Is ChatGPT safe for business? What to check

Most small businesses do not decide to use ChatGPT. They find out that they already do. Someone in billing drafts a hard email with it, someone at the front desk pastes in a customer complaint, and a manager asks it to tidy a spreadsheet with client names in it. The question is not really "is ChatGPT safe?" It is "is the way our team uses it safe?" This guide covers what to check.

Short answer: it depends on the plan and the rules around it

ChatGPT is a set of plans, and they treat your information differently. The biggest differences are whether your chats can be used to train the vendor's models, who controls how long chats are kept, and whether the business can see and manage who is using it. A business plan with sign-in controls and a written rule is a very different thing from a staff member's personal account on their phone.

So the safe answer for a business is: use a plan the business controls, turn on the settings that matter, write down what may and may not go in, and find out where people are already using personal accounts.

Data controls by plan

Here is what OpenAI's own pages say, as of October 2026. Plan names and terms change, so check the linked page before you rely on a line.

• Personal plans (Free, Plus and Pro). OpenAI's help center says that on these plans data sharing is enabled by default, and the user can opt out. The person who owns the account controls that setting, not your business.
• Business, Enterprise and Edu. The same help article says that with ChatGPT Business, Enterprise, Edu and the API platform, OpenAI does not use inputs and outputs to train its models by default.
• Single sign-on. OpenAI says SSO and domain verification are included with ChatGPT Business, so staff can sign in with their work account. OpenAI notes that SSO controls how people sign in, while workspace membership is managed separately, so when someone leaves, remove them from the workspace as well as from your sign-in system.
• Retention. OpenAI's enterprise privacy page says that for Enterprise, Edu and ChatGPT for Healthcare, workspace admins control how long data is kept, and deleted conversations are removed from its systems within 30 days unless the law requires otherwise.
• Audit logs. OpenAI's Compliance Platform is available to Enterprise and Edu workspaces. It gives access to logs and metadata that can be connected to eDiscovery, data loss prevention (DLP) or security monitoring tools.

If patient information is involved, the plan question is stricter, because it turns on whether there is a business associate agreement (BAA). See Is ChatGPT HIPAA compliant? for which plans come with one.

Training on your data: the setting to check

"Will it learn from our data?" is usually the first worry. On a personal plan, the setting is called Improve the model for everyone. OpenAI's Data Controls FAQ says that when it is off, new conversations are not used to train OpenAI models. It also describes Temporary Chats, which are not used to improve OpenAI's models and may be kept for up to 30 days for safety purposes.

Two things are easy to miss. Turning the setting off changes new chats, not old ones. And it only covers the account it was turned off in. You cannot check it for a personal account you do not manage.

Training is also not the only risk. A chat that is never used for training still sits in someone's history, in an account the business does not control, with no record your business can review.

Staff personal accounts are the real gap

A business can buy the right plan and still have its client data in personal accounts. It happens because the personal account is already open on someone's phone, and it is quicker. That chat is outside your admin controls, your retention settings and your logs. When the person leaves, the history leaves with them.

This is why "we bought ChatGPT Business" does not settle the question. You also need a rule that work information goes only into the approved workspace, and some way to find out whether the rule is being kept.

"Anyone can paste client data and we would never know"

That worry is common, and it is fair. A chat box accepts whatever is pasted into it. Controls on the plan decide where the data goes and how long it stays. They do not stop someone from pasting a client file in the first place.

What helps is a mix of a clear rule, individual sign-ins so usage belongs to a named person, and logs reviewed by someone who owns the job. On plans that offer them, audit logs can feed tools your IT provider already runs. On plans that do not, ask your IT provider what they can see from sign-in records and network logs.

ChatGPT vs Copilot: the same questions apply

People often ask whether Copilot is safer than ChatGPT. The honest answer is that the questions are the same for both. Which plan is it? Does it train on our data by default? Who controls retention? Can it see files our staff should not see? Who reviews the logs? The tool that answers those well for your setup is the safer one for you.

Write an acceptable-use policy

A one-page rule beats a long one nobody reads. It should name:

• The approved tools and the plan for each.
• What must never go in: client and patient details, passwords, payroll, anything under a confidentiality agreement.
• That work happens in the work account, never a personal one.
• That a person checks AI output before it reaches a customer.
• Who to ask when someone is not sure.

Our AI policy template gives you a first draft to edit.

Find the shadow AI you already have

Write the policy against what people actually do. Unapproved AI use is called shadow AI, and asking the team tends to under-report it. Sign-in records, expense claims and network logs show more. How to detect shadow AI walks through where to look and what to record.

When the work needs your own systems

A lot of what staff paste into a chatbot is really a task on your own systems: summarise this client's history, draft a quote from past jobs, pull together the weekly numbers. Copying that data into a chat window is where the risk comes from.

EdgeKeeper takes a different route. It is our software on a dedicated Microsoft Azure server we manage, with no public internet address and restricted outbound access, connected to the systems you choose. Each question runs read-only in a temporary sandbox, a person approves specific changes, and access is recorded for review. Your data stays yours, with full export rights. The security page explains the controls and their limits, and Private AI for business covers what "private" should mean.

This guide is educational. It is not legal advice and does not replace your IT provider or counsel.