Shadow AI in Accounting Practices: Risks and Fixes

Shadow AI is staff using AI for work outside your practice’s approved rules. Find it by asking staff, reviewing browser extensions and checking available sign-in records.

What Shadow AI is

Shadow AI includes AI tools used without approval, even when staff use them to help with an ordinary client task. The practice has not assessed that use or set its limits.

For example, a staff member might paste a client email into a personal ChatGPT account to draft a reply. Another might install an AI browser extension that can read pages containing client details.

An AI note taker can also become shadow AI when someone connects it to client meetings without approval. The recording and transcript need the same attention as a document uploaded to a chatbot.

An approved product can still be used outside its approval. Permission to draft generic website copy does not authorise uploading a client’s financial statements to the same account.

The National AI Centre’s shadow AI guidance links unapproved use to unmet needs and time pressure. Ask what staff are trying to finish before deciding which uses to allow.

The Risks for Client Data

Shadow AI can send client information to a provider before your practice has considered confidentiality or retention. The risk starts when staff share the information, even if nobody sends the AI’s answer to a client.

The Tax Practitioners Board’s AI guidance explains that entering client information into an AI tool can disclose it to a third party. This depends on the tool’s configuration and use.

Code item 6 requires client permission for third-party disclosure unless a legal duty applies. The TPB recommends explaining the proposed disclosure, including who receives the information and where it will be stored.

The Office of the Australian Information Commissioner (OAIC) recommends keeping personal information out of publicly available generative AI tools. Its advice gives particular attention to sensitive information.

Removing a name alone may leave someone identifiable through the remaining details. A client’s business, transactions and family circumstances can reveal whose affairs the prompt describes.

Personal accounts also make retention harder to manage. Establish whether the practice can retrieve or delete uploaded files, end access when someone leaves and obtain a record of what was shared.

A training opt-out answers only the training question. Assess storage, provider access and onward sharing separately. A paid account also needs assessment for the intended task and data.

An outright ban can leave the reason for using AI unresolved. If staff still face the same work pressure, a rule without a usable alternative can encourage hidden use. Pair restrictions with an approved route.

Finding Shadow AI in Your Practice

Combine a staff survey with checks of practice-managed devices and account records. Each reveals a different part of the picture, and none proves that you have found every use.

Ask About Tasks and Accounts

Invite staff to report existing use and explain how you will handle the answers. Ask them to describe the task without copying client information into the survey.

Use these questions for each tool:

• What tool or AI feature do you use for work?
• Which tasks do you use it for?
• Do you sign in through a personal or practice account?
• What kinds of information have you entered or connected?
• Does it access email, files, browser pages or meetings?
• What approved option would let you complete the same task?

Include occasional use and features switched on inside existing software. Staff may recognise a writing assistant without thinking of it as an AI tool.

Review Browser Extensions

Ask your IT provider to inventory extensions on practice-managed browsers. Review each extension’s permissions and the websites where it can operate. Include writing assistants such as Grammarly in that review.

Compare the installed extensions with your approved list. Disable an unapproved extension’s access to client work while you assess it. Preserve any evidence needed to investigate an earlier disclosure.

Check Sign-In Records

Review available sign-in logs and connected applications for practice accounts. Look for unfamiliar AI services or permissions to read files and email. Investigate the purpose with the staff member who granted access.

Sign-in logs show account activity, not necessarily the prompts or documents sent. Personal accounts and unmanaged devices can fall outside those records. Reconcile the logs with staff answers instead of treating an empty report as proof.

If client information has already been shared, activate your incident response procedure. Record what went where and restrict further sharing. Follow the OAIC’s breach response steps to assess harm and notification obligations.

Managing Shadow AI

Approve specific uses and account settings, then give staff a written rule they can apply during client work. Assign a practice owner to keep the approved list current.

1. Approve an option for a defined task. Assess the data it receives, retention terms, access and deletion arrangements. Start with work that uses public or fictional information where that meets the need.
2. Write the rule. Name approved accounts and permitted tasks. State what information staff must keep out, who reviews outputs and how staff request a new use. Use an AI policy template to record those decisions.
3. Train staff with practice examples. Show how an email, attached document or meeting transcript reaches a provider. Include what to do after an accidental upload and who receives the report.
4. Review quarterly. Repeat the staff check and reconcile installed extensions with the approved list. Review changed provider terms, new AI features and any incidents. Reassess sooner when a tool’s access or purpose changes.

Treat quarterly review as a practice schedule, not a legal interval. A new connection to client email needs assessment when proposed, even if the next scheduled review is weeks away.

For a hypothetical practice with three personal AI accounts, first stop client uploads and investigate what was shared. Then give each staff member an approved way to complete their task, with a named owner for unresolved uses.

At the next review, sample an actual task using fictional data. Confirm that staff select the approved account, follow its data limits and know how to report a mistake.