Client data is being copied into AI tools right now.
- A paralegal asks ChatGPT to summarize a draft contract.
- A medical assistant pastes appointment notes into an online writing tool.
- An accountant uploads a spreadsheet to “clean up” formulas.
- A partner says, “It was only one prompt.”
That is shadow AI: employees using artificial intelligence for business work without formal approval, security review, or visibility from leadership.
It is usually well-intentioned. It is also a growing cybersecurity, compliance, confidentiality, and business risk for law firms, medical practices, accounting firms, and other organizations handling sensitive information.
The question is no longer whether your employees are using AI. The question is whether they are using it with your permission, under your controls, and with a clear understanding of what happens to the information they submit.
The “We Didn’t Know” Problem
Shadow AI often begins innocently.
“Hey Margaret, can you put this into ChatGPT and make it easier to understand?”
“I got a guy who says this AI tool is secure.”
“Don’t worry. I removed the client’s name.”
But removing a name does not necessarily remove the risk. A contract may still identify a client through dates, transaction details, addresses, matter numbers, or unique facts. A medical note may remain identifiable through a combination of symptoms, age, location, and appointment information. A tax document may contain enough financial details to expose a person even after one obvious identifier is deleted.
The most common problem is not malicious behavior. It is a lack of awareness.
Your employee may not know:
- Whether the tool retains prompts or uploaded files.
- Whether the provider uses inputs to improve or train its models.
- Who can access stored conversations.
- Whether data is copied into logs, backups, or third-party systems.
- Whether the tool has a breach notification commitment.
- Whether the tool is covered by a business associate agreement or data processing agreement.
- Whether using it violates a client contract, engagement letter, or professional obligation.
If leadership discovers the activity only after an incident, “we didn’t know” will not be a sufficient response for a regulator, client, insurer, bar association, or malpractice carrier.
Why Shadow AI Creates Serious Exposure
Public AI tools are not automatically unsafe. The problem is using a tool without understanding its data practices or placing the wrong information into it.
1. Data retention and model training
AI providers use different settings, contracts, and retention practices. Some tools may retain prompts for abuse monitoring, troubleshooting, analytics, or other purposes. Some consumer plans may allow inputs to be used to improve services unless a specific setting or enterprise agreement changes that treatment.
You need clear answers to three questions before client information enters an AI system:
- Is the input retained?
- Is it used to train or improve a general model?
- Can the vendor contractually and technically restrict access, reuse, and deletion?
If you cannot answer those questions, the tool should not process client, patient, privileged, or financial information.
2. Breach notification and incident response
Once sensitive information leaves your controlled environment, your incident response process becomes more complicated.
You may need to determine:
- What information was submitted.
- When it was submitted.
- Which employee or account submitted it.
- Whether the data was retained.
- Whether the vendor shared it with subprocessors.
- Whether unauthorized access occurred.
- Whether clients, regulators, insurers, or law enforcement must be notified.
For organizations covered by the FTC Safeguards Rule, certain breaches involving the unauthorized acquisition of unencrypted customer information affecting at least 500 consumers must be reported to the FTC as soon as possible and no later than 30 days after discovery. The FTC’s Safeguards Rule guidance also requires covered financial institutions to maintain a written, risk-based information security program and oversee service providers.
That matters to many tax preparation and accounting firms. An AI provider processing customer information may need to be treated as a service provider within your broader GLBA compliance and FTC safeguards program.
3. Ethics, malpractice, and client confidentiality
For law firms, the risk is not limited to a possible cyberattack. It may also involve attorney-client confidentiality, privilege, competence, supervision, and informed consent.
The ABA’s Formal Opinion 512 addresses lawyers’ use of generative AI and applies existing professional duties to these tools. Lawyers must take reasonable steps to prevent unauthorized disclosure of information relating to a client’s representation. Depending on the tool and how it uses inputs, informed client consent may be required before entering client information.
A careless upload can create:
- A potential confidentiality violation.
- Questions about whether privilege was waived.
- A client notification obligation.
- Malpractice exposure.
- Reputational damage.
- Problems with professional liability coverage.
The same principle applies outside legal practice. Medical practices have HIPAA obligations. Accounting firms have client confidentiality duties and may have GLBA obligations. Every professional services firm has contractual commitments to protect the information entrusted to it.
How HIPAA and GLBA Change the Conversation
Medical practices: HIPAA compliance requires visibility
If an AI tool creates, receives, maintains, or transmits electronic protected health information, it belongs in your HIPAA risk analysis.
The U.S. Department of Health and Human Services explains that a cloud provider storing or maintaining ePHI generally qualifies as a business associate, even if the data is encrypted and the provider cannot view it. A covered entity must have an appropriate business associate agreement before sharing PHI with that provider.
Do not assume that an AI vendor is HIPAA-compliant because its website uses the word “secure.” Confirm:
- Whether a BAA is available.
- What data the vendor retains.
- Whether prompts are used for training.
