Gemini now keeps "history off" chats for about two months. Write down what staff may paste into an AI assistant
Someone on your team has pasted a customer's email into a chatbot this month, and turned history off so it would disappear. For Gemini's app, disappearing now takes about two months instead of three days. It's a good week to write the rule down.

Google says Gemini app chats with history off, and temporary chats, will now be kept for around two months instead of 72 hours, and it has split chat history from AI training into two settings. Whatever your staff use, write a short rule: what may go into a consumer chatbot, what only into a work account your admin controls, and what goes into neither. Then name one person who keeps it current.
What Google changed
On November 16, Google said it's giving people more control over whether Gemini chats train its models. There are now two settings where there was one: one for keeping your chat history, one for letting Google use your conversations to train its AI. You can keep your history and still opt out of training.
The other half of the announcement got less attention. If you turn History off, future chats will be saved for around two months. Google says they're kept to respond to you, process your feedback, and protect Google, its users and the public. TechCrunch, citing an archived copy of Google's Privacy Hub from November 12, reported the earlier window was 72 hours. The rollout reaches the Gemini app globally over the coming weeks. Google didn't tell TechCrunch what prompted the change, and says temporary chats aren't used for training.
For comparison, TechCrunch notes that OpenAI says it may keep temporary chats for up to 30 days, and Anthropic keeps incognito chats for 30 days by default, with both companies able to keep some data longer for legal or safety reasons.
Why a small business should care
Nobody is claiming Google is doing something wrong. The point is smaller and more useful. "I turned history off so it's gone" was already a shaky assumption, and for one of the most popular assistants it's now a two-month assumption.
Think about who in your business has typed a customer's name, a complaint, a draft contract, an employee's situation or a price list into a chatbot. They probably did it to save time, and they probably turned something off to feel careful. Turning history off changes what you see, not what the provider keeps.
The reasonable response isn't a ban that everyone quietly ignores. It's a short rule that tells people what's fine, what needs a different tool, and what stays out entirely.
Turning history off changes what you see, not what the provider keeps.
The rule, written out
Here's a version you can paste into a handbook and edit. The square brackets are for you.
Using AI assistants at [Company]. You may use [approved assistant, on your work login] for drafting, summarizing and brainstorming. Do not paste customers' or employees' personal details, health or financial information, passwords, API keys, signed contracts or unpublished prices into any consumer chatbot, with history on or off. If a task needs that information, use [the approved tool] or ask [name]. If you pasted something you shouldn't have, tell [name] the same day. Nobody gets in trouble for telling us. [Name] reviews this rule every [quarter] and when a provider changes its terms.
Read it line by line.
"On your work login." A consumer account and a work account aren't the same product. In Google Workspace, for example, an admin can allow or block temporary chats and chat deletion, and if the company uses Google Vault, its retention rules apply even to temporary and deleted chats. Users see a message saying their activity might still be visible to admins or Vault users. Google's announcement is about the Gemini app, and the sources we read don't say how the two-month window applies to Workspace accounts, so check with your admin before you assume either way.
"With history on or off." This is the line the news makes necessary. Without it, people read "off" as "safe". It also covers the next assistant someone installs, whose defaults you haven't read yet, because the rule is about what goes in rather than about any one setting.
"Nobody gets in trouble for telling us." The rule only works if people report slips. Say so in writing, and mean it the first time someone tests it.
"Reviews this rule every quarter." Providers keep changing retention, training and defaults. One setting became two this week; next quarter it may be something else.
How the work moves
Three questions before anything is pasted
- Is it about a person?
A customer, a patient, an applicant or an employee. If yes, stop and read the rule.
- Is it ours or theirs?
Contracts, quotes and unpublished prices belong to someone, so check who.
- Which account is open?
A personal login and a work login are different assistants with different keepers.
- Could you say it aloud?
If you wouldn't read it to a stranger on the phone, don't type it.
What to do with the rule
A rule that lives in a document nobody opens doesn't change anything. Put it where the work happens: in onboarding, in the channel where people ask AI questions, and in the template people use for customer replies.
Announce it with a short message rather than a policy memo. Something like: "Google now keeps Gemini chats for about two months even with history off. Until we say otherwise, customer and staff details only go into [approved tool]. If you've pasted something sensitive into a chatbot recently, tell [name] today; nobody is in trouble." That last sentence does more work than the rest. The person who pasted a customer complaint into a free chatbot last month is the one you most need to hear from, and they will only speak up if it's safe to.
When someone does speak up, write down what was pasted, into which tool and on which account, and decide together whether the customer needs to hear about it. Most of the time the answer is no. The record is what lets you answer calmly if a customer ever asks. Our note on the permissions test for AI assistants is a useful companion, because it asks what an assistant can do once it's connected to your accounts, and this rule asks what goes in before that.
Then do one thing most small businesses skip: ask which approved tool people will actually use. If the approved option is slower than pasting into the free app, the rule loses. Choose something people prefer, and set it up so you decide what's logged and for how long.
When to build your own
Some work really does need customer information in the loop: summarizing support tickets, drafting replies from an order history, extracting fields from documents. For those jobs a general chatbot is the wrong place to start, and a small assistant built into your own systems, with its own permissions, logs and retention, is the right one. Model providers' API terms have their own retention rules, so read those too. Designing that assistant, deciding what it can see and what it must never see, is what our AI applications and agents practice does. If you'd like to start with the rule and one real use case, talk to us.
Related work: Sutton|Place
Decision guide
Same chat, two kinds of account
| Consideration | Consumer Gemini app | Work account with an admin |
|---|---|---|
| Temporary chat | Kept for around two months by Google, per its announcement | Your admin turns temporary chats on or off; Vault rules can still retain them |
| Deleting a chat | You can review and delete history yourself | Your admin decides whether users may; Vault can keep the data hidden but retained |
| Who decides | Each person, in their own settings | Someone with the Gemini settings admin role |
| Who can look later | Google, for the retention window | Workspace admins or Vault users, per the message users see |
We build alongside the people doing the work.
Bring us the problem you’re thinking about.