When Clients Use AI, What Changes for the Accountant?

Not every client is using AI. Enough businesses and employees are using it—sometimes without a formal company process—that accountants can no longer treat AI-generated work as an edge case.

The latest U.S. Census Business Trends and Outlook Survey estimated that 23.2% of covered nonfarm employer businesses used AI in a business function during the August 10–23, 2026 reference period. A broader Federal Reserve survey found that 46% of employer firms said the business or its employees currently used AI.

Those figures measure different populations and behaviors. The Census estimate is a high-frequency population estimate covering a two-week period. The Federal Reserve result comes from a weighted convenience sample and includes use by either the business or its employees. They should not be treated as competing measurements of the same thing.

The practical conclusion is simpler: AI-generated questions, summaries, classifications, forecasts, and recommendations can now enter client conversations. The accountant’s core responsibility does not change. The path from source evidence to conclusion becomes more important.

‍

Accountants already need to understand where a number came from. AI adds another layer between the original record and the client’s conclusion.

A client may arrive with a polished cash-flow forecast, an expense classification, a tax explanation, or a summary of business performance. The presentation may look complete even when the underlying records, assumptions, or source citations are missing.

That changes the first question. Instead of immediately asking whether the output is correct, ask:

What records, assumptions, and instructions produced this result?

This is not a demand to audit every AI-assisted email or brainstorming note. The level of verification should follow the intended use and potential consequence. An AI-generated draft used to prepare meeting questions presents a different risk from a forecast used to approve hiring or a conclusion used in a tax filing.

AICPA’s 2026 Technical Questions and Answers section 400.02 is nonauthoritative, but its central point is useful: members need professional judgment to decide whether technology output is appropriate for its intended purpose. In federal tax practice, the IRS Office of Professional Responsibility is more explicit for practitioners governed by Circular 230: facts, citations, calculations, and conclusions in AI-created material require appropriate human review.

AI does not become the source merely because it produced the answer. The invoice, contract, ledger, report, calculation, or governing authority remains the evidence.

‍

Some client AI use will be visible. A client may send a chatbot transcript or say that a forecast came from an AI tool. Other use may be embedded in software or performed by an employee without a formal company process.

A 2026 Census working paper on AI diffusion illustrates that gap. During its supplement period, 18% of firms reported AI use in a business function, while 23% reported workers using AI in work-related tasks. Among firms reporting worker use, 36% showed no corresponding indication of formal firm-level use.

The firm and worker questions covered different reference periods, so that 36% is not a precise measure of unauthorized or “shadow” AI. It does show why a company-level answer may not reveal every place AI is affecting work.

Accountants can surface this without turning every client meeting into a technology audit. Add a short question to onboarding and recurring reviews:

Has anyone used an AI tool to prepare, classify, summarize, forecast, or interpret information that entered this workflow?

If the answer is yes, determine whether the output remained an internal draft or influenced the books, a filing, a financial statement, a recommendation, or a material operating decision.

‍

The Federal Reserve survey found that accuracy was the most commonly reported AI challenge, cited by 46% of firms using it. That does not mean 46% received incorrect answers. It means accuracy was a material concern among adopters.

The appropriate response is a clear verification threshold.

For a classification, return to the transaction and the accounting policy or chart-of-accounts logic. For a forecast, check the opening balance, time period, expected receipts, committed payments, formulas, and assumptions. For a cited rule, open the governing or authoritative source rather than relying on a generated citation.

A useful three-part check is:

1. Source: Which original records or authorities support the input?
2. Method: How were the figures, classifications, or conclusions produced?
3. Use: What decision or deliverable will depend on the output?

The more consequential the use, the stronger the evidence and review should be. A general discussion may need a reasonableness check. A filing position, accounting treatment, compliance conclusion, or decision dependent on entity-specific facts may require a qualified professional applying the relevant rules and framework.

This preserves proportionality. The goal is not to reject work because AI touched it. The goal is to prevent an unsupported output from acquiring authority as it moves through the workflow.

