ChatGPT Free can use the client information you submit for model training unless you switch off that use in your account's privacy settings. ChatGPT Business does not use business content for training by default. Neither position, by itself, gives you permission to submit confidential client information or guarantees that nothing is stored.
Paying for an individual ChatGPT plan is not the same as using ChatGPT Business. Consumer plans have training controls you must check. Start with the account and workspace actually open on the screen, then decide whether the information belongs there. A company payment card is not a privacy setting.
The answer changes with the account, not the prompt
Writing “confidential, do not train on this” inside a prompt is not how you set the service's data controls. Put those controls in place before submitting the information. A promise generated in the chat is not a substitute for the provider's terms or your account settings.
| Route | Training position | What an owner should check |
|---|---|---|
| ChatGPT Free | Consumer content may be used for training; an opt-out is available | The model-training switch in privacy settings, before client work |
| Individual paid ChatGPT plan | Paying does not remove the need to check consumer training controls | Actual settings and whether the account is approved for the task |
| ChatGPT Business | Business content is not used for training by default | The correct workspace, allowed data, sharing and enabled services |
| OpenAI API | API data is not used for training by default | The complete application, including other suppliers and their storage |
The API is the route software uses to call a model directly. It is a separate service from pasting text into ChatGPT. ChatGPT Plus or Pro subscriptions do not include API use; API billing is separate. You do not need to build an API workflow just to draft an ordinary email with approved information.
The table describes training, not blanket permission to process data. Even on Business, check your client agreement and internal rules. If the information is sensitive, or your obligations are unclear, ask your data-protection adviser or solicitor before using it.
Use the tutorial on switching off AI training on business data for a wider comparison across tools. Here, the important distinction is between consumer ChatGPT controls and the default protection for business content.
“Not used for training” answers only one question
Training means using material to help develop a model. Processing means using your input to perform the task you requested. Storage means keeping information for some period. Sharing concerns who else can access it. Those are different activities, so ask about them separately.
A no-training commitment does not mean an answer is produced without processing your input. It also does not establish zero retention, meaning no customer content is kept under the relevant service terms. Do not tell a client that their data is “never stored” because you have bought Business.
Official OpenAI documentation confirms the Business no-training default and explains that retention depends on the data category and product configuration. Conversations, uploaded files and connected systems need their own checks. Avoid applying one remembered deletion period to every feature.
Memory is another separate consideration: information an assistant uses to personalise later interactions is not the same thing as training the underlying model. Review the available memory controls in your account. Do not assume that changing the training preference also settles memory, sharing or deletion.
Temporary Chat is not a promise of instant disappearance. OpenAI's current guidance allows retention for up to 30 days. If a client's condition is that information must not be retained, do not use the word “temporary” as your approval evidence. Check the exact requirement and the service's current terms.
For example, an illustrative laboratory owner wants to summarise a client complaint. Removing the client's name leaves an unusual test, a precise collection time and a distinctive incident. That combination may still identify the client to someone with other information. Training switched off does not make unnecessary details necessary.
The owner can instead ask for a generic complaint-response structure using a fictional scenario. If actual case details are essential, the laboratory needs an approved processing route. The tutorial on what zero data retention really means explains the more demanding supplier questions.
An import-export business checks fifteen routine prompts
Consider an illustrative six-person import-export business. Four office staff use ChatGPT for email drafts and internal notes. The owner inspects fifteen recent task descriptions without copying their confidential contents into another AI tool: eight need only public or invented examples, four concern shipment delays, and three involve private pricing negotiations.
The eight generic tasks can use fictional data. For the four delay emails, staff need a delayed dispatch date and an agreed next-update time, not an entire purchase order. The three pricing negotiations remain outside the pilot while the owner checks client restrictions and whether those details should be processed at all.
