Artificial intelligence is moving deeper into business operations, but the controls surrounding its use do not always provide a complete view of the risk. A recent EY survey found that 98% of surveyed senior AI executives reported formal AI governance policies, yet 47% said their organizations had bypassed governance processes for urgent AI deployments. More than one-third, 36%, reported an AI incident or failure that caused materially negative consequences, including financial damage, operational disruption, or data loss.
For CPAs and financial professionals advising businesses, the concern extends beyond the existence of an AI policy or risk assessment. A documented control can show that a process has been reviewed. It does not necessarily show how a weakness in that process could affect cybersecurity, compliance, operations, and financial performance at the same time. As AI becomes embedded across business functions, risk assessments need to provide visibility into how exposures interact, where they can compound, and how much warning exists before the consequences become material.
When separate risks stop behaving separately
Traditional risk assessments often evaluate exposures according to defined categories. Cybersecurity may sit with information technology. Compliance may belong to legal or regulatory teams. Operational risk may be evaluated separately from financial exposure. Each category can have its own owner, controls, and reporting process.
The problem arises when the risk itself does not remain inside those boundaries.
Consider an AI tool used to process financial information or support a business decision. If the tool relies on inaccurate data, the initial issue may appear to be data quality. If that information feeds a financial forecast, the problem can influence management decisions. If sensitive information is exposed through the application, cybersecurity and privacy risks emerge. If the organization cannot explain how an automated decision was reached, compliance and governance concerns may follow.
One weakness can therefore create several risk events. Assessing each one independently can create a visibility problem because the assessment does not show how the events interact or how the consequences of one can amplify another.
This is becoming more relevant as finance leaders take greater responsibility for technology decisions. Deloitte’s September 2026 Finance Trends 2027 research found that 54% of surveyed finance leaders are leading enterprise AI and technology capital-allocation decisions, while 48% oversee AI and technology spending and cost controls.
For CPAs advising those organizations, the assessment cannot end with identifying an individual AI risk and assigning it a score. The next question is what else that risk can affect.
Risk ownership is as important as risk identification
One of the first questions a CPA can ask is simple: Who owns this risk?
That is different from asking who operates the process. An employee or department may manage an AI application without being accountable for the risks created by its data, decisions, or connections to other business functions. When ownership is unclear, a risk can fall between technology, finance, compliance, and operations without any one function having a complete view.
The same principle applies to controls. The existence of a policy does not establish that an organization can detect when the control is being bypassed or failing.
The EY survey illustrates the problem. Among respondents whose organizations use agentic AI, 26% said their organizations could not detect unauthorized AI agents operating internally. That raises a fundamental risk-management question: How would the organization know?
A control without meaningful detection provides limited warning. For CPAs, that means an assessment should examine not only which controls exist, but also what evidence demonstrates that those controls operate as intended.
Several questions can help expose these gaps:
- Who owns the underlying risk? The person running the process and the person accountable for its risk may not be the same.
- What else depends on this control? A weakness should be traced into the financial processes, data flows, vendors, and business functions that rely on it.
- How would the organization know the risk is developing? Monitoring should identify meaningful changes in access, outputs, exceptions, data movement, or other leading indicators.
- Where else could the same weakness exist? A control failure in one process may indicate that other processes were designed around the same assumption.
The last question is particularly important. Finding and correcting one weakness does not necessarily mean the underlying risk has been reduced. If the same control design or assumption exists across multiple processes, fixing one finding leaves the broader exposure intact.
Likelihood and severity do not tell the entire story
Traditional risk methodologies commonly evaluate two factors: likelihood and severity. Likelihood measures the probability that a risk event will occur, often informed by historical frequency. Severity measures the potential impact of that event.
Both remain necessary, but AI-related risk can expose what those two factors do not show.
A low-likelihood event may still require significant attention when its potential impact is severe, the organization has little ability to see it coming, and the underlying weakness exists across multiple processes.
This is where predictability and compounding become important to the assessment.
Predictability considers the strength of the warning signals available before a risk event occurs. A risk with weak or absent leading indicators presents a different management challenge from one that gives an organization enough warning to respond.
Compounding examines how a risk event can interact with other exposures, cascade into additional functions, or amplify consequences. For an AI-enabled process, a single control weakness could affect data integrity, financial reporting, privacy, cybersecurity, operations, or regulatory compliance.
A point-in-time score may identify the initial exposure. A broader assessment must show what that exposure can become.
CPAs should follow the risk beyond the initial finding
As CPAs take on larger advisory roles, their value increasingly includes helping clients understand how financial and nonfinancial risks connect.
That requires looking beyond the immediate finding. If an AI system produces unreliable information, the assessment should identify which decisions depend on that information. If an AI vendor experiences a security problem, the assessment should determine which data, processes, and financial activities rely on that provider. If employees bypass an AI governance process, advisers should ask why the bypass occurred, what controls were avoided, and where else the same behavior could happen. The objective is to distinguish between remediating a finding and reducing the underlying risk.
AI makes that distinction increasingly important because technology can connect processes that organizations still assess separately. Cybersecurity, compliance, operations, and financial exposure may have different owners, but a risk event can move across all four.
CPAs are well-positioned to challenge that fragmented view. Asking who owns the risk, how the organization would detect it, where the same weakness exists, and what other processes could be affected provides a clearer picture than an isolated risk score.
As AI becomes more embedded in financial and operational decisions, risk assessments will need to keep pace. Organizations that can identify how risks interact, recognize meaningful warning signals, and trace weaknesses across business functions will be better equipped to address an emerging exposure before it compounds into a larger financial or operational problem.
ABOUT THE AUTHOR:
Dr. Jeffrey L. Edwards is the founder and chief risk strategist of FFERM Technologies, a financial risk intelligence company serving community banks, credit unions and other financial institutions. He has more than 30 years of financial services experience, including two decades focused on operational, credit, market and third-party risk across audit, risk oversight, and business operations. A professor and former technology professional, Dr. Edwards developed the Four-Factor Enterprise Risk Management methodology to add compounding and predictability to conventional likelihood-and-severity scoring, turning static ratings into a clearer view of how risk behaves.
Photo credit: Sasun Bughdaryan/Unsplash
Sign in to get access to this free resource, and all of our whitepapers and reports.
Download this content today!
Register Now Already registered? Click here to Log In
