AI in Auditing: How Small CPA Firms Can Use AI Responsibly

Small and mid-size CPA firms face a familiar squeeze. Fewer accounting graduates are entering the profession, experienced seniors are hard to hire and harder to keep, and the audit busy season keeps getting longer as calendar-year audits run into employee benefit plan and single audit deadlines. AI tools are being offered as part of the answer.

This guide covers where AI can help in a nonissuer audit, where it cannot replace professional judgment, and how the firm's existing obligations under the AICPA quality management standards, GAAS and the Code of Professional Conduct apply when AI is used. It paraphrases those standards and is not a substitute for them.

Why firms are looking at AI now

The staffing gap

The AICPA's periodic Trends reports on accounting education and hiring have documented a decline in accounting graduates and CPA exam candidates over recent years. Many states have responded by adding alternative pathways to CPA licensure beyond the traditional 150-hour requirement. Whatever the long-term outcome, most small firms are doing the same work with fewer experienced people.

Busy season pressure

For a typical small firm, the busy season starts with calendar-year audits and reviews in the first quarter, overlaps with tax deadlines, and runs into ERISA plan audits due in summer and fall and single audit reporting packages due no later than nine months after the end of the audit period (2 CFR 200.512). Time pressure can affect audit quality, and removing low-value manual work is one lever a firm controls.

What AI can realistically do

Current tools are good at reading and summarizing documents, extracting structured data from unstructured files, matching and reconciling records, drafting first versions of routine text, and flagging anomalies in large data sets. They are not reliable at reaching audit conclusions, and they can produce confident, wrong output. The responsible approach is to use them where their output can be checked efficiently.

Where AI fits in an audit

The table below sorts common audit tasks by how suitable they are for AI assistance, assuming human review of the output.

Data extraction and tie-out

The best early use cases are the ones that consume staff hours but add little judgment: keying invoice details into a sample workpaper, matching bank statement lines to cash receipts, footing and cross-footing, and agreeing schedules to the trial balance. Each output can be checked against the source document, which makes review practical.

Working paper preparation

AI can draft the structure of lead schedules, supporting schedules and roll-forward workpapers from the trial balance and prior-year file. The preparer still performs and documents the procedures. Under AU-C 230, the file must show the nature, timing and extent of procedures, the results and the evidence obtained, and who performed and reviewed the work. A workpaper generated by a tool is still the preparer's workpaper.

Analytics and journal entry testing

AI-assisted tools can scan entire general ledgers for entries with characteristics associated with fraud or error, such as entries posted at unusual times, by unusual users, to unusual account combinations, or with round amounts. This supports the journal entry testing required by AU-C 240. The auditor still defines the criteria, evaluates the selections, and tests them.

Where analytics are used as substantive procedures, AU-C 520 still applies: the data must be reliable and the expectation precise enough. A tool's anomaly score is not an expectation in the AU-C 520 sense.

Risk assessment and materiality

AI can help gather information for risk assessment under AU-C 315, such as industry developments or changes in the entity's contracts, and can compute materiality under AU-C 320 consistently. The assessment of inherent and control risk and the choice of benchmark are the auditor's judgments. The materiality calculator shows the arithmetic, but the reasoning belongs in the file.

Sampling

AI tools can select samples, but sample design under AU-C 530 is the auditor's responsibility: the population, the sampling unit, tolerable misstatement and the method. If a tool selects items, the file should show how selection worked so the sample can be re-performed.

Supervision and review: AU-C 220 still applies

Under AU-C 220 as revised by SAS No. 146, the engagement partner takes overall responsibility for managing and achieving quality on the audit. That includes the direction and supervision of the team and the review of their work, and determining that the resources assigned to the engagement, including technological resources, are sufficient and appropriate.

Practical implications when AI is used:

• Review the output, not just the summary. A reviewer who signs off on an AI-prepared schedule without checking a sample of items back to source has not reviewed it.
• Be explicit about who did what. The file should show which steps were assisted by a tool and which were performed or verified by a person.
• Train staff to challenge the tool. Professional skepticism under AU-C 200 applies to tool output as much as to management's representations. Generated citations to standards, contract terms or amounts must be verified against the source.
• Watch for automation bias. Junior staff are more likely to accept tool output uncritically. Supervision needs to account for this.

Audit evidence considerations

AU-C 500, as revised by SAS No. 142, addresses the use of automated tools and techniques and the evaluation of information intended to be used as audit evidence, including its relevance and reliability. Information processed by an AI tool is still information that the auditor must evaluate. If a tool extracts data from a client's documents, the auditor considers whether the extraction is accurate and complete before relying on it.

