AI 并未创造审计风险,真正改变的是什么
AI didn't invent audit risk, so what actually changed?
投资审计长期对经技术处理的信息执行审计,AI 并非审计风险的真正来源;生成式 AI 改变的是部分固有风险的表现形态,审计师需将职业怀疑延伸至 AI 辅助的程序。AI 生成的分析可能外观酷似已完成的专业工作,传统可复算核对方法不再完全适用,审计目标未变但实现手段需调整。若投资管理人、基金行政管理人与外部审计师使用相似 AI 工具,原本独立的各方可能因技术趋同而暴露于同类错误。
Investment auditors were auditing technology-transformed information long before anyone called it artificial intelligence.
Processing Content
Consider a typical investment environment. Custodians provide holdings and transaction data. Portfolio accounting systems process that information. Pricing services supply valuations. Performance systems calculate returns. Firms aggregate portfolios into composites and ultimately present investment performance to clients and prospective investors.
As auditors, we rarely encounter information that has not passed through technology somewhere along the way. And we already know what can go wrong. A security can be mapped incorrectly. A corporate action can be treated improperly. A performance system can apply the wrong methodology. An Excel formula can propagate an error across thousands of records. Auditors can also place too much confidence in a system simply because its output looks authoritative.
None of these risks arrived with generative AI. As AI becomes more embedded in auditing, there is a practical question worth asking: What is actually new?
Old risk, new technology
Consider a simple example. An audit staff member obtains a custodian statement showing an investment balance of $10 million but mistakenly reads the balance as $1 million. Now suppose an AI-enabled extraction tool reads the same statement and makes exactly the same mistake.
From an audit perspective, the immediate problem is the same: The auditor may be performing a procedure using incorrect information.
Scale is not entirely new either. An incorrectly configured accounting system can process thousands of transactions incorrectly. A faulty spreadsheet formula can affect an entire population. A performance engine applying the wrong methodology can systematically miscalculate returns across multiple portfolios. Overreliance on technology is not new to our profession either.
That is why we should be cautious about treating every risk involving AI as a new "AI risk." The more useful question is whether AI changes the nature of a familiar risk enough that our audit response needs to change.
In some cases, it might. Take review, for example. An audit staff member who does not understand a complex investment agreement may leave the analysis incomplete, raise a question or ask a manager for help. A generative AI tool may instead produce a polished, logically structured and apparently complete analysis even when an underlying interpretation is wrong.
That does not mean the audit staff member is necessarily more reliable than the technology. The practical concern is different: An incorrect machine-generated answer can look remarkably like completed professional work.
As reviewers, that matters. If a tool analyzes thousands of transactions and reports "no exceptions identified," we still have to understand what was analyzed, evaluate whether the inputs were appropriate, and determine whether the resulting output supports the audit objective. The appearance of completeness cannot substitute for audit evidence.
Professional skepticism has always applied to information we obtain and evaluate. As we use AI ourselves, it also has to extend to technology assisting us in performing audit procedures.
A familiar lesson from investment performance
Investment performance provides a useful comparison. Performance professionals have long worked with information that moves through a series of systems:
Custodian data → Portfolio accounting → Performance calculation → Composite construction → Presentation
As practitioners, we would never conclude that a return is correct simply because a performance system calculated it. At the same time, we would not conclude that it is unreliable simply because technology was involved.
We ask practical questions. What information went into the calculation? What methodology was applied? Were the relevant policies and procedures followed? Can selected results be recalculated or otherwise tested?
The Global Investment Performance Standards provide one example of this approach. Verification provides assurance, at a firm-wide level, on whether relevant policies and procedures have been designed in compliance with the standards and implemented. It does not provide assurance about the accuracy of a specific performance report.
There is a useful lesson here for AI: Technology touching information is not, by itself, the risk. The question is whether we can understand the information, test it and determine whether it is appropriate for the purpose for which we are using it.
Where AI may test our existing approach
One area worth watching is reperformance. With a traditional calculation, we can generally inspect the formula, understand the inputs and reproduce the result. Generative AI does not always fit as neatly into that model. Its output can depend on the prompt, context, model configuration and model version.
Suppose an auditor uses AI to review a lengthy investment agreement and identify provisions relevant to valuation or classification. The underlying agreement may be reliable audit evidence, but if AI-assisted analysis contributes meaningfully to the audit procedure, familiar questions remain. Can another experienced auditor understand what was done? Can the conclusion be traced back to the underlying agreement? Can the work be adequately reviewed or reperformed?
The audit objectives have not changed. The mechanics of satisfying them may have.
There is another issue worth watching as AI adoption expands. Imagine an investment manager uses AI to extract a contractual term from an investment agreement. A fund administrator independently uses an AI-enabled process to classify the same instrument. The external auditor then uses AI-assisted analysis to evaluate the supporting documentation.
The parties remain independent and perform different functions. But could otherwise independent parties become exposed to similar errors through their use of similar technologies?
Other disciplines refer to related situations as "common-mode" failures: safeguards that appear independent but can fail for the same underlying reason. This is not necessarily a widespread audit problem today. But it raises a practical assurance question: Does independence among the parties necessarily mean independence in how the technologies they use can fail?
That question may become more relevant as AI is introduced at multiple points in the financial reporting and assurance process.
The principles have not changed
Recent PCAOB amendments to AS 1105 and AS 2301 reflect the increasingly technology-intensive environment in which audits are performed. Among other changes, AS 1105.10A addresses evaluating the reliability of external information provided by the company in electronic form, including understanding its source and the company's process for receiving, maintaining and processing it.
The broader lesson is straightforward. Technology changes, but the underlying questions auditors have to answer remain familiar.
Do we have sufficient appropriate audit evidence? Do we understand the information we are relying on? Have we appropriately evaluated its reliability? Can the work be reviewed or reperformed? Have we maintained professional skepticism?
AI does not make those questions obsolete. It may make some of them harder to answer.
Before creating an entirely new category of "AI audit risk," perhaps the profession should first ask a more disciplined question:
Which existing assurance principles already work — and where does generative AI genuinely test their limits?

Audit manager
Ashley Jiang, CPA, CIPM, is an audit manager, asset management, at Baker Tilly US. She serves clients who manage a wide range of asset classes... Read full bio
来源:Accounting Today(美国会计行业媒体) · accountingtoday.com