DEV Community

Cover image for Shortlisting AI for a family office: map data reach, permissions and training defaults first
PJ (Prakshal Jain)
PJ (Prakshal Jain)

Posted on Originally published at privatesuperintelligence.si

Shortlisting AI for a family office: map data reach, permissions and training defaults first

#privacy #fintech #ai #security

Disclosure: I co-founded Mitosis Labs, which runs Private SuperIntelligence. This is an adaptation of the original guide.

If you are the technical person at a family office, "which AI tool should we use?" usually lands on your desk as a product question. I find it easier to treat it as a data-access question. Every tool on this list reads some set of the office's information, under some permission model, with some default about training and retention. Draw those three things for each candidate and the shortlist gets much shorter.

Everything I say about vendors below comes from their own pages, or for Aurora from a Fortune report, accessed 2026-10-15. Sources are at the end.

Boundary 1: what can the tool reach?

A family office holds a portfolio, the principal's email, entity and trust documents, the household and travel. The tools on offer split into four groups by how much of that they reach.

Category Examples What it reads Where its scope ends
Enterprise AI seats ChatGPT Business or Enterprise, Microsoft Copilot, Claude Enterprise What the office connects The office sets it up and governs it under the provider's terms
Platforms adding AI Addepar, Masttro, Asora, Eton, Aleta The portfolio's system of record Answers come from the data in the platform
Private chat apps Proton Lumo, Tinfoil, Venice What one person types No link to the office's own information
Concierge services Knightsbridge Circle, Aurora Lifestyle and travel requests Outside the office's records and entities

Microsoft says Copilot works across email, chats, documents and meetings in Microsoft Graph. Anthropic lists Claude Enterprise connectors to Gmail, Google Drive, Slack and Microsoft 365. So with enterprise seats, reach is a configuration decision your office owns.

The platforms reach the other way: deep into portfolio data, and stop there. Masttro's DocAI ingests PDFs for liquid and illiquid assets and handles capital calls, distributions and valuation statements. Eton says EtonAI does document extraction across 250+ document types, with natural language querying and a knowledge graph. Aleta reads capital calls, K-1s and fund statements. Addepar, in a March 2026 post, says Addison "draws on Addepar's unified data model to answer questions, surface relevant guidance and direct users to the right workflows." Each assistant answers from what sits in its platform, so the principal's inbox, the household's calendar, deal email and travel stay outside it.

Boundary 2: MCP is moving where the answer is produced

Two of the platforms now expose portfolio data to outside assistants through the Model Context Protocol.

  • The Asora AI Connector lets you query Asora wealth data from Claude, ChatGPT, Gemini, Grok or any MCP-capable assistant. Access is read-only, scoped to existing permissions, and every query is audit-logged.
  • Aleta supports MCP so AI agents can query the family's wealth data in plain language.

This is good design, and Asora is clear about the trade: it says it never uses your prompts or data for training, and notes that once an answer reaches your assistant, that provider's terms apply. For your data-flow diagram, that means the trust boundary now runs through the assistant you connect, and that assistant's retention and training settings matter as much as the platform's.

Boundary 3: permissions are inherited, including the old ones

Copilot is the clearest example. Copilot shows each user any content that user has at least view permission for. If a trust folder was shared with the whole company years ago, Copilot can surface it to anyone in the company. Copilot is respecting the permissions you already have, so old sharing settings on trust and entity folders matter. Review them before rollout.

ChatGPT Enterprise and Claude Enterprise list controls for managing access. OpenAI says ChatGPT Enterprise supports SAML SSO, admin-set retention and a Compliance API audit log. Anthropic lists SSO, SCIM provisioning, role-based access control, audit logs, data retention controls and customer-managed encryption keys for Claude Enterprise.

Boundary 4: training and retention defaults, including feedback buttons

  • OpenAI says it does not train its models on business data by default, that you own your inputs and outputs where the law allows, and that it limits its staff's access to incidents, permitted recovery and legal requirements. Eligible new Enterprise customers can store content at rest in regions including the United States, the United Kingdom and Europe.
  • Microsoft says prompts, responses and Graph data are not used to train foundation models.
  • Anthropic states it does not use inputs or outputs from its commercial products to train models by default. A conversation rated with thumbs up or down can be stored for up to five years and used for training. An organization owner can turn the rating feature off. Put that toggle on your rollout checklist.
  • Proton says Lumo does not log chats, uses zero-access encryption so "no one (not even Proton)" can read them, and never trains on your data. Tinfoil encrypts data to a secure enclave Tinfoil itself cannot access, verifiable through hardware attestation. Venice keeps history on the device and relays requests without storing them.
  • Eton describes EtonAI as a private, single-tenant instance and lists ISO 42001, ISO 27001, ISO 27701, SOC 1 Type 2 and SOC 2 Type 2. Aleta lists SOC 2 Type II and encryption at rest and in transit.

The private chat apps make strong promises for a single person's conversations. For an office they are do it yourself: none of them connects to your custodians, entity records or shared drive on its own.

Concierge services

Knightsbridge Circle describes itself as "a dedicated extension of your personal office," with a personal manager who looks after four members. Fortune reported on 2026-10-11 that Aurora is an invite-only app pairing AI with human experts for members' dining, travel, luxury goods and wellness, and that members can connect calendars and wearables such as Whoop and Oura so Aurora can anticipate requests. Both fit lifestyle and travel. Neither is described as working on a family office's records, entities or investments.

The checklist

  1. Which information should the AI reach: the portfolio, the principal's email, entity and trust documents, or all of it?
  2. Who in your office will set it up, review permissions and keep it current?
  3. Does the provider train on your data, and what do its feedback buttons do?
  4. Where is the data stored, and who at the vendor can see it?
  5. Do the answers carry citations your team can check?

Answer the first two for your own office. For the training question, policies differ by platform, so ask each vendor for its written policy.

Where Private SuperIntelligence fits

Private SuperIntelligence is an AI concierge for family offices. It works across the family's own email, documents and data, answers with citations, and acts within permissions the family sets. Each deployment is built for one client and works alongside the platforms above.

Private SuperIntelligence takes a limited number of families at a time so every one gets the same standard of service. Book a private call (ref BAF1-DEV) to talk through your office with Alex Morris.

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

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

Originally published at privatesuperintelligence.si. Vendor pages accessed 2026-10-15.

Sources

Top comments (0)