- How access is logged.
- How subcontractors are handled.
- How data and model artifacts are deleted.
- How quickly the vendor reports security incidents.
Accounting firms: customer information does not belong in consumer tools
Tax returns, payroll records, bank statements, financial statements, Social Security numbers, and client contact lists can constitute nonpublic personal information.
If your firm prepares tax returns or performs covered financial activities, the FTC Safeguards Rule may apply. Your written information security program should account for AI platforms just as it accounts for email, file storage, accounting software, and other vendors.
That means documenting the risk, reviewing the provider, limiting access, monitoring use, training employees, and maintaining an incident response plan.
A Practical Shadow AI Playbook
You do not need to ban every AI tool. You need to govern its use.
1. Publish an acceptable use policy
Your policy should use plain language and answer:
- Which AI tools are approved?
- Which tools are prohibited?
- What information may never be entered?
- Are generic, non-confidential prompts permitted?
- Who approves new tools?
- How must employees report a mistake?
- Is human review required before AI-generated work reaches a client or patient?
Use a three-tier model:
- Approved: Enterprise tools reviewed by IT, legal, and leadership.
- Permitted with restrictions: Consumer tools for generic, non-confidential tasks only.
- Prohibited: Unvetted tools, browser extensions, plugins, or applications with unclear retention and data-use practices.
2. Approve enterprise tools with real data protection
If AI will support client work, select a business or enterprise offering with documented controls.
Look for:
- No training on your business inputs by default.
- Tenant or workspace isolation.
- Administrative controls.
- Audit logs.
- Role-based access.
- Retention and deletion settings.
- Encryption.
- Vendor breach notification commitments.
- A DPA, BAA, or comparable contractual protection where appropriate.
The goal is not to accept a vendor’s marketing promise. The goal is to document why the tool is appropriate for your risk profile.
3. Audit what is already connected
Before writing a policy, find out what is happening today.
Ask employees: “List every AI tool you have used for work in the past 90 days, even if you used it only once.”
Then review:
- Microsoft 365 and identity logs.
- OAuth application consent.
- Browser extensions.
- Web filtering reports.
- Credit card and expense records.
- File-sharing integrations.
- API keys and automated workflows.
- AI features recently enabled in existing software.
An IT audit or IT assessment can help identify tools that employees may not remember or may not consider “AI.”
4. Apply DLP, conditional access, and endpoint security
Policy alone cannot stop a rushed employee from pasting a document into the wrong website.
Technical controls can help:
- Use data loss prevention to detect sensitive content.
- Block uploads to unapproved AI domains.
- Restrict access from unmanaged devices.
- Apply conditional access based on user, device, location, and risk.
- Control third-party OAuth applications.
- Monitor browser extensions.
- Enforce endpoint security across laptops and workstations.
- Retain logs for investigation and compliance reporting.
These controls should be proportionate. A law firm may need special detection for matter numbers and privileged documents. A medical practice may focus on PHI identifiers. An accounting firm may prioritize tax IDs, financial account numbers, and tax forms.
5. Train employees without creating a culture of silence
If employees fear punishment, they may hide shadow AI use. If they understand the risk and have approved alternatives, they are more likely to report it.
Training should include realistic examples:
- “Can I paste this contract if I remove the client name?”
- “Can I upload a patient note if I use initials?”
- “Can I ask AI to explain this tax issue using a real client scenario?”
- “Can I install this browser extension if it saves me time?”
Make the answer clear: Do not enter sensitive information into an unapproved tool. If you already did, report it immediately.
Early reporting gives you a chance to investigate, contain, document, and respond.
The Pros and Cons of Managed AI Governance
There are costs to doing this correctly. You may need to pay for enterprise licensing, conduct an IT assessment, revise contracts, configure DLP, and provide recurring employee training.
There are also costs to doing nothing:
- Unpredictable incident response expenses.
- Client notification and legal fees.
- Regulatory investigations.
- Lost trust.
- Malpractice or professional liability claims.
- Disrupted operations.
- Emergency technology projects after an incident.
A managed services provider gives you economies of scale. Instead of asking one employee to be your security officer, compliance manager, vendor assessor, and helpdesk, you gain a team that can help manage identity, endpoint security, email security, cloud controls, policy, monitoring, and response.
A PC of Mind helps regulated organizations build that foundation through cybersecurity services, cloud computing, and IT consultancy.
Final Word
Shadow AI is not a future problem. It is an information governance problem already inside your office.
Your employees are trying to work faster. Give them a safe way to do it.
Create an acceptable use policy. Approve the right enterprise tools. Audit what is already connected. Use DLP and conditional access. Train your people. Document your decisions.
For law firms, medical practices, accounting firms, and other regulated organizations, AI can become a competitive edge, but only when client confidentiality, HIPAA compliance, GLBA compliance, and cybersecurity are designed into the workflow from the beginning.
Start with an IT assessment before a client, regulator, or insurer starts asking questions for you.
