Picture the day your team stops asking whether a file may go into the assistant, because the answer is written down. Everyone knows which account to use, which kinds of information belong there and who signs off on the rest. From then on the AI you already pay for does real work on real records.
Getting there starts with one question, asked in every industry: is it safe to put our data into ChatGPT, Claude, Copilot or Gemini?
This article describes how these products and our workflows work. It is not legal, security or compliance advice. The people who answer for your company’s obligations decide what they require.
The short answer.
It depends on four things, and your company controls all of them: the account and plan the work runs in, what the assistant is given, what the workflow tells it to do, and who decides. No vendor, and no workflow, makes that decision for you.
The rest of this article takes them in that order.
What the account and plan change.
The same assistant handles your information differently depending on the account you sign in with. A personal account and an account your company manages are covered by different terms, and each vendor says so on its own pages. The differences that matter most are whether what you type may be used to train models, whether a person at the vendor may read it, how long it is kept and what your administrator can control.
Here is what each vendor’s own page said when we read it on 16 November 2026. Vendors change these pages, so the linked page is the authority and this table is not.
| Assistant | Personal accounts | Accounts your company manages |
|---|---|---|
| ChatGPT | OpenAI sets out what applies to personal accounts in its data controls FAQ. Read it there. | OpenAI sets out its commitments for business plans on its enterprise privacy page. Read them there. For its API, OpenAI’s documentation says data sent is not used to train its models unless the customer opts in. |
| Claude | Anthropic’s privacy center says chats on its Free, Pro and Max plans are used to improve its models if you choose to allow it, or if a conversation is flagged for safety review. | Anthropic’s privacy center says that by default it does not use inputs or outputs from its commercial products, such as Claude for Work, to train its models. |
| Copilot | Microsoft’s privacy FAQ, which it says covers the older version of its Copilot app, says it uses Copilot conversations for AI training, except for certain categories of users and people who have opted out, and that you can opt out at any time. For Copilot on the web and the updated app, Microsoft points to its page for individuals. Read it there. | Microsoft’s documentation says prompts, responses and the company data Copilot reaches are not used to train foundation models, and that Copilot shows a person only the company data that person already has permission to view. |
| Gemini | Google’s Gemini Apps Privacy Hub says a setting controls whether chats are used to train its models, that human reviewers read some chats, and asks people not to enter confidential information they would not want a reviewer to see. | Google’s Workspace FAQ says your content is not used for model training outside your domain without permission, and that submissions from people with a Google Workspace with Gemini license are not reviewed by human reviewers. |
Two things follow. An employee using a personal account for company work is on different terms from the ones your company agreed to. And a business plan sets the terms without deciding which of your information belongs there. That decision is still yours.
What goes into an assistant, and what does not.
An assistant sees what a person gives it and what it has been connected to, and nothing else. There are three ways information arrives.
- Files and text a person supplies.Someone uploads an export or pastes a document. This is how most workflows start, and the person chooses what goes in.
- A connection to a business system.The assistant reads from the system and, where it is allowed to, writes back. What it can see and change is set in the connection and the system, and whoever approves your systems approves each one.
- What the assistant looks up for itself.Web search and similar tools send a query outside the conversation. Vendors can treat those queries differently from the conversation itself. Microsoft’s page on enterprise data protection, for one, describes web search queries separately.
Most work needs far less than people expect. A receivables summary needs the aging report, not the whole ledger. A draft reply needs the customer’s question, not the account history.
What a workflow’s guardrails do.
A workflow’s guardrails are written instructions that narrow what the assistant does with what it is given. Every workflow template names what the person supplies, what comes back, and who decides. A workflow can be told which fields to leave out, to stop and ask when something is missing, and to prepare a draft without sending it.
Guardrails shape behavior. They are not a lock. What an assistant is able to reach is set by the account, its permissions and your administrator’s settings, in the systems that enforce them. Our article on permissions and prompt injection covers what to inspect and restrict before a workflow touches a business system.
What a person still decides.
Three decisions stay with people in your company, whichever assistant you use.
- Which information the workspace is approved for.Your company decides which kinds of information may go into which account, and writes it down.
- Who may run the workflow.Your company decides who runs it, usually the people already allowed to see the records it works from.
- What happens to the result.A named person reviews what comes back and decides whether it is sent, filed or acted on. AI prepares. A person approves.
Every field asks it in its own words.
The question is the same everywhere, and each field asks it about its own records. A law firm asks about client confidentiality. A school asks about student records, and a medical practice about patient privacy. A financial adviser asks about the records the firm has to keep, a public agency about public records, and a hardware company about designs it has promised to keep private. People teams ask about employee records. Research teams ask whether sharing an unpublished result changes what they can later claim for it.
Those are questions about your obligations, and the people who answer for them in your organization are the ones to decide: your general counsel, your compliance or privacy lead, your records officer or your outside adviser. Whether a vendor’s terms fit your situation is a question for that person and the vendor.
The questions approvers ask.
These are the questions that come up when a company approves an assistant for real work.
- Which account will people sign in with, and is it one the company manages?
- What does the vendor’s own page say about training, human review and retention for that plan today?
- Which kinds of our information are approved for that workspace, and which are not?
- What does this workflow need, and what can it leave out?
- If it connects to a system, what can it read and what can it change?
- Who reviews the result, and who is allowed to send it?
For how we handle your information when we build a workflow with you, read Is our data safe? and our security page, which is written to share with the person who approves data use in your business.
Once those answers are written down, your team can put the assistant to work on the records that matter. That is how you 10X the ROI of your existing AI.