Atlanta CPA firms can use AI without exposing confidential client data by approving tools for specific tasks, keeping identifiable information out of unapproved prompts, and managing access through firm accounts. Start with synthetic or properly de-identified information, review how each service handles data, and require a qualified person to check output before it informs client work.
What does the reported AI data warning mean for CPA firms?
“John brought the right knowledge to my issue and resolved it in about a reasonable amount of time. I walked away confident the problem was actually fixed.”
CPA Practice Advisor reported on November 2, 2026, that pasting client returns, bank statements, or Social Security numbers into a free AI tool is a risky shortcut. The story presented the example as a deliberately bad tip in its "How to Get Breached" series. The practical question for a firm is what happens to information after someone enters it into a service.
This concern applies to partners, seasonal staff, administrators, and anyone handling client records. A prompt may contain sensitive information even when it looks like ordinary work instructions. A return, an email thread, or a client-specific question can reveal more than the user intended.
Smarter MSP's November 3 report, "The data loss risk hiding in AI prompts", discusses data loss risks involving prompts. For a CPA firm, AI use belongs in the firm's information-handling practices, not outside them as a convenience feature.
Keep identifiable client information out of an AI tool until the firm approves that tool and use for it.
The first decision is not whether AI is useful. It is whether the tool and workflow are approved for the information involved. That boundary protects confidentiality while leaving room to test lower-risk tasks.
Which AI tasks can a CPA firm start with?
Begin with tasks that do not need client-identifiable details or source documents. For example, staff can ask an approved tool to organize a generic checklist, draft a neutral meeting agenda, or explain a general accounting concept. Do not include client names, account numbers, tax identifiers, financial figures, or combinations of details that could identify someone.
Use synthetic examples when testing a workflow. Replace actual names, dates, amounts, and circumstances with fictional details, then check that copied text and attachments do not retain real information. De-identification involves more than deleting a name: context can identify a person or business, too.
Consider both information sensitivity and the consequence of the task. Drafting a general checklist differs from uploading tax workpapers or asking AI to interpret client records. More sensitive inputs and consequential outputs call for stronger approval, access, and review.
For each proposed use, record the task, the information it needs, the approved tool, the accountable person, and how a human will check the result. If sensitive client data is necessary, wait until the firm has reviewed the tool and workflow. Start with work that has a clear owner and a low-risk data boundary.
Atlanta firms should also consider the written security practices and professional obligations that apply to their work. Georgia's breach-notification law and industry requirements may inform decisions, but they do not make every AI use identical. Ask the firm's legal or compliance adviser to determine which obligations apply.
Choose a first use case that can be tested without client-identifiable information. That gives the firm a practical way to assess value before approving more sensitive work.
How should a firm approve an AI tool?
Evaluate the specific product and account type, not a vendor's general statement about responsible AI. Confirm what information the service collects, how long prompts and outputs are retained, whether they may be used to improve models, how deletion works, and which subprocessors or integrations may receive data. Record the answers and the firm's decision.
Review contract terms, privacy practices, administrative controls, and security documentation for the intended use. Confirm whether the firm can manage accounts centrally, limit access, configure retention, and remove users when their roles change or they leave. If important terms are unclear, do not treat that uncertainty as permission to upload client records.
Assign a business owner and a technical or security reviewer to each approved use. The business owner explains the need and the work supported. The reviewer checks identity, access, data handling, integrations, and logging. Name who will reassess approval if the product or its terms change.
GDS Technology's IT compliance services can support organizations reviewing technology controls and compliance-related practices. The firm remains responsible for its professional obligations and decisions about client information.
Make approval specific: record the tool, account configuration, permitted data classes, allowed tasks, user group, and review date. A broad statement that "AI is approved" does not give staff a useful boundary when they are working to a filing deadline.
Approval should name the task and the data it may use. A written boundary gives staff something concrete to follow and the firm something concrete to review.
What controls keep client information out of prompts?
Make the approved path easy to find and follow. Publish a short policy that names approved tools, prohibited information, permitted tasks, and the person to contact with questions. Cover browser tools, desktop applications, extensions, meeting assistants, and AI features built into software the firm already uses.
- Use firm-managed accounts. Separate professional use from personal accounts where possible. Apply centralized identity and access controls, and grant access according to work responsibilities.
- Set clear data boundaries. Do not upload returns, workpapers, client correspondence, credentials, or other confidential records unless the firm has approved that service and purpose.
