概述
They apply the same rules on evidence, supervision and documentation to technology-assisted procedures as to manual ones. An audit firm that uses AI is still responsible for three things: the data the tool reads must be reliable, the procedure must fit its purpose, and the work papers must let an experienced auditor understand and re-create what was done. This matters because investors rely on audit opinions, and a fast automated procedure does not make a weak conclusion any stronger.
深入探討
The Public Company Accounting Oversight Board (PCAOB) sets auditing standards for audits of companies registered with the SEC. It regulates the work, not the tool. Several standards govern any procedure, whether a person does it or software does. AS 1105 (Audit Evidence) requires evidence to be sufficient and appropriate, and appropriateness depends on relevance and reliability. AS 2301 sets out how the auditor responds to assessed risks. AS 1201 covers supervision. AS 1215 (Audit Documentation) requires work papers that would let an experienced auditor with no prior link to the engagement understand the work performed, the evidence obtained and the conclusions reached. In 2024 the PCAOB adopted amendments to AS 1105 and AS 2301 that deal directly with technology-assisted analysis of information in electronic form. They take effect for audits of fiscal years beginning on or after December 15, 2025. The amendments make three points. When an analysis serves more than one purpose, such as risk assessment and substantive testing, the auditor must meet the objective of each purpose. The auditor must evaluate the reliability of electronic information, including information the company received from outside sources. And items that a tool flags as meeting the auditor's criteria must be investigated. They cannot be set aside. The PCAOB has also said publicly that its staff is monitoring how firms use generative AI. Its broader quality control standard, QC 1000, treats technological resources as part of a firm's quality control system. Two misconceptions are common. The first is that testing 100% of a population removes the need for judgment. A full-population test built on incomplete or altered data is still unreliable, and the flagged items still need to be evaluated. The second is that AI output counts as audit evidence in its own right. The evidence is the underlying information plus a procedure that was designed, performed and reviewed properly. A summary or score produced by a model is only as good as the checks performed on it.
現實世界的實施
A team runs an analysis over every revenue journal entry for the year and flags entries posted on weekends by users who rarely post to revenue. The flagged entries must then be followed up, because running the analysis does not by itself count as evidence.
An engagement team uses a generative AI tool to pull renewal, termination and pricing terms out of 300 customer contracts. Before relying on the extracted terms, it checks a sample of them against the signed contracts.
Before feeding the company's system-generated aged receivables report into an analytics tool, the auditor tests whether the report is complete and accurate, because it counts as information produced by the company.
A reviewer's work papers record the tool version, the parameters used, the data source, and a reconciliation of the extracted ledger to the trial balance, so another auditor could re-perform the analysis.
The Future of PCAOB Standards and AI in Audits
The 2024 amendments give firms a clearer framework for data analytics, and the first audits under them will show how inspectors read the investigation requirement for flagged items. Generative AI raises questions the current standards answer only indirectly. Examples include how to document a model's role in drafting memos, and how much re-checking of extracted content is enough. The PCAOB has signalled interest through staff outreach, but a firm should not assume new guidance is coming on any particular timeline. For now, the safest approach is to treat every AI-assisted step as a procedure that must be relevant, reliable, supervised and documented under the existing standards.
付諸實踐
為什麼這很重要
- 風險與安全. 災難性和日常的人工智慧危害都取決於誰了解風險以及誰能夠採取行動。
- 更明確的決策. 民眾和專業素養決定強而有力的安全政策在政治上是否可行。
- 突破炒作. 清晰的解釋可以減少炒作、實驗室公關和模糊道德劇場的影響。
留意
- 將存在風險視為科幻小說,同時能力複合。
- 混淆了表面產品安全與高度自治下的對準。
- 只給非英語和非專業觀眾留下低品質的資源。
後續步驟
- 單獨的產品危害、誤用和失控/失調風險。
- 詢問哪些證據會改變您對時間表和嚴重性的看法。
- 比起行銷主張,更喜歡主要來源和具體評估。
- 確定一條行動路徑:職業、政策、資金或技能——而不僅僅是意識。
不斷探索
常見問題
What is PCAOB Standards and AI in Audits?
PCAOB standards neither ban AI nor give it special approval. They apply the same rules on evidence, supervision and documentation to technology-assisted procedures as to manual ones. An audit firm that uses AI is still responsible for three things: the data the tool reads must be reliable, the procedure must fit its purpose, and the work papers must let an experienced auditor understand and re-create what was done. This matters because investors rely on audit opinions, and a fast automated procedure does not make a weak conclusion any stronger.
2024 年關於技術輔助分析的修正案改變了哪兩項 PCAOB 標準?
2024 年修正案更改了 AS 1105(審計證據)和 AS 2301(對評估風險的回應)。它們闡述了審計員如何評估電子資訊以及如何使用技術輔助分析。
所有收入日記帳分錄的分析標示出 40 個符合審計長標準的分錄。修訂後的指導意見有何期望?
修正案明確規定,必須對確定為符合審計師標準的項目進行調查。產生標誌清單並不是一個完整的過程。
在將公司產生的帳齡應收帳款報告放入分析工具之前,審計師應該評估什麼?
公司製作的報告是公司製作的資訊。在審核員依賴它之前,必須對其可靠性(包括完整性和準確性)進行評估。
根據 AS 1215,誰應該能夠理解人工智慧輔助程式的工作底稿?
AS 1215 設定了經驗豐富的審核員測試。文件必須讓這樣的人了解所做的工作、所獲得的證據和得出的結論。
單一技術輔助分析用於風險評估和實質測試。審核員必須做什麼?
修正案規定,當一項分析服務於多種目的時,審計師必須達到每一個目的。實質測試通常比風險評估要求更精確。