Can employees use public AI tools with sensitive data? The practical answer is: not by default. A tool’s public availability does not tell you whether it is appropriate for customer, patient, or financial information. Before allowing sensitive inputs, a business should assess the specific tool, account, settings, contract, and intended use.
That distinction matters because employees may focus on whether a tool uses prompts to train its model, while overlooking retention, human or vendor access, connected apps, sharing, deletion, and the organization’s own responsibilities. A clear decision process helps teams use AI productively without treating confidential data as ordinary text.
Start with the data, not the tool’s popularity
Classify information by how much harm or disruption could result if it were exposed or misused. A practical starting point is to sort it into these categories:
- Public information: Material already approved for public release, such as published service descriptions.
- Internal information: Work material not intended for public release, such as internal procedures. Check your organization’s rules before sharing it.
- Confidential customer information: Information that identifies a customer or reveals private business, account, or transaction details.
- Regulated or highly sensitive information: Information such as patient data or financial records that may carry specific privacy, security, contractual, or compliance obligations.
Set a simple default: employees should not enter confidential customer information or regulated data into an unapproved AI tool. If a task can be completed with public information, synthetic examples, or details that have been properly removed, use that lower-risk input instead. Removing a name alone may not be enough if other details could identify the person or account.
Privacy and security are practical concerns for businesses adopting AI. In a 2025 survey, 42% of surveyed SMBs named data privacy and security as a drawback to using AI; respondents could select up to three answers.
Check more than model training
Turning off model training can be a useful safeguard, but it does not answer every data-handling question. Before approving a tool, understand what happens to prompts and uploaded files through their full lifecycle.
- Use and training: Does the provider use submitted information to train or improve models? Do the terms differ between account types or features?
- Retention and deletion: How long are prompts and files kept? Can the organization delete them, and what does deletion cover?
- Access: Who at the provider can access submitted information, and under what circumstances?
- Sharing and integrations: Can information flow to connected apps, plugins, or other services? Which users can enable those connections?
- Contract and commitments: Do the terms clearly describe how information is handled and what the provider commits to do?
Do not assume that a setting or a reassuring product description settles all of these questions. Review the terms for the exact product and account your employees will use. The NIST Generative AI Risk Management Framework profile can help organizations structure a review of potential risks and safeguards.
Take extra care with patient and financial information
Healthcare: Do not treat a public AI tool as approved for patient information just because it can summarize notes or draft messages. First determine whether the information is protected health information, what role the vendor would play in handling it, and whether the organization’s required agreements and safeguards are in place. If a vendor handles protected health information on behalf of a covered organization or business associate, ask qualified privacy or compliance professionals whether a business associate agreement is required before using the service for that purpose.
HIPAA questions are not answered simply by asking whether a tool is generative AI. The specific data, use case, vendor relationship, and safeguards matter. Zevonix lists HIPAA among the compliance requirements it supports, but organizations should have their own compliance or legal advisers review proposed uses.
Financial information: Account numbers, payment details, transaction records, and customer financial information should not go into an unapproved tool. Requirements vary with the organization, the information, the service, and the obligations that apply to the business. Confirm those obligations with the appropriate compliance or legal adviser, and verify that the tool’s terms and controls fit the intended use.
For either setting, minimize the information shared. If the task does not require a person’s identity, account details, or full record, do not include them. Use an approved workflow and have an accountable person review AI-generated output before it is used in a customer-facing or operational decision.
Give employees a usable decision rule
An acceptable-use policy should make it easy to pause before sensitive information is submitted. Spell out which tools are approved, which data categories are prohibited, who can approve an exception, and where employees should ask questions. Include examples relevant to everyday work, such as drafting a general email versus pasting a customer’s account history into a prompt.
The policy should also cover how to use approved settings, whether connected apps are allowed, how to check AI-generated output, and how to report a mistake. Train staff to use only the information needed for the task, and name an owner who can review tool requests and keep the approved-tool list current.
For a broader view of the systems and information your business needs to protect, see Zevonix’s data protection guidance and business IT needs assessment.
If sensitive data was submitted by mistake
Tell employees to report an accidental submission promptly through the organization’s established incident process. They should identify the tool and account used, the kind of information submitted, and when it happened. Avoid assuming that deleting a chat removes every retained copy or resolves the exposure.
The responsible team can then review the provider’s deletion controls and terms, assess what information may have been exposed, and decide whether additional steps are needed under the organization’s policies and obligations. If the data may be regulated or contractually protected, involve the appropriate privacy, security, or compliance contact.
Frequently asked questions
Can employees paste customer information into ChatGPT?
Not unless the specific tool and account have been approved for that data and use. If approval is unclear, leave identifying and confidential details out and ask the designated owner.
Does turning off AI model training make sensitive data safe?
No single setting answers every privacy question. Also review retention, deletion, access, sharing, connected apps, and the provider’s terms before allowing sensitive inputs.
Does an AI vendor need to sign a business associate agreement?
That depends on whether the vendor handles protected health information on behalf of a covered organization or business associate and on the specific arrangement. Get qualified compliance guidance before using the tool with patient information.
What should an employee do after submitting sensitive information?
Report it promptly through the company’s incident process. Do not assume that deleting the prompt is enough; the responsible team should assess the exposure and next steps.
Make AI use a deliberate business decision
Employees can benefit from AI tools, but sensitive data should only be used when the specific tool, task, protections, and obligations have been reviewed. A clear approved-tool list and practical employee guidance make that boundary easier to follow.
Zevonix provides managed AI services and serves businesses across Flagler County. If your organization is deciding how to adopt AI while protecting business information, contact Zevonix to discuss your IT needs.