The business chooses four ChatGPT Business Standard seats for the approved office workflow. The list price in USD is $25 per user per month on monthly billing, so four seats cost $100 a month. At $20 per user per month billed annually, the equivalent is $80 a month, or $960 a year. These are subscription costs, not a prediction of savings.
The owner allocates two hours to inspect the tasks, agree permitted data and confirm the account setup, plus thirty minutes with each of the four staff members. That totals four staff-hours. At an assumed internal rate of $30 an hour, the setup allowance is $120. Add any advice or supplier review the business needs; those costs are not included in this illustration.
The first approved prompt uses a fictional shipment. Staff draft the email, compare every promise with the facts supplied, and add the real recipient details in the normal business system. For the pilot, the owner checks all four delay-email drafts before they are sent.
The owner writes the permitted data down as a short table the staff can keep open, rather than a paragraph of policy. The filled-in version for this business:
| Task | Allowed in the Business workspace | Keep out |
|---|---|---|
| Shipment delay email | New dispatch date, cause in one phrase, next update time, [customer] placeholder | Purchase orders, prices, the customer's own correspondence |
| Supplier chaser | Order reference, expected date, what is late | Other suppliers' prices or terms |
| Internal meeting notes | Agenda, decisions, action owners | Individual staff performance or health matters |
| Pricing negotiation | Nothing yet; outside the pilot | Everything, until client restrictions are checked |
The "Keep out" column does most of the work. Staff rarely need telling what to include; they need telling which tempting attachment to leave behind.
Success is not just a lower subscription bill or a faster draft. Each task should have an approved account, a clear reason for the data supplied, a human check and no unnecessary attachment. Count how often staff still need to ask for clarification. That tells you whether the rule is usable.
Practise with a prompt that needs no client identity
Here is a separate illustrative example for a packaging supplier. A staff member initially plans to paste an email chain with names, a delivery address, prices, order references and comments from two other customers. The actual task is to write a calm update about one production delay.
Prepare the reduced facts manually in an approved system. Do not submit the full email chain to ChatGPT and ask it to anonymise the information after the upload. The disclosure would already have happened.
Draft a brief customer update using only these fictional facts.
Keep the tone calm and direct. Do not invent compensation,
delivery guarantees or reasons for the delay.
The printed cartons will be dispatched on Thursday instead of Tuesday.
The cause is a production delay.
We will provide the next update by 15:00 tomorrow.
Use [customer] and [order reference] as placeholders.
Ask the customer to tell us if this changes their required schedule.Illustrative output: “Hello [customer], I am sorry that dispatch of [order reference] has moved from Tuesday to Thursday because of a production delay. We will update you by 15:00 tomorrow. Please let us know if this affects your schedule. Your order will arrive by Friday.”
The last sentence must go. No arrival date was supplied. The owner checks the remaining promises against the real order before staff fill in the placeholders and send the message. Data protection and factual accuracy are separate checks; a prompt with no client identity can still produce a bad promise.
For a more careful preparation process, use removing identifying client details before using AI. Replacing a name with “Client A” alone is not enough if the rest of the text still identifies the person or business.
Check the screen and the task before the next upload
For an individual account, open its settings and find the privacy or data controls. Check the switch that allows conversations to help train models, and turn that use off if it conflicts with your policy. Setting labels can change, so read the explanation next to the control. Make the change before submitting client information.
For Business, confirm that the task is taking place in the approved business workspace. Do not use the presence of a paid subscription somewhere in the company as evidence about the account currently open. Record the workspace and the permitted task types in your internal instructions.
The commonest slip happens when one person has two accounts in the same browser. In an illustrative design studio, a staff member signed into both a personal Free account and the studio's Business workspace, and drafted a client brief in whichever one the account menu last showed. Nobody noticed until the manager went looking for the conversation in the business workspace and it wasn't there. Nothing on screen had warned anyone, because both accounts look almost the same. The practical fix is to keep them apart: a separate browser profile, or a separate browser, for the business account, and a habit of checking the workspace name in the account menu before pasting anything from a client.
