Accounting firms need a client-data check before using ChatGPT.
Before an accounting firm puts a client file into ChatGPT, a named reviewer should confirm the permitted account, the information that may be shared and the purpose of the task. A colleague’s access to the tool does not settle those questions. Start with a synthetic example while the client-data decision is unresolved.
The account is the doorway into the records room. Before handing over a file, establish who controls that doorway and where the information can go afterwards. Keep that decision with the approval so another person can review it later.
Identify the workspace and the task.
Write down the active workspace, the plan and the person who administers it. If an accounting application uses an API behind the scenes, ask its owner to identify that service and the settings that apply. The name on the button is insufficient evidence of the underlying arrangement.
OpenAI’s ChatGPT Work cloud security documentation describes Business, Enterprise and Edu workspace protections, including exclusion of business data from model training by default. It also explains that available controls depend on the plan and configuration. Those statements concern the covered business workspaces; they should not be carried over to a personal account. See the official Work cloud security documentation.
Use these questions to distinguish the account types without treating a plan label as approval:
- Personal workspace: Has the firm permitted this account for the proposed task? Who can change its data settings, and what restrictions apply to client information?
- Firm-managed workspace: Which workspace is active? Who controls access and settings? Does the approved use cover this particular task and data?
- API-based application: Which organisation and project handle the request? What does the application retain or send elsewhere? Who can inspect those settings?
A satisfactory answer identifies the actual account and its owner. It should not depend on a colleague remembering which login they used last week.
Separate training from retention and onward access.
Training policy, retention and connected-system permissions need separate entries in the review. OpenAI’s Work documentation describes different storage categories for conversations, files and connected-app content. It also distinguishes API-specific retention controls from Work retention. Do not use an API setting as the retention policy for a ChatGPT task.
For the proposed workflow, list where information is supplied and where an output may be saved. Include connected applications and recipients. If the application owner cannot explain a destination, leave that part of the decision unresolved.
A screenshot of a privacy toggle records one setting at one moment. Pair it with the account identity, the applicable terms and the firm’s permitted use. That gives a later reviewer something concrete to check.
Keep one account-and-data record.
Before the first client file, complete a short record:
- Task: The work to be done and why client information is needed.
- Account: The workspace or application, its administrator and the person using it.
- Data: The permitted fields and the information excluded from the task.
- Destinations: Connected applications, saved outputs and intended recipients.
- Evidence: The settings and terms checked, their source and the date of review.
- Decision: The authorised reviewer, conditions, unresolved points and next review.
Store references in that record where possible. Do not create another unnecessary copy of a client file merely to document its handling. The firm’s reviewer should resolve any contractual or professional obligations that affect the decision.
Give staff an instruction they can follow.
An instruction such as “use the firm’s approved workspace to summarise the permitted fields; save the reviewed result in the designated client folder” is more useful than a general permission to use AI. It names the operating arrangement. The written rule should also identify material that must stay outside the task and the person who can answer an exception request.
Review the decision when the account, application, feature or intended use changes. This is an operating checklist, not legal advice or permission to disclose a particular client’s information.
Does turning off training settle the client-data decision?
It addresses one issue. The firm still needs to establish permitted data, retention, access and onward destinations. Record each decision separately.
What should staff do while approval is unresolved?
Use a synthetic example with no client information, or pause the task. A successful upload does not establish that the firm permitted it.
Questions buyers ask.
Can an accounting firm put client data into ChatGPT?
A named reviewer should first confirm the permitted account, the information that may be shared and the purpose of the task. OpenAI documents different protections for Business, Enterprise and Edu workspaces than for personal accounts, so the account type is the first fact to establish.
Do ChatGPT’s business data protections cover a personal account?
OpenAI’s Work documentation describes protections for Business, Enterprise and Edu workspaces, including exclusion of business data from model training by default. Those statements concern the covered workspaces and should not be carried over to a personal account.
For the wider review method, read the CISO evaluation guide for AI workspaces.
Start the review with one task and its actual account. Open the procurement workbench.
Josh Olayemi · Founder, Handvantage · August 2026
