DEV Community: PJ (Prakshal Jain)

What AI providers' own policy pages say happens to your firm's data

PJ (Prakshal Jain) — Fri, 16 Oct 2026 09:27:55 +0000

Disclosure: we 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, RIA or venture firm, the AI privacy question tends to land on your desk as one line: "Is our data safe in ChatGPT?" The honest answer is a matrix, not a yes or no. It depends on the plan each person is on, a training toggle, a retention clock, a subprocessor list and whatever connectors someone switched on.

We read each provider's own pages on 16 October 2026. Here is what they say, written for the person who has to turn it into policy.

The short version

Business plans from OpenAI, Anthropic, Google and Microsoft do not train on your content by default. The exposure sits elsewhere: personal accounts used for work, retention after deletion, outside processors, and old sharing settings that AI can now read.

Training and retention, by plan

Plan	Trains by default?	Retention
ChatGPT Free, Plus, Pro	Yes, unless the user opts out [1]	Deleted chats removed within 30 days, with legal and security exceptions [3]
ChatGPT Business, Enterprise, API	No [1]	Enterprise admins set retention; deleted chats within 30 days [2][4]
Claude Free, Pro, Max	Only if the user allows it, except chats flagged in a safety review [7]	Deleted within 30 days; if training is allowed, kept chats can be held de-identified up to 5 years [8]
Claude for Work and API	No, unless you send feedback or opt in [9]	API inputs and outputs deleted within 30 days, with exceptions [10]
Gemini Apps (consumer)	Yes while Keep Activity is on [11]	18 months default; 72 hours with Keep Activity off; human-reviewed chats up to 3 years [11]
Google Workspace with Gemini	No, without your permission [12]	Admin-controlled [12]
Microsoft Copilot (consumer app, updated 18 Aug 2026)	No [15]	Activity history until the user deletes it [15]
Microsoft 365 Copilot	No [13]	Activity history under your Microsoft 365 terms [13]

Three edge cases worth encoding in your policy:

Feedback is a side door. On OpenAI's personal plans, a thumbs up or down can make "the entire conversation associated with that feedback" eligible for training, even after an opt-out. [1]
Human review outlives deletion. Google's Gemini page says reviewed chats "are not deleted when you delete your activity" and are kept for up to three years. [11]
Litigation overrides retention. OpenAI was ordered to preserve consumer ChatGPT and API content in the New York Times case. The order ended on 26 September 2025, but OpenAI says it still stores limited historical data from April to September 2025. Enterprise and Edu were excluded. [4]

The dependency graph behind one prompt

A prompt does not stop at the AI provider. OpenAI's sub-processor list, last updated 14 October 2026, names Microsoft, CoreWeave, Oracle and Amazon Web Services among the cloud hosts for ChatGPT Business and Enterprise, and TaskUs in the Philippines for support and for content moderation on the API and ChatGPT Business. [5]

Microsoft 365 Copilot now lists Anthropic as a subprocessor, and Microsoft says it "enables Anthropic models on by default for most customers in commercial cloud," outside Europe and the UK. [14] If you run a US tenant and have not looked at that admin setting, your Copilot traffic may already reach a second model provider.

Connectors inherit your permission debt

Microsoft 365 Copilot "only surfaces organizational data to which individual users have at least view permissions." [13] That is correct access control, and it also means a SharePoint site shared too broadly in 2019 is now searchable in natural language.

In ChatGPT Business, OpenAI says "many apps are enabled by default," while new Enterprise workspaces start with a selected set. Enterprise orgs can ask OpenAI Support to stop staff connecting work accounts to personal ChatGPT. [6] For agents that act (send email, file documents), decide up front which actions always require a human. We would put sending money and sending email on that list.

Five tests for "private"

A no-training clause answers one question. Before the trusts, the deal room and the principal's inbox go anywhere near a model, get written answers to these:

Where it runs. Whose hardware, which location?
Who runs it. Who operates it day to day, and who do they answer to?
What reports home. Content, logs, usage data, or nothing?
Who sees what. Per-person record access, and which actions need a human?
Where answers come from. Can every answer be traced to a source document?

For any hosted service, the plain answer to the first two is the provider's data centers and the provider's staff. Hosted plans can still score well on the rest.

A runbook for this quarter

Inventory every AI account used for work last month, personal ones included.
Turn off training on any personal account that touched office work, or move that work to a business plan. [1][7][11]
For each provider, read three pages: training, retention, subprocessors. Put the numbers in your AI policy.
Fix sharing before you connect anything. [13]
Name the data that should never leave an environment the family controls.

Where we fit

Private SuperIntelligence is the service we built for the families who want all five tests answered in their favor. Each deployment is built for one client, from the hardware to the model provider to where the data is stored; a client deployment captures none of the client's data and sends no telemetry back to Mitosis Labs, and it can run air-gapped with the client holding the keys. [16][17][18] It answers with citations, drafts within permissions the office sets, and leaves sending and payments to a person. We take a limited number of families at a time. [18]

To see where your own office stands, take the AI data exposure check: five questions, under a minute.

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. Sources checked 16 October 2026.

Sources

Hosted workspace or managed deployment: choosing family-office AI when you are the one who has to run it

PJ (Prakshal Jain) — Fri, 16 Oct 2026 03:53:15 +0000

Disclosure: I co-founded Mitosis Labs, which runs Private SuperIntelligence.

If you are the technical person in a family office who gets handed the AI decision, you are not choosing a model. You are choosing an operating model. ChatGPT Enterprise is a hosted workspace that your administrators configure and govern. Private SuperIntelligence is a deployment that Mitosis Labs builds for one family and runs for it, from the hardware to the choice of model provider to where the data is stored, delivered in weeks. 1 3 7 8 They put operational responsibility in different hands, and that is the real decision.

