Who Answers When Your Accounting AI Gets It Wrong?

The accountability gap nobody is talking about

Your firm deploys Microsoft Copilot to draft variance commentary on a client's financial statements. The tool writes an explanation for a 12% revenue variance that's technically accurate by the numbers but misses the business context — a one-time event that won't recur. Your staff reviews it quickly and signs off. The client relies on it in their own reporting. The number was right. The story was incomplete. Someone will eventually ask: who is responsible for the incompleteness?

Your firm is. Not Microsoft. Not the software vendor. You.

This isn't a theoretical problem. It's a contractual fact baked into every major AI tool's terms of service. Microsoft Copilot's license agreement states that Microsoft does not warrant the accuracy of AI-generated output. Intuit's Accountant Suite terms disclaim liability for AI recommendations. OpenAI's terms are even starker: "We do not control or endorse any AI-generated content." Salesforce, Google and others follow the same pattern. The tool is a tool. The responsibility is yours.

What happened this week to change the temperature

Two separate institutional signals arrived this week marking that the accountability gap is moving from abstract problem to concrete regulatory concern.

First: Accounting Today's senior editor published an op-ed on May 25 calling for an "AI Pecora Moment" — a reference to the 1933 Senate hearings that overhauled Wall Street accountability after the crash. The editorial argues that as AI moves from a support tool to a financial output tool, the regulatory architecture around accountability hasn't caught up. Who is liable when the AI system produces wrong outputs used in financial reporting? The edit board says the answer should be clearer than it is today.

Second: This same week, an institutional governance document addressed AI accountability at a global level as a structural governance concern requiring formal frameworks. The document doesn't create law, but it signals that accountability for AI outputs is now an institutional governance question, not just a technology ethics question. When governance bodies start treating something as a structural problem, regulators follow.

Neither signal was covered widely in the accounting press. Together, they mark the moment when accountability for AI in financial workflows stopped being a vendor problem and started being a governance problem.

The contractual reality

Here's what the contracts actually say. Microsoft's Copilot terms state: "We do not guarantee that the Services will be uninterrupted, timely, secure, or error-free, or that defects will be corrected. We do not guarantee that the Services or the information provided will be accurate." By clicking accept, your firm agrees that Microsoft bears no liability for accuracy. You do.

Intuit's Accountant Suite agreement mirrors the pattern: "Intuit does not warrant that any AI-generated content...is accurate, complete, or suitable for any particular purpose." The language is slightly different. The liability structure is identical. You warrant accuracy. They don't.

OpenAI's terms are the most explicit: "OpenAI may be unable to guarantee accuracy of any output, and you acknowledge that you will not rely on any AI-generated content for critical use cases." This is the vendor saying "we know this can be wrong and we're not responsible if it is."

These aren't edge cases or buried fine print. They're the standard contractual form for every major AI platform. The business model of the AI vendor is to shift liability downstream — to you, the deployer.

What the standards say about your responsibility

The AICPA Code of Professional Conduct, Section ET 1.300, states: "A member shall not sign or permit the member's name to be associated with any financial statement or report unless the member has first examined it and is satisfied as to the manner of its presentation and as to the matters it contains." The code makes no exception for AI-assisted work. The CPA is responsible.

In practice, this means: if your Copilot-generated reconciliation misses a variance, your firm missed it. If your AI tool writes variance commentary that's technically correct but contextually incomplete, your firm wrote it. If your variance analysis tool produces output based on stale data or a misunderstanding of business context, your firm is the point of professional failure.

The liability structure creates an asymmetry: the vendor builds the tool, the vendor gets paid for the license, but the vendor carries zero liability for output. You carry all of it. This is why some of the most sophisticated firms are treating AI adoption as a governance question, not a productivity question.

What regulators are signaling

The SEC's 2026 Exam Priorities explicitly include language about AI output accuracy: "Examiners will review for accuracy registrant representations regarding their AI capabilities and the accuracy of output from AI systems used in financial reporting or client advisory roles." This isn't a warning. It's a scope statement. The SEC is making AI accuracy part of the examination standard.

The PCAOB's Technology Innovation Alliance has been working since 2025 toward standardized frameworks for how auditors should document and evaluate AI use in audit workflows. The TIA's working group is focused on four areas: data standardization, audit documentation standards, quality assurance protocols and auditor training requirements for AI tools. They all lead to the same place: AI output in regulated workflows will require formal, auditable documentation of how the AI was used, what checks were applied and who reviewed it before it was relied on.

Regulators aren't banning AI in accounting workflows. They're moving toward formal accountability frameworks. The accountability question — who is liable when the AI is wrong — is moving from contractual obscurity to regulatory explicitness.

What this means for firms that have already deployed AI

If your firm is already using AI tools in client work — reconciliation, variance analysis, draft reporting or data analysis — you need to close three gaps immediately.

Documentation. Create a documented quality assurance process for any AI-generated output before it goes into client work. Document what AI tool was used, what data was fed into it, what human review was applied and who approved the output before it was delivered to the client. This documentation is your defense if someone asks later "how did you verify this was right?"

Engagement letter clarity. Your engagement letter or SOW should disclose that AI-assisted analysis is part of your process. This doesn't need to be alarmist. It should be factual: "We may use software tools including AI-assisted analysis to improve efficiency in reconciliation, variance analysis and workpaper documentation. All output is subject to professional review and CPA judgment before delivery to you." That sentence puts the client on notice that AI is involved and that professional judgment is still the filter.

Insurance carrier notification. Your malpractice carrier should know that you're deploying AI in client work. You don't need permission, but you need disclosure. Some carriers have updated their policies to address AI use. Some haven't. The conversation needs to happen now, not after an audit finding or client dispute.