Quality management: technology is a resource

SQMS No. 1 treats technology as part of the resources component of the firm's system of quality management. The firm should address quality objectives and risks related to how technological resources are obtained or developed, implemented, maintained and used. For AI tools, a small firm's response might include:

• Approval before use. A short list of approved tools and approved use cases. Staff should not use unapproved public tools with client data.
• Fitness for purpose. Testing a tool on prior engagements before using it on current ones, and documenting the results.
• Training. Guidance on what the tool is and is not suitable for, and how to document its use.
• Monitoring. Including AI-assisted engagements in the firm's monitoring and inspection, and tracking errors attributable to tools.
• Service provider oversight. Where a vendor provides the tool, the firm evaluates the vendor as a service provider under SQMS No. 1.

See our guide to SQMS 1, SQMS 2 and SAS 146 for the wider framework.

Data confidentiality and third-party service providers

Client data is confidential under the AICPA Code of Professional Conduct's Confidential Client Information Rule (ET 1.700.001). Sending client data to an AI vendor, including a general-purpose chatbot, can be a disclosure to a third-party service provider.

Two interpretations are directly relevant:

• Disclosing Information to a Third-Party Service Provider (ET 1.700.040). Before disclosing confidential client information to a third-party service provider, a member should have a contractual agreement with the provider requiring it to maintain confidentiality and be reasonably assured that the provider has procedures to prevent unauthorized release. If no such agreement exists, the member should obtain the client's specific consent before disclosing.
• Use of a Third-Party Service Provider (ET 1.150.040). Before sharing confidential client information with a third-party service provider used to help provide professional services, a member should inform the client, preferably in writing, that a third-party service provider may be used.

The Code also addresses the member's responsibility for the work of a third-party service provider under the general standards, including being competent to supervise and evaluate it.

Practical steps:

• Read the vendor's terms. Confirm whether client data is used to train models, how long it is retained, where it is stored, who can access it, and how it is deleted.
• Prefer enterprise agreements with confidentiality and data processing terms over consumer accounts.
• Update engagement letters to inform clients that third-party service providers, including technology providers, may be used.
• Minimize data. Send only what the task needs, and remove personal information such as Social Security numbers when it is not required.
• Consider other rules. State privacy laws, client contracts and, for any tax work, IRC section 7216 can impose additional requirements.

A sensible adoption path for a small firm

1. Pick one painful, checkable task, such as tying out PBC schedules or extracting sample support, and measure the time it takes today.
2. Choose a tool with acceptable confidentiality terms, and document the approval under your quality management system.
3. Pilot on a completed engagement and compare the output with the existing workpapers.
4. Write a one-page procedure covering how to use the tool, what to check, and how to document its use.
5. Expand gradually to the next task, and include AI-assisted files in monitoring.

What AI will not fix

AI does not replace the experience needed to understand a client's business, assess risk, challenge management's estimates, or decide what the evidence means. It also does not fix a weak methodology. If the firm's audit programs do not link risks to procedures, automating them makes the same problem faster. The most effective use of AI is to free experienced staff to spend more time on judgment, review and client conversations.

Frequently asked questions

Can auditors use public AI tools with client data?

Only if the firm meets its confidentiality obligations. Under ET 1.700.040, that generally means a contractual confidentiality agreement with the provider or the client's specific consent, and under ET 1.150.040, informing the client that a third-party service provider may be used. Consumer terms often do not provide adequate protections.

Does using AI change who is responsible for the audit?

No. The engagement partner remains responsible for managing and achieving quality under AU-C 220, and the firm remains responsible for its system of quality management under SQMS No. 1. Tool output must be reviewed like any other work.

Will AI solve the accountant shortage?

AI can reduce manual work and help existing staff handle more engagements, but it does not replace the judgment and experience the profession needs. Most firms will still need to hire, train and retain people.

How should AI use be documented in workpapers?

Record which tool was used, for what step, what inputs it received, and what the preparer and reviewer did to verify the output, so an experienced auditor could understand and re-perform the work.

CORAA prepares ready-to-review audit files for US CPA firms from QuickBooks books; you review and sign. See /us.

Sources

• AICPA, SAS No. 146, Quality Management for an Engagement Conducted in Accordance With Generally Accepted Auditing Standards (AU-C Section 220)
• AICPA, Statement on Quality Management Standards (SQMS) No. 1, A Firm's System of Quality Management
• AICPA, AU-C Section 200, 230, 240, 315, 320, 500 (as revised by SAS No. 142), 520 and 530
• AICPA Code of Professional Conduct, ET 1.700.001 (Confidential Client Information Rule), ET 1.700.040 (disclosing information to a third-party service provider) and ET 1.150.040 (use of a third-party service provider)
• Internal Revenue Code Section 7216 and Treas. Reg. 301.7216-1 through -3 (disclosure or use of tax return information)
• 2 CFR 200.512 (single audit report submission): https://www.ecfr.gov/current/title-2/section-200.512
• AICPA, Trends in the Supply of Accounting Graduates and the Demand for Public Accounting Recruits

AU-C references are to AICPA Professional Standards, paraphrased here; read the standard itself before relying on a requirement. This page is guidance, not professional advice. Report an error