家族辦公室使用 AI 前的十題準備度檢查表

Ten yes/no questions. Give your office one point for each yes you can back with a document or a test record. The total matters less than the open answers: they are the list of things to settle before AI touches live family email, documents or reporting.

Prakshal Jain (PJ), Co-founder and CTO, Private SuperIntelligence (by Mitosis Labs)

Alexander Morris, Co-founder and CEO, Private SuperIntelligence (by Mitosis Labs)

We wrote this for the operator who will be asked, sooner or later, "are we ready to use AI on the family's information?" by a principal who has already seen a colleague do it. Answer on paper or in your head. This page has no answer fields and no scoring script, so it cannot collect, store or send anything you decide.

Does this sound like your office?

• Some staff are trying AI before anyone owns the decision. Citi's 2025 Global Family Office Report found 22% of family offices use AI for operational tasks or investment analysis, up from 13% in 2024. [1]
• Nobody in-house feels qualified to judge the tools. In the same Citi survey, 57% of family offices named lack of internal expertise as the biggest barrier to adoption. [1]
• There is no written AI policy yet. IBM's 2025 breach study found 63% of breached organizations had no AI governance policy or were still developing one. [2]
• Someone may be using a tool the office never approved. IBM found one in five organizations reported a breach involving shadow AI, and only 37% had policies to manage AI or detect shadow AI. [2]
• Reporting is the first job people want to hand over. Citi reports 16% of family offices use AI for investment performance reporting, more than double the year before. [1]

If two or more of these fit, the check below is worth ten minutes.

How do I score it?

One point for each yes you can support with a current document or test record. No and unsure both score zero. The total runs from 0 to 10, and every question points the same way: yes means the evidence exists.

The ten questions

1. Have we written down one task and the person accountable for it?

Yes if a named person can show the task, the records it may use and what the output should look like. Citi finds the AI uses that succeed in family offices are everyday ones: summarizing documents, transcribing meetings, managing email and automating reports. [1] Starting with one of those, with one owner, keeps the rest of this check concrete.

2. Do staff have a written rule for which records may go into the chosen AI tool?

Yes if the rule names permitted and excluded records and someone owns exceptions. Citi calls data privacy non-negotiable for family offices. [1] Among the breached organizations IBM studied, 63% had no AI policy or were still writing one. [2] Our one-page family office AI policy is a starting template.

3. Can someone explain the training terms for the exact plan the office uses?

Yes if you have the provider's terms for that plan and know any opt-in setting. Terms differ by plan. OpenAI, for example, states that data sent to its API is not used to train its models unless the customer opts in, which says nothing on its own about a personal chat account. [3] Our guide to whether your office's data is being exposed through AI sets out each major provider's terms by plan.

4. Have we separated what is kept from what is trained on?

Yes if your evidence lists every place the proposed setup keeps data. On its API, OpenAI does not train on your data by default, yet it keeps abuse monitoring logs, which can include prompts and responses, for up to 30 days. Excluding content from those logs currently needs OpenAI's prior approval. [3] "Not used for training" and "not kept" are different answers.

5. Do we know every service that receives the task's data?

Yes if you can name every company that sees the data: the assistant, the model provider and anyone else they pass it to. A brand name on a contract does not tell you that. Every extra recipient is another set of terms to check, so question 3 applies to each one.

6. Have we tested restricted access with harmless records?

Yes if a test with invented records shows the expected result for each role. OWASP recommends minimum permissions, acting in the user's own context and enforcing authorization in the downstream system rather than trusting the model. [4] A polite refusal in the chat window is not that test.

7. Have we tested that sending needs the intended approval?

Yes if test records show nothing went out unapproved. OWASP's own example is a mailbox assistant that only needs to read mail, so it should not be able to send or delete, and it recommends a human approve high-impact actions. [4]

8. Can a reviewer trace a reported figure back to its accepted source?

Yes if a sample report keeps the source, the period and who accepted it, or if reporting is outside the pilot. Addepar describes automated verification checks that flag data issues for its team to investigate. [5] Any AI that drafts reporting should go further and show where each figure came from.

9. Is a qualified person assigned to settle missing or conflicting evidence?

Yes if open items have an owner and cannot quietly become accepted answers. Citi notes that family offices keep final decisions, especially on investments, with experienced people. [1]

10. Do we have an agreed response when deletion cannot be checked directly?

Yes if the office records what it can inspect, what rests on contract or attestation, and who decides whether that gap blocks use. OpenAI's own page shows data can sit in more than one store. [3] A search that returns nothing does not prove erasure.

What does my score mean?

These bands are our editorial guidance, not research findings.

• 0 to 3: start with a written policy and one task. Name the owner, the permitted records and who approves. Keep live family information out until that exists.
• 4 to 7: close the open answers. Give every no or unsure an owner and a date. Test with invented records while terms, permissions and retention are checked.
• 8 to 10: review the evidence, not the total. Have the accountable operator, whoever handles security, and counsel read the open limits before any live pilot. Reporting work also needs its reporting owner or accountant.

At any score, a single unresolved answer on access, unapproved sending or data handling can be a reason to stop. A yes elsewhere does not offset it. Run the check again when the provider, the plan, the task or the data changes.

For the vendor conversation itself, use our five questions to ask any AI vendor before it reads family email. For reporting, read consolidated reporting across custodians.

Where does Private SuperIntelligence fit?

Private SuperIntelligence is private AI set up across the family's own email, entities and documents, deployed in an environment built for that family alone. It drafts freely while sending and payments need a person, and consolidated reporting cites every figure, so questions 7 and 8 become settings you can test rather than promises. [6] It does not replace questions 1, 2 and 9: those decisions stay with the office.

Speak with the founders

Private SuperIntelligence takes a limited number of families at a time so every one gets the same standard of service. Book a private call to talk through your office with Alex Morris.

Disclosure: Prakshal Jain and Alexander Morris co-founded Mitosis Labs, which runs Private SuperIntelligence.

Frequently asked questions

Does this page calculate or save my score?

No. There are no answer fields, no automatic scoring and nothing to submit. You count your own yes answers.

Does a score of 10 mean our AI setup is safe?

No. The check is a planning aid, not a validated test or a certification. Qualified people in the office still decide whether a proposed setup meets its requirements.

How is this different from the AI data exposure check?

The AI data exposure check looks at where information goes under a provider's published terms for the tools your staff already use. This readiness check looks at the office's own evidence before it connects any AI to live records.

Who in a family office should run this check?

The person who will own the AI decision, often the COO, chief of staff or general counsel, with the principal seeing the result.

Sources

[1] Citi Institute. AI in the Family Office: Privacy, Efficiency and Institutional Rigor. May 11, 2026, citing Citi Wealth's 2025 Global Family Office Report. Accessed 2026-10-16.

[2] IBM Newsroom. IBM Report: 13% of organizations reported breaches of AI models or applications. July 30, 2025. Accessed 2026-10-16. Figures describe breached organizations studied by IBM, not family offices specifically.

[3] OpenAI. Data controls in the OpenAI platform. Accessed 2026-10-16. API terms only, not every OpenAI product.

[4] OWASP Gen AI Security Project. LLM06:2025 Excessive Agency. Accessed 2026-10-16.

[5] Addepar. Why Addepar. Accessed 2026-10-16. Vendor's own description, not an independent assessment.

[6] Private SuperIntelligence. AI for family offices. Accessed 2026-10-16.

Originally published at privatesuperintelligence.si.