The written instruction matters as much as the setting. Before: "Use ChatGPT for customer emails, but be careful with sensitive data." After: "Use the studio's Business workspace, opened from the work browser profile. Check the workspace name before you start. Paste only the facts listed for that task in the permitted-data table. If the task isn't on the table, ask the office manager first." The second version names the account, the check and the fallback, so staff can follow it without guessing what "careful" means.
Keep client material out of optional feedback and data-sharing submissions unless those uses have been separately approved. Read what you are agreeing to before submitting a conversation as feedback. Do not assume a general training preference answers every question about a separate voluntary submission.
Connected apps deserve their own look. ChatGPT's connectors are now called apps, and on business plans an admin decides which ones staff can switch on. An app linked to a shared drive can reach whatever that staff member can open there. Picture an admin enabling a drive app so staff can ask for "the latest price list"; the answer also draws on a folder of another client's contracts that was shared with the whole team years ago. No training was involved, but a client's documents reached a chat they had nothing to do with. Enable only the apps a task needs, and tidy the drive permissions before connecting anything.
- Confirm permission. Does the client agreement and your own policy allow this information to be processed through this route?
- Reduce the input. Remove unrelated people, attachments, private comments and facts the task does not need.
- Check settings and connections. Verify the account's training controls and any services that could receive the information.
- Set the review. Decide who will check the output and where the approved result will be kept.
- Use the retention process. Follow your agreed record and deletion rules. Do not invent a universal expiry period.
In an illustrative home-care provider, a manager asks for a friendly reminder about an upcoming staff meeting. There is no reason to attach the weekly care rota. The approved task card says: “Use the meeting time, agenda and fictional recipient placeholders. No client records, visit details or care notes.” That is more useful than simply telling staff to avoid “sensitive data”.
A filled-in record might say: “Task: staff meeting reminder. Workspace: approved business account. Input: meeting details only. Reviewer: operations manager. Final record: staff communications folder.” The record explains the choice without reproducing private content in a second place.
An app using OpenAI still needs its own supplier check
“Powered by OpenAI” does not describe every company that handles the information. A third-party service may receive your document before calling the API and may keep its own records afterwards. Ask that supplier about its processing, storage and subcontractors; the model provider's training default is only one part of the answer.
For example, an illustrative spare-parts manufacturer considers a quotation assistant that sends order text to OpenAI through an API. The vendor says the API does not train on customer data. The manufacturer still asks whether full drawings, prompts and outputs are stored in the quotation service, who can access them, and how deletion works. Until those answers are clear, the trial uses fictional part numbers and quantities.
OpenAI's API data-control documentation distinguishes training from retention. It says abuse-monitoring logs may contain customer content and are retained for up to 30 days by default, subject to stated exceptions. Some features also keep information needed to run the application. Special retention controls require eligibility and approval; they are not a default consequence of paying for API calls.
Keep this check proportionate. You need answers about the actual route your data takes, not a technical diagram of every possible AI service. Ask the person supplying your application to explain that route in plain language.
If client information has already gone to the wrong account
Stop further submissions and record the account, approximate time, kind of information and any sharing or connected service involved. Do not paste the same material into a new chat to ask whether it was confidential. Tell the person responsible for data protection in your business.
Review the provider's current deletion and support processes. Changing a training setting now should not be treated as proof that earlier handling has been reversed. If the incident might affect a client commitment, get advice on the response and any notification requirements. Avoid promising the client an outcome you cannot verify.
Then fix the reason it happened. Was the approved workspace unclear? Did a staff member need a fast answer when the authorised tool was unavailable? Did the instructions say “use ChatGPT” without naming the account? The tutorial on preventing unapproved AI use by staff helps turn the incident into a practical change.
