accessed 12 October 2026. ↩︎
  10. Microsoft. “Microsoft Copilot activity history.” Microsoft Support. support.microsoft.com/en-us/privacy/microsoft-copilot/activity-history. Accessed 12 October 2026. ↩︎
  11. Microsoft. “Privacy FAQ for Microsoft Copilot.” Microsoft Support. support.microsoft.com/en-us/topic/privacy-faq-for-microsoft-copilot-27b3a435-8dc9-4b55-9a4b-58eeb9647a7f. Accessed 12 October 2026. ↩︎
  12. United States District Court for the Southern District of New York. “Order.” In re OpenAI, Inc., Copyright Infringement Litigation, No. 25-md-3143, ECF No. 551, 13 May 2025. blogs.duanemorris.com/classactiondefense/wp-content/uploads/sites/56/2025/10/In-re-OpenAI-Inc.-Copyright-Infringement-Litig.-5.13.25-Order-ECF-No.-551.pdf. Accessed 12 October 2026. ↩︎
  13. United States District Court for the Southern District of New York. “Stipulation on preservation of output log data.” In re OpenAI, Inc., Copyright Infringement Litigation, No. 25-md-3143, Doc. 922, 9 October 2025. blogs.duanemorris.com/classactiondefense/wp-content/uploads/sites/56/2025/10/In-re-Open-AI-Inc.-Copyright-Infringement-Litig.-10.9.25-Order-ECF-No.-992.pdf. Accessed 12 October 2026. ↩︎
  14. OpenAI. How we're responding to The New York Times' data demands in order to protect user privacy. Updated 22 October 2025. openai.com/index/response-to-nyt-data-demands/. Accessed 12 October 2026. ↩︎
  15. OpenAI. “ChatGPT shared links FAQ.” OpenAI Help Center. help.openai.com/en/articles/7925741-chatgpt-shared-links-faq. Accessed 12 October 2026. ↩︎
  16. Fortune. “Hundreds of thousands of Grok chats are now public on Google search.” Fortune, 22 August 2025. fortune.com/2025/08/22/xai-grok-chats-public-on-google-search-elon-musk. Accessed 12 October 2026. ↩︎
  17. Personal Data Protection Commission Singapore. Personal Data Protection Act overview. www.pdpc.gov.sg/overview-of-pdpa/the-legislation/personal-data-protection-act. Accessed 12 October 2026. ↩︎
  18. Personal Data Protection Commission Singapore. Data protection obligations. Updated 3 April 2023. www.pdpc.gov.sg/overview-of-pdpa/the-legislation/personal-data-protection-act/data-protection-obligations. Accessed 12 October 2026. ↩︎
  19. Personal Data Protection Commission Singapore. Advisory Guidelines on Use of Personal Data in AI Recommendation and Decision Systems. 1 March 2024. www.pdpc.gov.sg/-/media/files/pdpc/pdf-files/advisory-guidelines/advisory-guidelines-on-the-use-of-personal-data-in-ai-recommendation-and-decision-systems.pdf. Accessed 12 October 2026. ↩︎
  20. Monetary Authority of Singapore. “MAS sets out supervisory expectations on responsible AI adoption by financial institutions.” Media release, 7 October 2026. www.mas.gov.sg/news/media-releases/2026/mas-sets-out-supervisory-expectations-on-responsible-ai-adoption-by-financial-institutions. Accessed 12 October 2026. ↩︎
  21. Monetary Authority of Singapore. Guidelines on Artificial Intelligence Risk Management. 7 October 2026. www.mas.gov.sg/regulation/guidelines/guidelines-on-artificial-intelligence-risk-management-for-financial-institutions. Accessed 12 October 2026. ↩︎
  22. OpenAI. Introducing gpt-oss. August 2025. openai.com/index/introducing-gpt-oss/. Accessed 12 October 2026. ↩︎
  23. Qwen Team, Alibaba. Qwen3: Think deeper, act faster. 29 April 2025. qwenlm.github.io/blog/qwen3/. Accessed 12 October 2026. ↩︎
  24. Google. “Gemma models overview.” Google AI for Developers, updated 7 October 2026. ai.google.dev/gemma/docs. Accessed 12 October 2026. ↩︎
  25. Meta. “Llama 4 Scout 17B-16E Instruct model card.” Hugging Face, 5 April 2025. huggingface.co/meta-llama/Llama-4-Scout-17B-16E-Instruct. Accessed 12 October 2026. ↩︎

Topics

Privacy and AI SingaporeCross-border

Related guides

Keeping Your 13O or 13U Incentive in Force

The conditions a Singapore single-family-office fund must meet to keep its 13O or 13U tax exemption, when each one is tested under MAS's August 2026 rules, and what happens if one is missed.

Foreign Income Tax in Six Jurisdictions

Six jurisdictions tax residents on income earned abroad, and automatic reporting lets tax authorities see it. What changed in China, Taiwan, Malaysia, Singapore, Indonesia and the US, and the records a household may need to produce.

Counterparty Checks After a Singapore Seizure

In late October 2025 Singapore police seized or froze more than S$150 million of assets linked to a sanctioned regional conglomerate. This post covers how households and small firms that dealt with such a party can be exposed, and the checks and records that limit the exposure.