Customer data belongs in ChatGPT only when your organization has approved that specific data, purpose, product, account, and workflow. A business subscription or a training opt-out is not blanket permission. At Applied Margin, we start with the minimum information a task needs, verify the applicable terms and controls, and keep human review before customer-facing or consequential use. When those conditions are unresolved, use public or genuinely synthetic examples instead.
Write a short policy that tells employees which workspace to use, what information is permitted, who approves exceptions, and what to do after a mistake. Check retention, access, connected services, and required agreements separately from model training. Patient information, confidential legal matters, payment credentials, and other restricted records need qualified review of the actual workflow. Neither a personal account nor a paid work account settles that review by itself.
What counts as customer data in an ordinary task?
The input is often more revealing than the task sounds. “Help me write a scheduling reply” might include a person's name, home address, health details, and a forwarded thread. “Summarize this complaint” might include payment information and an employee dispute. The assistant does not need all of that merely because it arrived in one email.
In an illustrative HVAC office, a coordinator can practice rewriting a general arrival-window message with invented names and no real address. In a dental office, an appointment note connected to an identifiable patient can raise different obligations. The business type matters, but so do the specific information, purpose, relationships, and applicable rules.
Ask what the tool must know to complete the proposed task. Keep extra details in the approved system of record. Do not remove a name and assume the remaining material is anonymous: an unusual event, precise location, dates, or combinations of details may still identify a person. If the task cannot work with less information, that is a reason to examine the workflow more carefully.
How can a traffic-light rule guide the team?
Use this table as a starting policy classification, then have the appropriate owner adapt it to your contracts and obligations. Green means permitted under your approved policy, not risk-free. Yellow means a named person must authorize the use. Red means stop the ordinary chat workflow. A specialist must confirm that the proposed use is permitted and approve the appropriate system and controls.
| Category | Illustrative information | Working rule |
|---|---|---|
| Green: approved general material | Your published service description, public opening hours, an invented practice example | Use the approved workspace and check the answer before use |
| Yellow: internal or identifiable information | A customer correspondence excerpt, a staff procedure, a nonpublic agency brief | Confirm purpose, minimum fields, permissions, terms, and retention before uploading |
| Red: restricted or consequential records | Patient records, confidential legal matter files, credentials, identity documents, payment authentication data | Keep out of ordinary chat; escalate to the responsible specialist and approved system |
A real customer's text does not become green because it seems harmless to the person pasting it. Customer agreements may restrict disclosure; another employee may not have permission to see the record; an attachment may contain information the task does not need. A published service menu may be used in a personal assistant if your policy explicitly permits that account and purpose. Your policy should identify permitted uses clearly enough to prevent both careless sharing and unnecessary confusion.
What does turning off model training actually settle?
It addresses whether eligible content is used to improve a model. It does not establish that the content is never stored, never accessed for support or safety, never shared with a connected service, or approved under your customer agreement.
OpenAI documents a training opt-out for personal ChatGPT accounts and explains that saved chats remain after training is turned off. It separately documents a no-training default for managed business workspaces. Those are useful distinctions, but neither one replaces your organization's approval of the data and task. ChatGPT data controls.
Check the exact plan and workspace with an administrator. Record who can access submitted information, how long chats and files are retained, how deletion works, and which exceptions apply. Do not infer those answers from an API description when employees are using a chat product, or from one plan's contract when the team purchased another.
Connections deserve their own check. OpenAI's app guidance states that information sent to a third-party service is subject to that provider's storage, processing, privacy, and residency terms. Approval of the assistant therefore does not automatically approve every connected app or action. OpenAI app security and compliance guidance.
Do Claude, Copilot, and Gemini change the rule?
They change the documents and settings you must inspect, not the need to inspect them. Anthropic's commercial policy describes a no-training default and exceptions for certain voluntarily shared data, including feedback. Its retention policy separately covers saved conversations, service-specific retention, and exceptions. Do not treat a consumer Claude account, a commercial workspace, and an API integration as identical. Anthropic training policy, retention policy.
Microsoft's Copilot documentation distinguishes enterprise protections from the terms for web-search queries and agents. Google distinguishes qualifying Workspace core-service use from additional-service use. In each case, check the exact account and feature before applying a broad vendor assurance to the task. Microsoft enterprise data protection, Google account and service distinctions.
For tool selection, use the ChatGPT, Claude, Copilot, and Gemini comparison. For data approval, require a short written record of the actual configuration. “We use a business account” is not enough for an employee to know whether they may upload this particular file.
What belongs in a one-page AI policy?
Adapt the following starting text with the person responsible for privacy, security, and any professional obligations. Replace every bracketed field before issuing it. Keep a separate approval register with product details and contract evidence; employees need a short daily rule, while administrators need the supporting record.
Purpose and owner. We use approved AI tools to assist defined tasks. [Named owner] maintains this policy and the approved-use register. Version: [date]. Questions go to [internal contact].
Approved tools and uses. Use only [product, plan, workspace, and sign-in] for [listed tasks]. Public or synthetic examples may be used as specified in the register. Customer and internal information may be used only within an explicitly approved data category and purpose. Do not move work into another account because a limit or permission stops you.
