AI Ethics in Accounting

The Ethical Rule

The responsible person must still:

1. 1Understand the purpose and consequence of the work.
2. 2Use an appropriate tool and approved environment.
3. 3Evaluate the quality and limitations of the data.
4. 4Assess whether the output is fit for purpose.
5. 5Challenge errors, omissions, bias, or unsupported claims.
6. 6Protect confidential information.
7. 7Explain the work honestly.
8. 8Retain appropriate evidence and review.
9. 9Escalate when the risk exceeds their authority or competence.

Soft-label· IESBA technology guidance states that professional accountants remain responsible for judgments and decisions made in their work regardless of the level of automation or sophistication of the technology.

The Ethical Framework: Five Principles

The following table translates the profession’s ethical principles into everyday AI decisions.

Soft-label IESBA guidance identifies technology as a source of ethical risk across competence and due care, objectivity, confidentiality, and independence. It requires accountants to assess whether technology outputs are fit for purpose, understand relevant assumptions and limitations, evaluate data quality and bias, and determine the appropriate extent of reliance.

1. Integrity: Do Not Turn a Plausible Answer into a Fact

Generative AI can produce text that is fluent, specific, and confident even when it is inaccurate, incomplete, or fabricated. Soft-label· NIST materials call this risk confabulation: confidently stated but erroneous or false content that may mislead or deceive users.

The ethical risk appears when an accountant presents that output as though it were verified.

Examples of integrity failures

• Sending an AI-drafted tax explanation without verifying the authority, facts, and calculation.
• Saying a financial-reporting conclusion is “GAAP-compliant” because an AI tool said so.
• Repeating a vendor’s claimed AI accuracy percentage without knowing the test population, conditions, or limitations.
• Calling a workflow “fully automated” when staff still make material classification, review, or approval decisions.
• Inserting AI-generated citations into a memo without opening and checking the underlying source.
• Reporting an AI-generated business explanation as a known fact when the ledger only shows a numerical change.

Example: the unsupported variance narrative

2. Objectivity: Resist Automation Bias and Human Pressure

Automation bias is the tendency to favor technology-generated output even when other information raises questions about its reliability. Soft-label IESBA materials now explicitly name automation bias as a bias that can impair objectivity and professional judgment.

Automation bias does not require a sophisticated AI system. It can happen when a person accepts an auto-categorized transaction, a matching recommendation, an audit anomaly score, an AI-generated legal explanation, or a draft disclosure because the output looks authoritative.

Two opposite mistakes

Pressure can also impair objectivity

• A manager wants the close done before the review is complete.
• A client wants a favorable tax answer.
• A partner wants to advertise that the firm is “AI-powered.”
• A vendor promises a time-saving result that management wants to believe.
• A team wants to avoid admitting that an automated workflow failed.

The ethical response is not to be difficult for its own sake. It is to identify the threat, document the concern, seek appropriate review, and refuse to turn pressure into an unsupported conclusion.

3. Competence and Due Care

Competence does not mean becoming a machine-learning engineer. For most accountants, it means understanding the tool and workflow well enough to use them safely for the assigned task.

Soft-label IESBA technology revisions require professional accountants to understand, be able to explain, and evaluate technology used in their professional activities, and to keep pace with relevant technology developments.

Due care requires review

Using AI responsibly is not “prompt once, send immediately.” Due care includes:

1. Defining the task and required output.
2. Using approved data and tools.
3. Checking source data and assumptions.
4. Reviewing output against authoritative or controlled evidence.
5. Testing calculations and edge cases.
6. Documenting limitations and open questions.
7. Obtaining review at the appropriate level.

Competence warning signs

4. Confidentiality: Data Has a Lifecycle

The issue is bigger than “don’t share passwords.” AI can create new paths for information to move: prompts, uploads, chat history, browser extensions, connectors, APIs, model providers, logs, support tickets, backups, and generated outputs.

Soft-label IESBA technology guidance says confidentiality obligations extend across the full data lifecycle: collection, use, transfer, storage, dissemination, and lawful destruction. It also states that using client data for purposes such as training AI models requires proper authorization and clear boundaries around consent.

Before uploading any client or firm data

Third-party service providers

Soft-label Soft-label· AICPA Code of Professional Conduct materials address use of third-party service providers. Before disclosing confidential client information to a third-party provider, a member should have a contractual arrangement requiring confidentiality and reasonable protective procedures or obtain specific client consent. The client should be informed, preferably in writing, that a third-party provider may be used; if the client objects, the member should not use the provider or should decline the engagement.

