Audit firms handle some of the most sensitive information inside a business. Working papers can contain payroll data, customer information, bank details, contracts, forecasts, internal control documentation, and evidence about potential misstatements or fraud. When artificial intelligence is added to the workflow, that information can move into new systems, accounts, and vendor environments.
Audit data privacy is therefore not just a question of whether an AI tool is secure. It is a question of what data the tool receives, why it needs that data, where the information goes, who can access it, and how long it remains there.
Map the data before approving the AI workflow
Start by documenting the information that will move through the AI-enabled process.
For a document summarization tool, that may include uploaded contracts, invoices, or meeting notes. For an AI assistant connected to a document repository, the potential data set may be much larger because the system can retrieve files automatically.
The NIST Privacy Framework recommends inventorying and mapping data processing activities, including the systems involved, data elements processed, purpose of processing, third parties, and processing environment.
For audit teams, a simple data-flow map can answer practical questions: What leaves the firm's environment? Does the AI provider receive the full document or only an excerpt? Is another model provider involved? Are prompts or files stored elsewhere?
Privacy controls become easier to design once those paths are visible.
Minimize client data before it reaches the model
An AI tool should receive only the information required for the task.
If the team wants help categorizing transaction descriptions, it may not need customer names or account numbers. If a tool is being tested for document summarization, a redacted or synthetic sample may be sufficient. If only one section of a contract matters, uploading the entire agreement may create unnecessary exposure.
This also means looking beyond obvious personal information. Confidential audit data can include unreleased financial results, acquisition plans, internal control deficiencies, pricing information, litigation details, and management discussions.
Treat confidentiality as a full-lifecycle obligation
Privacy does not end once information has been entered into an approved system.
The International Ethics Standards Board for Accountants states that the fundamental principle of confidentiality applies regardless of the technology used. Its technology guidance extends the focus across the data lifecycle, including collection, use, transfer, storage, dissemination, and lawful destruction.
Ask how long prompts and uploaded files are retained, whether they are used for model training or improvement, which employees or subprocessors can access them, where they are stored, and what happens when an account is deleted.
If client data may be used for another purpose, such as training an AI model, the firm should not assume the original engagement use automatically authorizes that secondary use. IESBA's technology materials specifically address the use of client data for AI training and the need for appropriate authorization.
Review integrations separately from the core AI product
An approved AI application can become much higher risk when connected to other systems.
A standalone chatbot might see only what an auditor manually enters. The same product connected to email, cloud storage, or a working paper repository could potentially reach thousands of documents.
Review each integration based on the information it can access. Use least-privilege permissions so the AI receives only the repositories or data required for its approved purpose. A vendor approval should not automatically approve every connector the vendor releases later.
Make vendor privacy questions specific
Generic questions such as "Is your platform secure?" rarely produce useful answers.
Ask whether prompts, files, and outputs are retained, where they are processed, whether customer information is used for model training, which subprocessors participate, whether retention can be configured, and what incident notifications are provided.
PCAOB staff have reported that audit firms recognize the need for strong supervision of generative AI use to address risks including data privacy and security.
Contractual commitments should match the technical and privacy claims made during procurement.
Protect the new records AI creates
AI systems can create sensitive records in addition to the original client documents.
Prompts may reveal the questions an engagement team is investigating. Outputs may contain summaries of control deficiencies or financial information. Monitoring logs may record sensitive data that an employee attempted to submit.
Those records need appropriate access controls and retention periods. A privacy program can undermine itself if a monitoring system copies confidential prompts into a broadly accessible log that is kept indefinitely.
NIST's Privacy Framework treats privacy risk across the entire data-processing lifecycle, from collection through disposal, rather than focusing only on the original source data.
Plan for mistakes before they happen
Even with good controls, someone may upload the wrong file, use an unapproved account, or discover that an integration exposed more information than expected.
The firm should have a clear escalation process. Employees need to know whom to contact, what information to preserve, and which teams need to assess the event. Depending on the circumstances, that may involve audit quality, information security, privacy, legal, risk, or client leadership.
Early reporting is easier when employees know that a mistake should be escalated rather than hidden.
Reassess as AI use expands
An AI workflow that begins with public information may later be used with live engagement data. A vendor may add memory, new integrations, agents, or different subprocessors. Those changes can alter the original privacy assessment.
IAASB's recent work on technology and quality management emphasizes governance and risk management as emerging technologies become more integrated into audit and assurance engagements. Its 2025 roundtables included more than 240 stakeholders and focused specifically on maintaining quality as AI-enabled tools become more common.
Audit data privacy should therefore be treated as an ongoing control rather than a one-time procurement exercise.
A practical privacy review can be summarized with five questions: What data enters the AI system? Does the system need all of it? Where does the data go? Who can access it? When is it deleted?
If an audit firm can answer those questions for every meaningful AI workflow, it has a much stronger foundation for protecting sensitive information while still benefiting from the technology.
Each of the five questions above assumes the data can still be controlled, but once a prompt is submitted it cannot be recalled. A payroll extract pasted into a summarization request, a full working paper uploaded when one section was enough, or a bank detail dropped into a categorization prompt has already left the engagement team's hands. This is the principle Sanitized Ai is built on, catching and redacting client identifiers and other sensitive engagement information before the prompt reaches the tool, so the minimization decision your team made is the one that actually holds.
This quarter, choose the one AI workflow your audit team uses most and answer the five questions for it in writing, starting with what data enters the system. Mark the identifiers and figures that can be removed without changing the result, and make that removal the default step before submission. If you want to see how redaction can happen before a prompt is submitted, request a demo and walk through your own workflow with us.