InsightsProfessional Finance Insights › Generative AI in Audit: What It Can Draft, What Still Requires Evidence

Generative AI in Audit: What It Can Draft, What Still Requires Evidence

By CA Nikhil Gupta · 21 July 2026

Generative AI can create a polished audit memo in seconds. It cannot make unsupported information become audit evidence. The distinction matters because audit quality depends on the relevance and reliability of evidence, professional scepticism and documented judgement—not on how confidently a paragraph is written.

Finin2min Summary

The most productive audit uses are often unglamorous: mapping trial-balance accounts, extracting contract clauses, drafting request lists, comparing policy text and preparing first-pass narratives from verified analytics. These tasks can reduce mechanical effort. The auditor must still design procedures, evaluate contradictory evidence, assess management bias and conclude whether evidence is sufficient and appropriate.

A draft is not a workpaper until it is supported

An AI-generated description of a control or transaction should be linked to the process walkthrough, document, system report or interview from which it was derived. Reviewers must be able to reproduce the reasoning. Unsupported statements should be removed, not retained because they sound plausible. Where the model summarises a large population, the team should validate extraction accuracy on a representative and risk-sensitive sample.

Protect client confidentiality and privilege

Uploading ledgers, contracts or personal data into an unapproved public tool can breach engagement terms and data-protection obligations. Firms should approve specific models and deployment environments, control user access, define retention and training settings, and restrict export. Prompts and outputs can themselves contain confidential information and therefore belong within the information-security perimeter.

Preserve professional scepticism

Models optimise for likely responses, not for challenging management. They may smooth contradictions, overstate certainty or anchor on the wording supplied by the user. Audit prompts should explicitly request inconsistencies, missing evidence and alternative explanations. Even then, the auditor must decide whether contradictory evidence has been resolved and whether additional procedures are necessary.

Document the role of AI in the procedure

The workpaper should state the task assigned, source data, model or approved tool, parameters or prompt where relevant, validation performed, exceptions identified, human reviewer and final conclusion. This makes the process auditable and helps the firm distinguish a productivity aid from an automated audit procedure that may require more formal validation.

What the Viral Version Usually Misses

A social post may say AI will audit an entire company. That confuses information processing with assurance. AI can scan more documents and identify unusual patterns, but it cannot obtain management representations, exercise legal authority, observe all relevant conditions or assume the auditor's responsibility. It may expand coverage while simultaneously creating new risks in confidentiality, completeness and explainability.

Worked Scenario: AI-assisted lease review

An audit team uses an approved model to extract lease term, renewal options, escalation clauses and discount-rate references from 240 contracts. A validation sample finds 98% field accuracy but only 87% accuracy for clauses containing cross-references. The team therefore uses AI extraction for straightforward fields, requires human review of cross-referenced clauses and reconciles the contract population to the lease register. The model's output accelerates the procedure; the contracts, reconciliation and reviewer conclusions remain the evidence.

Practical Decision Checklist

Article-Specific Q&A

Can an AI-generated memo be included in the audit file?

Yes, after it is reviewed, corrected and linked to the underlying evidence. The file should make clear what the tool did and what the auditor verified.

Can AI select samples?

It can assist, but the population, risk objective and selection logic must be validated. A model-generated sample is not automatically representative or compliant with the designed procedure.

Should the client be told that AI was used?

Follow engagement terms, firm policy, confidentiality commitments and applicable regulation. Material use affecting data processing or service delivery may require transparency or consent.

Can AI provide accounting-standard citations?

It can help locate candidate guidance, but citations and quoted requirements must be checked against the current authoritative standard. Fabricated or outdated references are a known failure mode.

Does higher extraction accuracy remove the need for review?

No. Error concentration matters. A 99% average may hide systematic failure on the exact clauses or transactions that carry the highest audit risk.

Who signs off an AI-assisted audit conclusion?

The qualified auditor. Responsibility cannot be delegated to the model, software vendor or internal technology team.

Sources and Verification Trail

Editorial note: This article is for education and general awareness. Verify the latest primary source and obtain professional advice before acting.