Is Claude Safe for Business Data? The Settings and the Canadian Rules
11 min read
Claude is safe for business data when three things are true: the training switch is set on purpose, sensitive details stay out of the prompt, and you know which plan's terms you agreed to. This lesson covers all three, plus what Canadian privacy law expects.
Owners tend to land in one of two camps. Some paste the whole customer file into a free account without a thought. Others refuse to touch AI because "it steals your data". Both skipped the ten minutes of reading that settles it.
Does Claude train on your data?
It depends on the plan.
| Plan | Are your chats used to train models? |
|---|---|
| Free, Pro, Max | Yes, unless you turn the setting off |
| Team, Enterprise | No, by default |
Read from Anthropic's privacy pages in October 2026. The pricing page describes model training on personal plans as "opt-out", which puts the burden on you.
The switch is called Help Improve our AI models, under Settings > Privacy. With it on, Anthropic keeps chats for up to five years. With it off, its standard 30-day retention applies. Changing the switch affects new chats and old ones you reopen. A few details worth knowing:
- Deleting a conversation keeps it out of future training.
- Incognito chats aren't used for training and aren't saved to memory or history, whichever way the switch is set.
- Clicking thumbs up or thumbs down sends that conversation to Anthropic as feedback, and feedback can be kept for up to five years. That applies on business plans too.
- Chats flagged for safety review can be used to improve Anthropic's safeguards.
Which position for the switch?
For a business, off. You gain nothing commercially from adding your quotes and customer emails to a training set, and the shorter retention is a point in your favour if a client asks how their information is handled. Set it on every account your staff use, including the free one someone opened at home.
What stays out, whatever the settings
A setting controls what Anthropic does with your chats. The information has still left your building. Keep these out of any prompt, on any plan, unless you have a signed agreement that covers them:
- Card numbers, banking logins, passwords
- Social insurance numbers and other government ID
- Health details, and anything about a minor
- A client's confidential documents, where your contract or your regulator says they stay with you
For everyday customer messages, strip the identity and keep the situation. "Customer's furnace failed two days after our service call, she wants the visit refunded" drafts the same reply as the version with her name, address and phone number. Protecting customer data when you use AI covers the stripping habit in detail.
The Canadian rules on one page
PIPEDA holds you accountable for personal information wherever it goes. Three points apply to Claude.
A prompt is a transfer to a processor. Anthropic is a US company, and as of October 2026 the Claude apps give you no setting for where chats are processed. Canadian law allows you to use a processor outside the country. You stay responsible for the information, and your privacy policy should say that some of it may be processed outside Canada.
De-identify where you can. The federal and provincial privacy regulators published principles for generative AI in December 2023. They tell organizations to use anonymized or de-identified information in prompts wherever the job allows.
Quebec asks for more. Under Law 25, sending personal information outside Quebec calls for a privacy impact assessment first. If you serve Quebec customers and want real names in your prompts, talk to a privacy professional before you start.
This is orientation, not legal advice. A clinic, a law office or a bookkeeping practice answers to a regulator with its own rules, and those rules outrank this lesson.
Connectors raise the stakes
Once you connect Gmail or Google Drive in Chapter 3, Claude can read what's in them when you ask. That's the reason to connect, and it also means a request like "summarize this week's customer emails" now moves real names through the assistant. Decide the training switch before you connect anything. Connect a work account only, and leave out the inbox that holds your family's medical appointments.
When the business plan is the honest answer
If staff paste customer details into Claude every day, a switch on each person's account is a weak control. People forget, and new accounts appear. The Team plan puts every seat under commercial terms with training off by default and an owner who can see the accounts. That's the one place this course suggests paying for privacy reasons. The comparison is in Chapter 5.
A five-minute check
- Open Settings > Privacy on your own account and set the switch.
- Ask each person who uses Claude for work to do the same and send you a screenshot.
- Write down what stays out of a prompt. The one-page AI policy template has the list ready to fill in.
- Add one line to your privacy policy about processing outside Canada.
Try it now
Do step 1 before the next lesson. Then scroll back through last week's chats and look for a customer's full name. If you find one, you've found the habit to change this week.