Compliance

How to stop employees pasting client data into ChatGPT

Stop client data reaching unapproved AI tools with approved accounts and tested controls. Obsidian Ridge explains how to put an AI policy into practice.

SMB

Obsidian Ridge offers the assessment described below. This guide explains our recommended enforcement checks; it does not claim that a technical control alone satisfies a professional or legal duty.

Short answer: A small law firm, CPA firm or dental office should stop staff pasting client data into unapproved AI tools by approving a work tool and account, restricting personal AI accounts on managed work access, and testing controls that block prohibited pastes and uploads. A written AI policy alone does not enforce those decisions, and support for each restriction must be verified in your setup. Check the devices and connected tools as well, so desktop apps and agents are included in the review.

For the written rules staff should follow, start with our AI usage policy for law firms and dental offices. This guide covers putting those rules into practice.

Key facts

  • How do we stop client-data pastes? Our recommendation, October 23, 2026: pair the written rule with tested controls and a usable approved route; a policy is not a software restriction. Policy and control guidance.
  • Law firms: ABA Formal Opinion 512, issued July 29, 2024, addresses AI competence, confidentiality, supervision and training. ABA opinion.
  • Dental practices: HHS guidance requires an appropriate business associate agreement and safeguards when a cloud provider handles electronic protected health information on a regulated practice's behalf. HHS guidance, checked October 23, 2026.
  • Removing names is insufficient: De-identification must meet the applicable method, not just remove the patient's name. HHS de-identification guidance, checked October 23, 2026.
  • Discovery has a defined scope: The assessment reviews AI apps and agents on staff devices, their access and risky actions. Scope checked October 23, 2026.

What exactly should our firm approve?

Approve a tool, a work account and an allowed task together. Our recommendation is to record the approved combination in a short register that the office manager and IT provider can both use.

For example, approval to rewrite public website copy should not be treated as permission to upload a client file. Use examples that match the practice: a client letter at a law firm, a taxpayer spreadsheet at an accounting firm, or a treatment summary at a dental office. These are illustrative tasks, not accounts of client incidents.

For each approved use, record who owns the decision, what information is allowed, which account must be used and where staff ask for an exception. Review the business terms and data handling before approval. A familiar brand or a paid account is not a substitute for that review.

Free 30-day AI security assessment

Obsidian Ridge's free 30-day AI security assessment finds the AI apps and agents running on staff devices, shows what client data and systems they can reach, and flags risky actions. A CISSP walks you through the findings and recommended fixes. Thirty days, no charge, no card. Implementation and any paid tools are separate.

Built for teams of 10 to 100.

Request my free assessment

How do we enforce the rule on work devices?

Have your IT provider configure the supported restrictions and demonstrate the result with invented data. Ask them to test personal-account access, text paste and file upload separately, rather than reporting that the AI site is simply allowed or blocked.

A managed work browser can block supported paste actions into unapproved AI sites when the relevant controls are licensed and configured. Vendor documentation. That is one possible control, not a claim about every browser, application or device.

Our acceptance checklist is a record of what actually happens:

Test with invented dataWhat the business should record
Open the approved tool with a work accountWhether the allowed task works
Switch to a personal accountWhether the approved account restriction holds
Paste prohibited text or upload a test fileWhether the intended action is blocked
Try an alternate browser or desktop appWhether this is covered or remains a gap
Use a connected extension or agentWhich permissions and actions need a separate decision

Do not put real client information into a test just to see whether a control catches it. A useful result includes allowed paths and uncovered paths. If an alternate route still works, record the gap and decide whether to close it or restrict access to the underlying work information.

What changes for law firms, CPA firms and dental offices?

The allowed data and approval process must reflect the practice's own obligations. For lawyers, ABA Opinion 512 calls for understanding the tool's relevant risks and handling of client information; whether informed consent is required depends on the circumstances. Do not turn that into a blanket statement that every AI task always needs the same consent.

For a HIPAA-regulated dental practice, review whether a provider will handle electronic protected health information, commonly called ePHI. HHS cloud guidance explains the business-associate relationship and agreement requirements. A paste restriction does not create the necessary agreement or replace the practice's risk analysis.

Do not describe a patient record as de-identified just because a name was removed. HHS describes Expert Determination and Safe Harbor, each with its own requirements.

For a CPA firm, our practical recommendation is to identify taxpayer records and confidential financial information in the approved-use register. Keep the technical test separate from the firm's decision about which information can be disclosed. This guide makes no claim that an AI control, by itself, establishes tax-law or professional compliance.

How can staff get their work done without bypassing the rule?

Give them an approved route and a clear exception contact. Our recommended block message says which task is restricted, where the permitted tool is, and who can decide an exception. It should be readable by someone trying to finish a document, not just by an administrator.

Ask employees to demonstrate the work they need to complete with invented data. If the approved tool cannot do the task, decide that openly. Keep a short log of exceptions with an owner and a review date. Use those requests to improve the policy and controls together.

Include connected mail, drive access and installed agents in the discussion. A rule focused only on text pasted into a website leaves the business without an answer about other approved or unapproved ways a tool obtains information.

What if someone has already pasted client information?

Record the tool, account, approximate time and information involved, then use the firm's incident process. Our recommendation is to stop further sharing and tell the person responsible for privacy or security promptly. Do not promise that deleting a conversation reverses a disclosure.

The owner should work with the relevant IT, privacy and professional advisers to decide containment and any notification duties. Keep the response factual and make reporting easy; staff should know exactly where to bring a mistake.

To improve the starting inventory, use the assessment scope below. Its findings can inform enforcement decisions. Putting the fixes in place is separate work.

Free 30-day AI security assessment

Obsidian Ridge's free 30-day AI security assessment finds the AI apps and agents running on staff devices, shows what client data and systems they can reach, and flags risky actions. A CISSP walks you through the findings and recommended fixes. Thirty days, no charge, no card. Implementation and any paid tools are separate.

Built for teams of 10 to 100.

Request my free assessment

Sources

Last updated

October 23, 2026. We refresh this content as the threat landscape and tools evolve.

FAQ

Questions readers usually ask next

How can a small law firm or dental office stop staff pasting client data into ChatGPT?

Approve a work tool, account and allowed data use. Configure supported restrictions on personal accounts and prohibited pastes or uploads, then test them with invented data. Review desktop apps, extensions and agents separately.

Is a written AI policy enough?

No. A policy states the rule. Supported technical controls, an approved working route and checks of actual behavior put it into practice.

Should we block every AI website?

Choose restrictions from the tasks and information your business has approved. Give staff a usable permitted route and test alternative access paths; do not assume a website block covers installed apps or agents.

Does removing patient names make a record safe to paste?

Not by itself. HIPAA de-identification has requirements beyond names. A dental practice should use the applicable HHS method and its own approved process.

Does the free assessment implement the restrictions?

The assessment produces findings and recommended fixes. Implementation and any paid tools are separate.

Full bio & provenanceSee related service

Related reading