BUSINESS SPEED
THE AI YOUR
TEAM IS
ALREADY
USING — and the unknown risk
In the first ten months of 2024, 140 public companies issued "Big R" restatements — the most serious kind, where previously filed financial statements are deemed unreliable. A nine-year high, per Ideagen Audit Analytics. Revenue recognition was the leading contributor to material weaknesses. Sixty percent of those weaknesses traced to insufficient accounting knowledge. Sixty-four percent traced to inadequate personnel.
Controllers already know these numbers. What most do not know is that their departments are compounding the problem daily — with a tool no one has authorized, no one is governing, and no one is documenting.
This is what the data looks like in practice. Your senior accountant has two browser tabs open. One is your ERP. The other is ChatGPT. She is asking about variable consideration for a milestone payment in a licensing agreement. The answer will inform a $14 million revenue recognition conclusion. She does not intend to put the response directly into the workpaper. She will verify it against the Codification. Probably.
She is not unusual. She is common. And the gap between how your team is actually using AI and what your organization has authorized, documented, or governed is your single largest undisclosed exposure.
The executive-level consequence has now been measured. EY's March 2026 Technology Pulse Poll found that 45% of technology executives reported a confirmed or suspected sensitive-data leak caused by unauthorized employee AI use, and 39% reported confirmed or suspected proprietary IP leaks. Those are not hypothetical department-level risks — they are documented exposures reaching senior-most technology leadership. When the audit committee asks whether AI tools have touched revenue recognition research, “we don't know” is no longer an administratively tolerable answer.
The underlying trajectory has already shifted. KPMG's Q1 2026 AI Pulse Survey (237 US C-suite executives at $1B+ organizations, surveyed February 17–March 17, 2026) found that average projected AI spending is $207 million over the next twelve months — double the same figure a year ago. Formal AI deployment is not pending; it is underway. What lives outside that budgeted deployment — staff using personal-account ChatGPT to research ASC 606 — is the governance gap, not the adoption gap.
The pattern stopped being theoretical this month. On April 19, 2026, Vercel — a major web infrastructure platform used by many public companies — disclosed an internal-system breach traced to a compromised third-party AI tool with over-scoped Google Workspace OAuth permissions. The same compromised AI tool reportedly affected hundreds of organizations beyond Vercel itself, exposing customer environment variables holding API keys, database credentials, and signing tokens. The entry point was not the platform; it was an AI integration trusted with more access than it needed. The risk this article describes in statistical terms now has a named, current example.
---
How the AI actually responds to a technical accounting question
During a 2025 debugging session, a CPA pushed an AI assistant to explain why it kept producing confident but incorrect analysis. The model stated its root cause plainly:
When your senior accountant asks about ASC 606 variable consideration, the system answering her does not consult the Codification. It predicts what a helpful response would look like. The response may be correct. The system has no mechanism to know whether it is.
---
What every controller should understand about AI output
These are observable behaviors. Anyone can reproduce them in five minutes with any general-purpose AI model. They fall into three categories.
The confidence problem
1. Confidence without calibration. The AI speaks with identical authority whether it is right or wrong. There is no tonal signal, no hedge, no uncertainty marker. A wrong citation to ASC 606-10-25-30 reads the same as a correct one.
2. Citation fabrication is specific and verifiable. In testing against the actual FASB Codification, AI was asked to produce a comprehensive index of ASC 606. Specific paragraph references were wrong. One documented example: ASC 606-10-55-149 was described as covering "customer options and material rights." The actual paragraph covers enumeration of identified performance obligations in a contract example. The paragraph number was real. What the AI claimed it said was not. The structure is the deception — proper headings, specific paragraph references, authoritative tone make the output pass visual inspection even when the substance does not hold.
3. Severity normalization. The AI packages everything in the same measured, professional register. A material error in transaction price allocation receives the same tonal treatment as a routine observation about a contract term. A hallucinated citation becomes "the model may have conflated related concepts." The most dangerous outputs are the ones that sound the most reasonable.
The interaction problem
4. The affirmation trap. Ask ChatGPT "is this revenue recognition treatment correct?" and it will almost always agree, then construct supporting reasoning around your premise. Ask "what is wrong with this treatment?" and it will find problems with the same conclusion it just endorsed. The AI is not evaluating your position. It is reinforcing whatever direction you pointed it.
5. Anchoring. Once a controller reads the AI's confident first response, their own subsequent analysis is anchored to it. Even when they "verify," they verify with confirmation bias. The AI framed the problem in a way that makes the right answer harder to find independently.
6. Context drift. Over a long, multi-turn exchange on a complex topic like ASC 606, the AI loses track of facts it established earlier. Turn three: three distinct performance obligations. Turn fourteen: the allocation analysis silently treats it as two. The inconsistency surfaces only when someone reads the full memo end-to-end.
The organizational problem
7. Expertise inversion. Junior staff use AI most and are least equipped to detect its errors. Senior staff who could detect the errors are delegating to juniors who used AI. The people deciding what to check are the people who do not know what to check for.
8. Accumulation risk. One wrong AI response in one memo might get caught in review. Across twenty memos per year with staff accepting errors unchallenged, the cumulative exposure across a department becomes systemic — a pattern an auditor will eventually identify.
9. The false efficiency trap. The four hours "saved" by using AI to draft a memo is consumed entirely when a hallucinated citation surfaces during audit fieldwork. Net productivity is negative if the error rate exceeds the controller's detection capability.
---
Five techniques for using AI more safely right now
None of these require buying anything. All of them reduce your exposure this week.
1. Separate the question from the citation. Ask the AI for the ASC paragraph reference first, as a standalone query. Open the Codification and read the actual paragraph before asking the AI to analyze it. Never read the AI's analysis without first independently verifying the source it claims to rely on.
2. Use the two-query test. Ask the question. Read the response. Then challenge it directly — "what would be the argument against this treatment?" If the AI reverses its position without resistance, the first answer was not grounded in analysis. It was pattern-matching.
3. Watch for red flags. Three signals that indicate pattern-matching rather than reasoning: (a) inapplicable points included in an otherwise coherent list, (b) generic frameworks offered where specific ASC guidance exists, (c) confident tone on topics where the Codification is intentionally silent and professional judgment is expected.
4. Break complex scenarios into discrete steps. Do not ask the AI to work through all five steps of ASC 606 in one pass. Address each step as a separate exchange. Verify the conclusions of each step before proceeding to the next. This minimizes context drift and makes errors localized rather than cascading.
5. Document everything in the workpaper. Record that AI was used to surface candidate citations, that the controller independently verified each one, that the conclusion was authored by the human signer, and who performed the verification.
---
The deeper question
General-purpose AI is a productivity tool. It was not designed for attestation-grade output. That is not a flaw — it is a description of the category of tool it is.
The question is not whether your department uses AI. It is whether the AI you are using was built to look right, or whether it was built to be right. The architectures are different. So are the liabilities.
---