AI reporting tools face scrutiny over auditability risks

Financial reporting teams adopting generative AI may face rising control and liability risks if tools cannot trace disclosures back to standards, transactions and management decisions, according to Vallikat Peethamber, founder of VectorPeak Technologies. In a Finextra opinion post, Peethamber argued that productivity gains from large language models are arriving at the same time as closer regulatory and audit scrutiny.

Peethamber said the central weakness is not the fluency of model output, but its lack of evidential structure. Financial statements require disclosures and accounting policies to be traceable, consistent across periods and entities, auditable by third parties and defensible before auditors, regulators or courts, he wrote.

Large language models, by contrast, generate text by predicting likely word sequences from prompts and training data, according to Peethamber. He said that design can produce polished responses without retrieving a defined source, checking historical consistency or preserving the reasoning behind a conclusion.

Hallucination and control concerns

Peethamber cited research from 2024 and 2025 that found leading language models hallucinated in up to 41% of finance-related queries. He said the risk for reporting teams is that such errors can appear coherent and technically plausible, making them harder to detect during routine review.

He pointed to a documented example in which a leading model incorrectly described ISA 620, an auditing standard concerning the use of experts, as relating to fair value measurement. The answer was wrong but presented in a manner that could appear credible to a non-specialist reviewer, he said.

In a reporting cycle, Peethamber wrote, a misstated reference to an IFRS paragraph or an inconsistent description of a liability may remain embedded in a lengthy annual report until audit work takes place months later. By that stage, accounts may already have been published and related regulatory or investor communications issued.

The Public Company Accounting Oversight Board identified AI use in audit as a focus in its 2025 inspection priorities, according to Peethamber. He also said the US Securities and Exchange Commission has reiterated that companies remain responsible for the accuracy of their financial statements regardless of the tools used to prepare them.

Regulators seek traceability

Peethamber said the problem extends beyond false answers to the absence of a causal record. In his example, a contract can drive revenue recognition judgements, performance obligation assessments, transaction price allocation, timing of recognition, balance sheet presentation and related disclosures. Each step needs to connect to the next.

A language model does not inherently preserve that chain, he wrote. If an auditor asks why revenue was recognised in one quarter rather than another, a generated explanation does not provide the same assurance as a retained record of the transaction, assessment and approval path.

Peethamber argued that financial reporting systems should be built around retrieval from verified knowledge bases, causal encoding of decisions and calibrated statements of what the system does and does not know. Under that model, a tool would cite the relevant standard or paragraph, show the path from event to disclosure and decline to answer when a query falls outside its verified domain.

He linked that approach to incoming regulation. The EU AI Act will treat AI systems used in areas including credit assessment, financial compliance and reporting as high-risk from August 2026, with obligations covering technical documentation, logging, human oversight and transparency to deployers, according to Peethamber. He also noted that the Financial Conduct Authority has begun a formal review of AI’s role in financial services.

For finance chiefs and controllers, Peethamber said the practical test is whether an AI tool can identify the exact accounting source behind an output, show the chain from transaction to disclosure, separate supported answers from guesses and produce an audit trail that external auditors can verify.

This story draws on original reporting from Finextra Research.