Skip to content
Back

Using Generative AI Without Exposing Client Data

Evaluate AI tools before entering accounting records, and design safe prompts and review procedures that protect client confidentiality.

Xian Hui

Xian Hui

9 November 2026 · 4 min read

An accountant keeps a client folder closed and slides only an anonymised sheet towards an AI chat on a laptop

In brief

How can I use generative AI for accounting work while protecting client data?

Use an approved tool and submit only the data needed for the task. Check provider terms, retention settings and access arrangements before entering client information, and consider abstract questions, synthetic examples or aggregated totals. Verify outputs against original records and promptly report any information entered into the wrong tool.

What you'll learn

  • Recognise sensitive data in prompts and attachments
  • Check provider terms and settings before use
  • Review AI outputs before using them in client work

Using Generative AI Without Exposing Client Data

A team member wants help describing changes in operating expenses. They consider pasting a client's full general ledger into a public AI chat because it would save time. The ledger contains employee names, bank references and supplier information that the tool may not need to produce a useful explanation.

Artificial intelligence can assist with drafting and analysis, but convenience is not authorisation to disclose information. A safe workflow starts by deciding which data is necessary and whether the chosen tool is approved for it.

What might be sensitive in an AI prompt?

Prompts may contain more than visible names. A ledger can reveal customer relationships, pricing, staff salaries, payment details and commercially sensitive transactions. An uploaded file may also contain hidden worksheets, comments or metadata. Removing the company name alone may not prevent re-identification in a narrow or unusual context.

Before submitting information, ask whether an abstract question, synthetic example or aggregated totals would be sufficient. For a discussion of expense trends, a summary table with non-identifying categories may be more appropriate than transaction-level personal data.

Use the same checks when classifying and sharing client records.

Why must the firm review the AI provider?

AI services vary in their contractual terms, data-retention settings, administrative controls, access arrangements and supported safeguards. A consumer service may not have the same approved business terms as an enterprise environment. The firm must establish what has actually been agreed, not assume that all services treat submitted data identically.

For the wider provider review, see Reviewing Cloud Applications, Integrations and Vendor Access.

The NIST Generative AI Profile is a voluntary resource for identifying and managing generative AI risks. Singapore's PDPC guidance library is relevant where personal data is processed. These resources do not mean that every AI prompt is prohibited. They support assessing purpose, exposure and controls.

What is an appropriate approval process?

A small practice can approve AI uses by risk level. For example:

Proposed useInitial decisionReason
Rewrite a generic accounting explanation with no client dataOften suitable for an approved toolLimited confidentiality exposure
Summarise an anonymised fictional caseReview the prompt for identifying context'Anonymised' needs practical checking
Upload identifiable payroll recordsRequire explicit business, legal and security approvalPersonal and confidential data may be disclosed
Permit an AI agent to edit bank or client master dataRequire a separate controlled design and human approvalThe tool could change financial records

This is an example policy table. The firm's actual approvals must reflect service terms, client obligations and applicable law.

Why is output review still necessary?

A model can produce convincing but inaccurate references, totals or explanations. It may omit key facts or misinterpret an instruction. AI output should be checked against original accounting records before it becomes part of client advice or financial statements.

Consider an original exercise: a model summarises a list of expense movements and says every increase is recurring. The accountant should verify the transactions and investigate unusual items rather than repeat the model's confident statement. The professional remains responsible for the conclusion.

What should staff do if client data was entered into the wrong tool?

Report the event promptly. Record what information was submitted, which service and account were used, and whether a file was uploaded or merely referenced. Do not assume that deleting the visible chat removes all copies held by the provider. The incident lead should review the applicable service terms and determine what containment or notification may be required.

Key takeaways

  • Use the minimum data needed to complete the task.
  • Check service approval, provider terms and access settings before submitting client material.
  • Treat AI-generated text and numbers as working drafts that require verification.
  • Report accidental AI disclosures through the firm's established incident process.

Last source check: 9 November 2026.

Next in the course: AI Agent Security: The Risk of an Unsecured Entry Point

Sources

Frequently asked questions

This information has been prepared for general informational purposes only and is not intended to be relied upon as accounting, tax, or other professional advice.

Related articles