The useful purchasing choice is the one that supports an approved workflow. Free with training switched off may suit public or fictional exercises under your policy. Business supplies a no-training default for business content. Client information still needs a justified purpose, an authorised route and a person responsible for checking what goes in and what comes out.
Questions owners ask about ChatGPT and client data
What does ChatGPT Business cost for a sole trader?
ChatGPT Business requires at least two seats. At the Standard seat list price in USD, that means $50 a month on monthly billing, or $40 a month billed annually. An individual plan may cost less, but its training settings and business suitability need a separate check. Buying two seats does not itself authorise processing any particular client's information.
Can I ask ChatGPT whether a particular client message was used for training?
Do not treat its conversational answer as evidence about internal data handling. Record the account, applicable settings and time of submission, then use the provider's support or privacy process for the question. If the issue could affect client commitments, involve your data-protection adviser. Avoid pasting the confidential message again just to ask what happened to it.
Does annual billing change the Business training default?
Monthly and annual billing are payment choices for the Business subscription, not different training protections. Business content is not used for model training by default. Choose the billing term around your staffing needs and commitment, then check the actual workspace, enabled services and client permissions separately. Do not infer additional privacy controls from the size of the invoice.
Further reads
- What to Check in an AI Tool's Privacy Policy and Terms — Read supplier terms without mixing up separate data controls.
- Questions to Ask Before Buying AI That Touches Client Data — Prepare questions about any tool that receives client work.
- How to Keep Customer Data Private When Your Team Uses AI — Build wider controls around day-to-day customer information.
- Who Owns AI Chat History When an Employee Leaves? — Plan access and records when staff leave.
- Reusing Past Client Work With AI Without Leaking Client Data — Decide when past work can become a reusable reference.
- Is It Safe to Put Customer Data Into ChatGPT? — Which ChatGPT plans can take customer data, which data never goes in, and a two-minute routine for stripping identifiers when you must use a personal plan.
- Shopify Magic vs ChatGPT for Product Descriptions — Shopify Magic is free and sits in the description box; ChatGPT handles bulk rewrites and brand voice. A decision table, a fair test and sample outputs.
- Is AI Worth It for a One-Person Bookkeeping Business? — The break-even sum for a solo bookkeeper, three practice profiles with different answers, and a 30-day test before you commit to any AI tool.
- ChatGPT or a Legal AI Tool: Which Should a Small Firm Use? — Which legal tasks a business ChatGPT plan handles safely, which need a legal research tool, and how a three-lawyer firm split the budget.
- Advice-Specific AI or ChatGPT: Which Should an Advice Firm Use? — What adviser-specific AI adds over a business ChatGPT plan, one meeting written up both ways, and how a four-adviser firm split its budget.
- Is It Safe for an Accountant to Use ChatGPT With Client Data? — Which ChatGPT plans an accounting practice can defend for client data, what never goes in, and the engagement-letter wording to add.
- Can Solicitors Use ChatGPT Without Breaching Confidentiality? — Confidentiality, privilege and data protection are separate tests for a law firm using ChatGPT. What passes each, and the settings to fix first.
- Can Financial Advisers Use ChatGPT? What Compliance Allows — A traffic-light list of ChatGPT uses for advice firms, which plan passes a compliance review, how to keep records, and a one-page policy to adapt.
- Free vs Paid AI Tools for a Small Trades Business — Which trade jobs free AI handles fine, the three moments it stops being enough, and the cheapest paid upgrades to try first.
- Patient Data and AI: A Confidentiality Checklist for Small Practices — A 30-item checklist, with a filled-in example register, for keeping patient information confidential when a small practice uses AI tools.
- AI Tools and AI Development: The Complete 2026 Guide — the AI hub, including every tutorial in the AI-for-business series.
Sources: official OpenAI documentation, ChatGPT Work cloud security and Data controls in the OpenAI platform, reviewed 27 September 2026. ChatGPT prices and consumer privacy facts checked 27 September 2026. Examples and internal time allowances are illustrative.