An AI application may be a third-party service provider when it receives or can access confidential client data. The fact that it is called an “assistant,” “copilot,” “chatbot,” or “productivity tool” does not change the confidentiality analysis.

5. Professional Behavior: Do Not Mislead About AI

AI-washing means exaggerating, misrepresenting, or making unsupported claims about AI use or AI performance. It can appear in client pitches, websites, engagement proposals, investor communications, performance claims, internal dashboards, or job descriptions.

Ethical ways to describe AI use

Professional behavior also includes honesty about limits

Do not hide:

• That a result was AI-assisted where the fact matters to the client or engagement.
• That the output is a first draft or candidate list rather than a conclusion.
• That the tool could not access all relevant documents.
• That a number is estimated, provisional, or subject to review.
• That an exception remains unresolved.

6. Fairness and Bias: Accounting Systems Can Affect People

Bias is not only a concern for consumer lending or hiring algorithms. Accounting and finance workflows can influence which vendors are flagged, which customers receive credit attention, which employees’ expenses are scrutinized, which transactions appear suspicious, or whose data is treated as an exception.

Soft-label Soft-label· NIST Generative AI Profile materials identify harmful bias and homogenization as risks, including performance disparities across subgroups or languages caused by nonrepresentative training data and incorrect assumptions about performance.

Questions to ask

• Does the tool work equally well across names, languages, locations, document formats, currencies, and business types represented in the data?
• Are certain customers, vendors, employees, or transactions more likely to be flagged because of incomplete or biased data?
• Is a classification rule using a proxy for a protected or sensitive characteristic?
• Are error rates being measured separately for meaningful groups?
• Does a reviewer have enough context to challenge a flagged item?
• Could a false positive cause reputational, financial, employment, credit, or compliance harm?

Example: vendor matching

7. Transparency and Explainability

Explainability does not require revealing every model weight. For accounting work, a practical standard is simpler:

Minimum documentation for an AI-assisted workflow

Soft-label NIST AI RMF materials call for documentation of test sets, metrics, tools used for evaluation, human-oversight processes, and evidence that systems are valid and reliable for the conditions in which they are deployed.

8. Accountability and Supervision

A common ethical failure is diffused responsibility:

• The preparer says, “The tool generated it.”
• The reviewer says, “The preparer used the approved tool.”
• The manager says, “The system had a high confidence score.”
• The vendor says, “The output is only a suggestion.”

None of those statements identifies who actually decided that the output was appropriate to use.

Assign roles explicitly

The accountability test

The AI Ethics Decision Tree

Interactive stepper through eight questions before deploying or relying on an AI-assisted output.

ABC Coffee Shop: Ethics at Month-End

Compare a safe variance-narrative workflow with an unethical shortcut that overclaims “AI-verified” analysis.

Ethical Prompts Versus Risky Prompts

Expand each risky prompt to reveal a better, copyable alternative.

Using only `tbl_Approved_PnL`, identify the account-level changes that mathematically explain the decline in operating income. Label unsupported causes `Needs Evidence` and do not infer business reasons.

Rely, Verify, or Reject

Knowledge Check

Five questions on responsibility, automation bias, evidence vs summaries, confidentiality beyond training, and incomplete evidence.

Final Ethics Checklist

Sources & Further Reading

Selected soft-label IESBA, NIST, AICPA, IRS, SEC, and FTC materials. Guidance and enforcement pages change; recheck official sources for current requirements. This page is educational and does not replace legal, regulatory, privacy, cybersecurity, tax, audit, or professional-standards advice.

• IESBA — Final Pronouncement: Technology-related Revisions to the Code
• IESBA — 2024 Handbook of the International Code of Ethics for Professional Accountants
• NIST — Artificial Intelligence Risk Management Framework (AI RMF 1.0)
• NIST — Artificial Intelligence Risk Management Framework: Generative AI Profile (NIST AI 600-1)
• AICPA Code of Professional Conduct — confidentiality & third-party service providers
• IRS — Disclosure or use of tax return information by preparers (IRC §7216 overview)
• IRS — Section 7216 information for tax professionals
• IRS OPR — Introductory Guidelines for Responsible AI Use in Federal Tax Practice (Alert 2026-19 PDF)
• SEC — Press Release 2024-36: Charges against Delphia and Global Predictions (AI-washing)
• FTC — Safeguards Rule: What your business needs to know
• FTC — Final order against Workado, LLC (unsupported AI detection accuracy claims)

What's Next?

Continue with related career and ethics lessons, deepen tool evaluation, or return to the AI pillar hub.