Information we do not paste into ordinary chat. Do not enter passwords, access keys, payment authentication data, government identifiers, patient records, or confidential legal files. Other private customer or employee information requires the authorization in the register. Share only the fields necessary for the approved task; keep the authoritative record in [approved system].
Connections and sharing. [Administrator] approves connected apps, file access, external actions, and shared links. Do not connect an inbox or drive, publish a chat, or submit confidential conversations as product feedback without approval. Access must match the person's role and be reviewed when responsibilities change.
Review and judgment. [Named role] checks facts, sources, calculations, promises, and sensitive content before customer-facing work is sent or published. Qualified people retain medical, legal, financial, hiring, and other consequential decisions. If the output cannot be verified, stop and use the ordinary process.
Mistakes and changes. Report a suspected disclosure or incorrect external action immediately to [incident contact]. Stop further sharing and follow the incident process; do not conceal the event or assume deleting a chat resolves it. Review the policy at [interval] and whenever the tool, terms, connections, or permitted use changes.
A short policy works only when its bracketed answers are real. If nobody owns access removal, nobody can describe retention, or the incident contact is unavailable, those are operating gaps to close before expanding use. The guardrails pattern in the eight-patterns field note is a useful conversation starter.
What requires special attention in dental, legal, and payment work?
Dental and other healthcare settings. Where HIPAA applies, HHS explains that a cloud provider handling electronic protected health information on a covered entity's behalf generally requires a compliant business associate agreement, alongside the organization's own risk analysis and safeguards. Verify the covered product and features with your compliance lead before use. A BAA does not make every possible workflow appropriate. HHS cloud-computing guidance.
Legal practices. The ABA's Formal Opinion 512 addresses competence, confidentiality, communication, supervision, and other professional duties when using generative AI. Counsel should assess the specific tool, access to inputs, disclosure risks, applicable rules, and whether informed consent is needed in the circumstances. Do not assume a vendor's enterprise label preserves confidentiality or resolves privilege questions. ABA Formal Opinion 512.
Payments. Keep payment credentials in the approved payment process. PCI SSC says card verification codes must not be stored after authorization, even encrypted. Copying payment details into a retained chat creates another place to manage them and can violate restrictions that a training opt-out does not address. PCI SSC guidance on card verification codes.
These examples identify questions for the qualified person responsible for your business. The policy is a working starting point, not a substitute for the legal, security, or professional assessment your circumstances require.
How do you turn the policy into a working habit?
Use a short practice session with public or synthetic examples. Ask each person to classify an input, identify what can be removed, find the approved workspace, and explain who reviews the result. Include an example they should refuse to upload and a mistake they should report. Check what they do, not whether they can repeat the policy heading.
Thirty minutes can introduce those habits; it does not establish compliance or prove that every employee can handle every exception. Repeat practice with the actual task boundaries and update the instructions when staff find unclear cases. Buying a subscription alone leaves those decisions unresolved.
The AI Workflow Sprint can help one team practice approved tasks and verification. Its $1,500 fixed scope covers up to five people, a short leadership kickoff, two working sessions, two to three practical workflows, reusable instructions, a prioritized opportunity map, and one follow-up session. Session lengths and scheduling are agreed before booking. Software, integrations, and subscriptions are separate; it is not a legal or security certification and has no automatic build credit.
For a defined system, implementation can be scoped directly; deeper paid discovery remains optional. Let’s talk about your business and the work you want to improve. Describe the task in general terms and keep confidential records out of the initial intake.
FAQ
Can employees use a personal ChatGPT account for work?
Your policy may permit public or synthetic material in a personal account, but customer information needs explicit approval for the exact service and use. A paid personal plan is still different from an approved organizational workflow. Employees should know which account to use, what data is permitted, and who decides an exception before they copy a real customer message.
Does turning off training make customer data safe?
It resolves only one part of the data-use decision. You still need to examine storage, retention, access, connected services, contracts, and the purpose of the processing. OpenAI also explains that disabling training does not delete saved chats. Approve the specific workflow with the responsible person rather than treating a setting as a general confidentiality or compliance guarantee. ChatGPT data controls.
Is removing the customer's name enough?
Not necessarily. Addresses, dates, unusual circumstances, account details, and combinations of information may still identify someone. Remove information the task does not need, but do not label a record anonymous based on a missing name alone. For restricted information, use the de-identification or data-minimization process approved by your privacy or compliance lead, or practice with genuinely synthetic examples.
Do we need a BAA before using AI at a dental office?
Where HIPAA applies and the service handles electronic protected health information on your behalf, HHS guidance generally requires an appropriate BAA along with your own risk analysis and safeguards. Confirm that the agreement covers the exact product and feature. Public, nonpatient practice examples are a different use case. Your compliance lead should approve the intended workflow before patient information is submitted. HHS guidance.
What should someone do after pasting the wrong information?
Stop further sharing and promptly contact the named incident owner. Record what was submitted, which account and feature were used, and when it happened without spreading the sensitive content into more places. Follow the organization's preservation, deletion, and notification process with the appropriate specialists. Do not assume deleting the visible chat erases every retained copy or ends the assessment.
Put it to work
Install the steps, prompts, human approval rules, and scoreboard in AI Use Policy and Security Week.