Does this describe your office?

The team's week goes on summarizing documents, answering correspondence and rebuilding the same reports. 6
Nobody has the time or the expertise to set up AI and keep it running. 6
The answers you need sit across custodians, fund administrators, accountants and the inbox. 5
Staff, advisers and family members each need to see different parts of the family's records. 5
Before anything touches the trusts or the deal room, you want to know exactly where that information goes. 5

If two or more apply, keep reading.

Two operating models

ChatGPT Enterprise	Private SuperIntelligence
Setup: your administrators configure the workspace, invite users and manage integrations 3 4	Setup: Mitosis Labs builds the deployment around the family's entities, advisers, email and documents 7
Daily operation: your administrators review access and monitor use 4	Daily operation: Mitosis Labs runs it, with a named contact 7
Processing location: OpenAI's service, hosted with major cloud providers 1	Processing location: an environment built for this one family, from isolated cloud to air-gapped 7
Training on your data: off by default for business content 2	Training on your data: none. The provider keeps no copy of the work and trains nothing on it 8
Processing and telemetry: content is processed in OpenAI's service to answer each request 1 2	Processing and telemetry: a client deployment captures none of the family's data and sends no telemetry back to Mitosis Labs 8
Scope: your team's work in the workspace and the tools you connect 3	Scope: the whole office: reporting, entities and trusts, correspondence, memos and private questions 7

Illustration. Sources: OpenAI Trust Portal and admin guides 1 3 4 ; Private SuperIntelligence family-office service 7 .

OpenAI's protections are real and carefully documented: no training on business content by default, encryption in transit and at rest, and administrator controls. 2 They are the right protections for a hosted tool. The operator's question is different: govern a hosted tool yourself, or have an environment built for the family and run by people who answer to it.

The processing boundary

A no-training promise answers one question: will the provider learn from our content? An operator's second question usually matters more: where is the work processed, and who can see it?

With ChatGPT Enterprise, the work is processed in OpenAI's service, which OpenAI states is hosted with major cloud providers. Business content is not used for training by default. 1 2

With Private SuperIntelligence, the work stays in an environment designed for one client, down to the hardware, the model provider and the storage location. The provider keeps no copy of the work, trains nothing on it and cannot see the outputs, and the deployment sends no telemetry back to Mitosis Labs. 8 Deployments range from isolated cloud to air-gapped. 7

The boundary matters because of what sits behind it. Deloitte's 2024 report found that 43% of family offices had suffered a cyberattack in the previous 12 to 24 months, rising to 62% of offices with more than $1 billion in assets under management. 9 RBC and Campden Wealth's 2026 report identifies cybersecurity and data breaches as the top near-term operational concern, cited by 59% of respondents. 10

Source: Deloitte, Family Office Cybersecurity Report, 2024. Survey of 354 single-family offices, September to December 2023; overlapping groups 9 .

Who holds the pager

Buying seats is the easy part. The work is mapping the family's information, connecting the right sources, setting who sees what, testing on real office work and keeping it right as the office changes.

Citi's May 2026 article, citing its 2025 Global Family Office Report, says that 57% of family offices cite lack of internal expertise as their biggest barrier to AI adoption. 6 J.P. Morgan reports that average annual operating costs exceed $6.6 million for offices with more than $1 billion in assets. 11

With ChatGPT Enterprise, OpenAI's guidance assigns onboarding, access reviews and monitoring to the office's own administrators. 4 With Private SuperIntelligence, Mitosis Labs sets up the service and runs it for the family, with a named contact. The office keeps every judgment and every approval; the team handles the technology. 7 For an office that wants the whole service operated for it, that is the evidence that points to Private SuperIntelligence.

Answers you can open

An AI the office relies on has to show its work. Private SuperIntelligence answers with citations, so you can open the record behind any answer. 7 8 Mitosis Labs also publishes research on Cortex, the memory system behind the service. Its research page reports 96% correct answers, 0.0% incorrect answers and a 4.4-second answer time for Cortex in the Agentic Memory Index's 272-task evaluation. 12

Source: Mitosis Labs research, reporting the Verging Labs revision of 3 August 2026 12 .

When ChatGPT Enterprise is the right call

Plenty of offices should buy ChatGPT Enterprise. If your office wants a general assistant for staff, is comfortable with a hosted service, and has someone in-house to configure and govern it, ChatGPT Enterprise is an excellent choice. OpenAI documents its protections carefully, training on business content is off by default, and the tooling for inviting users, managing integrations and reviewing access is built for exactly that administrator. 2 3 4 It is a strong option for everyday drafting and summarizing.

If instead the office wants AI it can trust with everything the family knows, in an environment built for the family, set up and run for it, Private SuperIntelligence is built for exactly that. 7 8

Five questions for any provider

Ask these of us and of anyone else.

Training: will our content ever train or improve a model?
Processing: where exactly is our information processed, and who operates that environment?
Visibility: what can the provider's own staff see, and what does the deployment send back to them?
Permissions: can we decide which staff, advisers and family members see which records, and which actions need a person?
Operation: who sets it up, who keeps it running, and who do we call when something changes?

Expect clear, written answers to all five from whichever provider you choose.

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.

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

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

Private SuperIntelligence is powered by Mitosis Labs.

Adaptation A1 of the original article, first published at privatesuperintelligence.si on 15 October 2026.

Sources

Sources checked 15 October 2026. OpenAI statements are taken from OpenAI's own pages.

Source context: the Cortex figures describe a memory-system evaluation, not measured results from a Private SuperIntelligence client deployment or a comparison with ChatGPT Enterprise. The selected model provider and any external processing must be identified in the deployment design. Deloitte's percentages describe historical survey groups; RBC's release describes 155 responding offices across the Americas.