How Family Offices Can Use AI for Research Without Revealing Holdings

For a family office, some of the most sensitive data is what the family owns, how much, at what cost, and what it plans to buy or sell next. AI research tools are useful for exactly the work family office analysts do, and every prompt is a small disclosure to a third party. Done well, AI for family offices delivers the research value without handing over the portfolio.

Why are holdings a special risk when a family office uses AI?

Holdings combined with intent identify the family and can move prices. A prompt that asks how to exit a large position in a thinly traded name, or how a pending private investment affects the family's tax picture, reveals far more than a general research question.

Some positions do become public. Under the SEC's Form 13F guidance, an institutional investment manager with investment discretion over $100 million or more in Section 13(f) securities files a quarterly report, due within 45 days after quarter end. That covers listed equity positions after a delay. It does not cover private company stakes, real estate, cost basis, entity structure or trades in progress, which are the details AI prompts tend to leak.

The regulatory backdrop is also thinner than at an RIA. A family office that meets the SEC's family office rule (advice only to family clients, wholly owned by family clients and controlled by the family, and not holding itself out to the public) is excluded from the definition of investment adviser. That means fewer prescribed safeguards, so the office's own policies carry most of the weight. Where the family also uses a registered adviser or broker-dealer, those firms have their own obligations, such as mapping sensitive customer information under Reg S-P, and counsel should confirm which rules reach the office.

What happens to the information analysts type into AI tools?

It depends on the account and the vendor's terms. Consumer accounts on many AI services can retain conversations and may use them to improve models unless the user opts out. Business and enterprise plans usually exclude customer data from training and give administrators control over retention, but the details vary by vendor and change over time. The same trade-offs apply to advisers using ChatGPT with client data.

The FTC's Office of Technology warned in February 2024 that companies quietly changing their terms to allow AI training on user data could be acting unfairly or deceptively, a sign that terms do shift. NIST's Generative AI Profile (NIST AI 600-1) lists data privacy and information security among the risks generative AI creates or makes worse, including leakage and unauthorized disclosure of sensitive data. Treat anything typed into an AI tool as shared with the vendor under whatever terms apply that day.

How can analysts research with AI without revealing what the family owns?

Separate the research question from the family's position. AI is good at summarizing filings, comparing business models and stress-testing a thesis, and none of that requires knowing who is asking or how much they hold. Practical rules analysts can follow:

• Ask about the company, sector or thesis. Avoid phrasing like "our position" or "we are considering."
• Leave out position sizes, cost basis, entry prices, account numbers, and the names of trusts, partnerships and family members.
• Never upload holdings reports, capital account statements, K-1s or investment committee memos to a general AI tool.
• Keep pending trades, deal terms and timing out of prompts entirely.
• Check confidentiality agreements on private deals before discussing a target or fund with any outside tool, AI included.
• When modeling, use illustrative figures unrelated to the real position.

Which controls make AI for family offices safe to use?

Good controls make the safe path the easy one and catch mistakes before they leave the office. Policy alone depends on every analyst remembering every rule on every prompt. The controls that matter most:

• One approved account: a business or enterprise AI plan with no training on your data, a defined retention period, and sign-in through the office's identity system with MFA. Personal AI accounts blocked on office devices.
• Labels on holdings files: sensitivity labels on portfolio reports, statements and investment memos so tools can recognize them.
• Data loss prevention: rules that block labeled files, account numbers and entity names from being uploaded or pasted into AI sites.
• Browser extension control: AI sidebar extensions can read whatever page is open, including portfolio accounting and custodian portals. Allow only approved extensions.
• Connector review: no AI tool connected to email, file shares or the portfolio system without a documented review of what it can read.

Does a family office need a written AI policy?

Not by regulation, but yes in practice. No SEC rule requires one for an excluded family office, and a short written policy is the practical way to set expectations and show the principals what is protected. One page can cover the approved tools, the data that never goes into a prompt, who approves new tools and connectors, and how often the policy is reviewed. Staff should acknowledge it, and new hires should see it before they get access to research tools.

What should a family office do this quarter?

Find out what is already in use, pick one approved path, and put guardrails around the holdings data. A workable sequence:

1. Inventory the AI tools, accounts and browser extensions staff already use on office devices.
2. Choose one business AI plan, read its training and retention terms, and save a copy of them.
3. Label holdings reports and investment memos, then turn on data loss prevention rules for AI sites.
4. Write the one-page policy with the prompt rules above and walk the investment team through it.
5. Schedule a review in six months to check what the controls caught and adjust.

How does s90 help?

s90 sets up AI guardrails for family offices: an approved AI account tied to your identity system, sensitivity labels on portfolio files, and data loss prevention rules that keep holdings, account numbers and entity names out of AI prompts. We also watch for new AI apps and extensions so the approved list stays accurate. If you want to give your investment team AI research tools without exposing the portfolio, we would be glad to talk it through.