The PracticeThe PracticeThe PracticeOperating infrastructure

The Practice · Guides

Seven mistakes agents make with client data in AI tools

Verified October 29, 2026 against Anthropic's and OpenAI's own help, privacy and pricing pages. Labels change.

The task

Find and fix the seven ways client information slips out of an AI account through settings and habits. Each fix uses a control Anthropic or OpenAI documents on its own pages, so you can check it yourself.

This is practice guidance from a working practitioner, not legal advice. Your brokerage's policy and your broker's direction come first.

Who it is for

Any agent or team lead who has pasted client details into Claude or ChatGPT, or is about to. Each of them can happen on the first day an account is open.

Before you start

Open your account's settings in one tab and this page in another; the review takes about 15 minutes. Have your brokerage's AI policy at hand if one exists.

The seven mistakes

Each mistake below has what it looks like, why it matters, and the fix. They run from settings to habits, and the last one matters most.

1. Client work in a personal account with training left on

What it looks like: an offer summary pasted into a personal Claude or ChatGPT account whose training switch nobody has touched.

Why it matters: on personal plans the training switch can be on without your having chosen it. Anthropic says that with "Help Improve our AI models" on, it may keep data in de-identified form for up to five years in its training pipelines; OpenAI turns "Improve the model for everyone" on by default on its personal plans.

The fix: decide the setting before the first client detail goes in (Claude: Settings, "Privacy"; ChatGPT: Settings, "Data controls"). Know its edges: turning it off does not pull data out of training already under way, and in ChatGPT a thumbs-up or thumbs-down can send that whole conversation to training after you opt out. Rate responses only in chats with no client details.

2. Brokerage data in an account the brokerage does not know about

What it looks like: transaction files and the team's pipeline in your own personal subscription, while the brokerage pays for a workspace or has a written AI policy you have not read.

Why it matters: on Claude Team and ChatGPT Business, the vendors say they do not train on content by default, and the organization manages the account and its data. A personal account sits outside the brokerage's settings and outside its view.

The fix: ask your broker which AI accounts the brokerage provides or allows, and keep brokerage data there. Keep your own sphere and the brokerage's records in separate accounts, so a move between brokerages never means splitting one account in two.

3. A share link with a client in it

What it looks like: sending a co-agent a link to the chat where you drafted a counteroffer summary, or setting a ChatGPT Project to "Anyone with a link."

Why it matters: a public link from a personal Claude account lets anyone with the link read the chat as it stood when you shared it, and Anthropic says that although it asks search engines not to index shared chats, a link posted elsewhere can be found by anyone. In ChatGPT, a link from a personal account opens for anyone who has it, and a shared conversation can include uploaded files and images.

The fix: copy the text you need into an email instead. Review old links: Claude, Settings, "Privacy," under "Shared chats"; ChatGPT, Settings, "Data controls," then "Manage" next to "Shared links." Deleting a link stops new access; it does not remove copies someone already saved.

4. Pasting the whole file

What it looks like: the full contract packet, a pre-approval letter with account numbers, or a client's entire text thread, pasted in to get a three-line summary.

Why it matters: everything pasted becomes part of a conversation the vendor stores under its own retention rules. Anthropic removes a deleted conversation from its back-end systems within 30 days, OpenAI may keep temporary chats up to 30 days for safety, and OpenAI's own sharing page says not to put financial details, passwords or account numbers in shared content.

The fix: give the assistant the facts the task needs and nothing else: the dates and the price, not the packet. Remove account numbers, birth dates and Social Security numbers before anything goes in. If a task seems to need the whole document, it belongs with your broker or the closing team, not in a chat.

5. Letting memory carry one client into the next

What it looks like: a draft for one buyer that mentions another buyer's inspection deadline, because memory picked it up last week.

Why it matters: Claude's memory is on by default on Free, Pro and Max, and ChatGPT can reference saved memories and past chats. Both build memory from what you said before, which is useful for your preferences and wrong for a client's file.

The fix: keep client work inside a Project. Claude keeps a separate memory per Project (turn off "Use account memory" for full separation), and a ChatGPT Project created with project-only memory keeps your saved memories out. For a one-off matter, use Claude's incognito chat or ChatGPT's temporary chat; neither creates memories. Once a month, read what memory holds: in Claude, "Manage" next to "Your memory"; in ChatGPT, Settings, "Personalization," "Memory."

6. Connecting everything, approving everything

What it looks like: email, files and calendar all connected on the first day, and every approval prompt clicked through without reading it.

Why it matters: connected services let the assistant read your mail and files and, on both platforms, take actions in those services. Anthropic says connected services process data under their own terms, and OpenAI says ChatGPT may ask for approval before reading information or completing an action.

The fix: connect one service when a task needs it, not all of them on day one. Read every approval prompt, and approve nothing that sends, deletes or shares until you have read what it will do. Disconnect what you stopped using: Claude, "Customize," then "Connectors"; ChatGPT, Settings, then "Plugins."

7. Sending the draft before reading it

What it looks like: a market update or client email sent with a square footage, a school name or a date the assistant supplied and nobody checked.

Why it matters: an assistant states a wrong fact in the same confident tone as a right one. The client holds you to what you sent, not to what the tool wrote.

The fix: read every client-facing draft against its source before it leaves: the listing data, the contract dates, your own notes. Keep a standing instruction that anything the assistant cannot support is flagged, not filled in. Anything that turns on what a contract means goes to your broker or a lawyer, not to the assistant.

How you know it worked

Fifteen minutes with your account open answers it. Your training setting reads the way you chose, your shared-links list holds nothing with a client in it, and your memory holds no client details.

Your connected services are the ones you actually use, and your broker has told you which accounts are for brokerage work. If any line fails, the mistake above it has the fix.

Next rung

The Guardrails Pack is the free baseline for any professional's AI, and it arrives with the Letters. Subscribe once.

Revision history

1.0, October 29, 2026, first version. Settings and retention figures verified on Anthropic's and OpenAI's help and privacy pages.

Educational, not professional advice. Nothing on this page is affiliated with or endorsed by its publisher.