‍

Client AI use can change where sensitive information travels. An owner may paste a profit-and-loss statement into a public chatbot. An employee may upload invoices, payroll information, customer records, or tax documents without knowing how the provider stores or uses that information.

AICPA & CIMA’s 2026 ethics guidance recommends understanding an AI tool’s limits, where data is stored, and what safeguards protect sensitive information. The exact professional and legal duties vary by credential, engagement, jurisdiction, information type, and service.

The accountant does not need to become the client’s software administrator. But the firm should be able to ask:

• Was the tool approved for this type of information?
• Did the prompt or upload contain confidential, personal, payroll, tax, or customer data?
• Who can access the output and conversation history?
• Is there a retained record of the source, prompt, assumptions, and human review?

For federal tax practitioners, additional confidentiality and due-diligence rules may apply. Those requirements should be evaluated under the applicable law and engagement rather than converted into one universal policy for every accountant or bookkeeper.

‍

When AI-generated material affects the books, a forecast, a tax question, or an operating decision, five questions can restore a usable evidence trail:

1. Where was AI used? Identify the tool and the specific task.
2. What data entered it? Check whether confidential, personal, tax, payroll, or customer information was exposed.
3. What output entered the workflow? Separate brainstorming from classifications, calculations, forecasts, journal support, or advice.
4. What source evidence supports it? Return to invoices, contracts, ledgers, reports, calculations, or governing authority.
5. Who reviewed and approved it? Assign responsibility before the output reaches a filing, financial statement, recommendation, or material decision.

Consider a client who brings an eight-week cash forecast produced by a public chatbot. The forecast may look reasonable, but the client cannot identify the prompt, the receivables report used, whether payroll and tax payments were included, or who checked the formulas.

The immediate task is not to debate whether the chatbot is good or bad. It is to rebuild the evidence trail:

• Confirm the current cash position.
• Identify expected receipts and their timing risk.
• List committed payments and excluded obligations.
• Recalculate the forecast using explicit assumptions.
• Explain what the forecast can and cannot support.

The same five questions work whether the output came from a public chatbot, an AI feature inside business software, or an internal model. The required review changes with the consequence.

‍

When clients can generate a definition, summary, or plausible recommendation in seconds, access to general information becomes less distinctive. That does not make accountants less relevant. It changes where their value is easiest to see.

The harder work is often:

• Recognizing that two systems use different definitions or periods.
• Finding the transaction or assumption that caused a result.
• Distinguishing an unusual item from a genuine trend.
• Explaining what evidence supports a conclusion.
• Identifying when the available information is insufficient.
• Applying professional judgment within the relevant engagement and standard.

AI can reduce preparation work. It can also increase the volume of polished material that arrives without a complete source trail. Accountants are well placed to verify evidence, investigate exceptions, design sensible controls, and explain the implications in business terms.

That opportunity is not automatic. Firms still need clear scope, adequate competence, reliable review, and honest communication about what people and technology each contributed.

‍

Start with one recent client workflow: monthly reporting, cash forecasting, expense review, bookkeeping cleanup, tax preparation, or an advisory meeting.

Ask where AI may already be present. Then define:

1. Which AI-assisted outputs the client should disclose.
2. Which source records must accompany a consequential output.
3. Which information may not enter an unapproved tool.
4. Who reviews the result before it affects a deliverable or decision.

Keep the process proportionate. A short intake question and a clear evidence threshold are more useful than a long policy nobody follows.

For firms that want owners and advisors to work from the same source-backed business context, Veltrix for accounting and advisory workflows brings connected business data into a shared workspace, shows the reports behind AI answers, and never changes the source tools. It supports the accountant’s judgment; it does not replace it.

Sources

• U.S. Census Bureau, Business Trends and Outlook Survey
• U.S. Census Bureau, latest national BTOS workbook
• U.S. Census Bureau, The Microstructure of AI Diffusion
• Federal Reserve Banks, 2026 Report on Employer Firms
• AICPA, TQA section 400.02: Using the Output of Technology
• AICPA & CIMA, Ethics, accountancy, and AI-powered tools
• IRS Office of Professional Responsibility, Responsible AI Use in Federal Tax Practice