- Review access and integrations. Remove access when staff leave or change roles. Check who can use the tool and whether integrations or shared workspaces expose prompts and outputs more broadly than intended.
- Train with realistic examples. Show how amounts, dates, industry details, or quoted text can identify a client even without a name. Give staff a safe alternative and a clear escalation route.
- Check outputs before relying on them. Treat AI material as a draft. A qualified person should verify calculations, citations, client-specific conclusions, and content used in a filing or sent to a client.
Set a response process for accidental disclosure. Staff should know whom to notify, what details to preserve, and how to prevent further sharing. The firm's incident and legal processes, not an employee acting alone, should guide decisions about containment and notification.
Accountability spans the person entering information, the service receiving it, the firm's access controls, and the reviewer relying on the output. GDS Technology approaches technology as an environment to build, secure, and govern over time. For CPA firms, that means connecting AI policy to the systems and routines staff actually use.
A professional-services client described the value of clear technical support this way: "John brought the right knowledge to my issue and resolved it in about a reasonable amount of time. I walked away confident the problem was actually fixed." The testimonial is not about AI, but it reflects the value of giving operational questions a clear owner instead of leaving staff to invent workarounds.
GDS Technology's podcast episode "Closing the Building Data Custody Gap" discusses accountability and aligning access with roles. Those same questions help firms determine who can access AI prompts, outputs, and connected files.
Controls need to work during busy season, not only when someone reads the policy. Test the process with a realistic prompt and confirm staff know what to do when a proposed use is not approved. A policy people can follow is a practical control.
How can a CPA firm introduce AI without disrupting busy-season work?
Avoid introducing a major workflow change when the firm has little capacity to absorb it. Choose a limited pilot, assign an owner, and define what success means before inviting staff. Begin with non-client information, then assess whether the tool saves time without adding avoidable review or data-handling work.
Bring together the people doing the task, the partner or administrator accountable for it, and whoever oversees technology and security. Map where information starts, what the AI tool receives, where the output goes, and who checks it. Include email, document storage, client portals, and connected applications when they are part of the workflow.
Before expanding the pilot, check whether staff followed the data boundary, reviewers could validate the output, and information was retained or shared as expected. Record exceptions and adjust the process. A demonstration may not reflect the demands of real deadlines and multiple client matters.
Keep an inventory of approved tools and use cases. Revisit it when a vendor changes its terms, the firm enables an integration, or staff propose a more sensitive task. Retire approvals that lack an owner or a clear purpose. Make the review part of technology governance, not one employee's informal memory.
For firms with multiple offices or hybrid teams, make the same policy available to everyone. The aim is not to prohibit useful tools by default; it is to prevent informal workarounds from becoming the firm's data-handling standard.
GDS Technology provides technology services for CPA firms and IT support in Atlanta. A useful first conversation starts with the work: Which AI task are staff already attempting, what information does it touch, and who owns the approval decision?
Start with one workflow, make its data boundary explicit, and test the review process before expanding. The firm can then decide from experience whether the time saved justifies the oversight and controls required.
Begin with one low-risk task, a named owner, and a review date. Expand only when the firm can explain what information the tool receives and how a person will verify its output.
Frequently Asked Questions
Can CPA staff paste client data into a free AI chatbot?
Not unless the firm has reviewed and approved that specific tool, account, and use for the information involved. Free access does not establish how prompts are retained, used, or shared. Until those details and firm controls are confirmed, staff should use synthetic examples or an approved alternative.
What should a CPA firm do if an employee already entered client information into AI?
Ask the employee to notify the firm's designated security or incident contact promptly. Preserve relevant details, including the tool, account, information entered, and approximate time, without sharing the data further. Follow the firm's incident response and legal review process to assess containment, notification, and documentation.
Does removing a client's name make an AI prompt safe?
Not necessarily. Dates, amounts, industry details, unusual transactions, quoted messages, and combinations of facts can identify a client without a name. Use synthetic information for testing, and have the firm assess whether a proposed prompt is genuinely de-identified before using it in an AI service.
What should a CPA firm check before expanding an AI pilot?
Confirm that staff followed the approved data boundary, the tool handled information as expected, and a qualified reviewer could verify the output. Record any exceptions, assign an owner to address them, and revisit the tool's access, retention, and integrations. Expand only when the firm can explain and support the workflow.
Follow GDS Technology on Facebook, and subscribe on YouTube for new videos.
