Auditing has always involved a strict set of rules, standards and controls. Whether an organization is scrutinizing its finances, technology or other assets, the goal is to establish a predictable series of events that leads to a predictable set of conclusions.
In 2026, however, AI is rewriting the rulebook. Predictability is no longer a given: Ask an LLM the same question multiple times, and it's likely to produce a different answer at each time. This isn't a bug; it's a feature. Yet, for CIOs striving for safe and responsible AI, the lack of regularity can be maddening.
Today, AI agents score loan applications, assess inventory and shape the numbers that appear in financial statements. They also interact with IT and operational systems. All of these AI use cases require consistency to be effective and trustworthy, yet current audit models don't account for AI's inherent variability.
"There's still a bit of a disconnect on how to audit AI, because AI is moving faster than frameworks are catching up — and auditors like repetition," said Lauren Kornutick, a senior director analyst in the analytics and AI group at Gartner.
Related:T-Mobile CTO on injecting AI into network resiliency
The problem becomes clear when auditors ask how various AI decisions came about and many CIOs can't explain. Conventional audit trails track a defined set of activities and events, and they presume a human signed off on each step. LLMs work differently. Prompt logs and model outputs reveal what went into a system and what came out, but not the process in between or whether the results are reliable.
As auditors push for defensible AI decision-making, CIOs are in the crosshairs. There's a need for new tools and new approaches. Organizations "should be thinking about how they create and retain that record as the system operates — rather than trying to reconstruct it later," said Jennifer Kosar, AI assurance leader at PwC US.
Friction points in the current auditing process
Until recently, AI audits mostly focused on taming hallucinations and extracting accurate results from GenAI models. "Today, the focus extends to understanding how AI is specifically being used in the process," said Mike Sawyer, audit partner and U.S. audit business leader for AI and innovation at KPMG.
This includes understanding what information went into an AI model, how AI output is generated and used, and what level of reliance an organization places on the information. "Auditors are looking for a clear record connecting the technology's specific use case and agent design to the process management it has in place to validate the AI judgment," Sawyer said.
Related:CISOs shouldn't sleep on banning smart glasses
At least, that's the goal. In many cases, auditing challenges don't originate with the model; they surface when work moves among teams. The technology group may understand how a system operates, while the finance team understands what a number means — but there's no record to connect the two. The result is a fuzzy concept of ownership that undermines oversight and incident response. Monitoring tools flag exceptions that go unresolved because no team can claim full responsibility.
The AI model may be working perfectly, but the audit will still fail if the organization can't prove it.
The numbers paint a troubling picture. Gartner's 2026 AI Governance Survey of 697 AI leaders found that business leaders rated AI bias and fairness monitoring "very" or "moderately" difficult 67% of the time. Output accuracy mechanisms followed right behind at 66%, and responsible AI policy landed at 64%.
Starting from scratch
It isn't a lack of effort on the part of CIOs and auditors. The problem, Kornutick said, is that there's no established playbook. Auditing frameworks and standards, including cybersecurity, have been built atop years of expertise, but the controls specific to AI — drift, bias monitoring, output accuracy — don't have that advantage.
Related:The CIO's maintenance backlog is now a business risk
"There are not a lot of precedents around to determine what good looks like," she said.
In many cases, auditors wind up trying to force-fit AI into a GRC process, Kornutick said. "It's almost like using waterfall development processes in a hyper agile world, because AI even moves faster than an agile process does."
What's more, with generative AI, there's no way to understand exactly how the response was formed because there's no step-by-step path to trace from prompt to output. "You can't describe how we get from point A to point Z," Kornutick said. "You can only interpret that the intention of the person writing the prompt is aligned to the output."
Adopting a boutique approach to AI models doesn't resolve the issue either. It's possible to control every variable on the input side and still end up with unpredictable output and unverifiable results, which once again undermines confidence in the AI audit.
Knowing which questions to ask
An effective AI auditing framework revolves around a basic question: how much information should the organization collect and keep? Because real-time inference logs are expensive, many organizations use continuous monitoring only for their highest-risk systems and rely on point-in-time audits for the rest. While this lowers costs, it results in lower visibility — and may even leave enterprises in violation of external regulations.
The temptation to focus solely on technical issues is misguided, Kornutick said. AI traceability and data retention are largely rooted in legal factors. Even when an organization optimizes thresholds for what's defensible, regulatory penalties can follow. The more comprehensive approach is to pursue a path of due diligence and have mitigating controls in place.
At that point, the question becomes: "Could we have prevented it? And then if we couldn't prevent it, how quickly could we detect and remediate it?" Kornutick said.
Defensible AI hinges on four key questions, Sawyer said. "What data was used? How was the output evaluated? Who was responsible for reviewing it? And how did it influence the final decision?" A CIO who can answer these questions clearly — and map answers to identified risks — is better equipped to support AI-enabled decision-making, he said.
Building a framework around accountability
To address the issues that arise from unclear ownership, enterprises should consider implementing a framework that addresses responsibility.
Kosar explained that a successful accountability framework revolves around three layers of protection:
The first layer addresses the design of the AI system. This includes an assessment of related risks and associated controls, as well as the determination of whether certain risks are acceptable.
The second layer periodically reviews those decisions and raises challenges as risks evolve.
The third layer independently assesses whether the first two layers performed correctly.
When implemented effectively, a framework of accountability minimizes time-consuming and costly forensic audits. With a clear audit trail of decisions, tests, approvals, exceptions, and changes logged, visibility improves.
Several vendors have recently entered the space, suggesting that this may be a growing area of enterprise investment in the years to come. These range from audit management platforms that deliver general governance frameworks, such as Optro (formerly AuditBoard), TeamMate, Workiva and Archer, to tools like Credo AI, IBM watsonx.governance and Holistic Ai, which track bias, accuracy and drift as they appear.
The key is not to delay the AI audit strategy. Too often, Kosar said, CIOs believe that AI doesn't yet fall within the scope of auditing because their organization hasn't yet rolled out a large-scale AI framework. Yet AI features increasingly pop up in enterprise software that business teams already use, she noted. As a result, existing controls may no longer work — and without an audit, there's no way to detect a problem.
Owning the risk
CIOs must stop treating auditing as a point-in-time exercise, Kosar said. When organizations get the audit equation right, they know where AI is running, recognize when usage patterns change, and can continuously test and monitor systems that carry the greatest risk. At that point, they are equipped to govern AI proactively.
According to Sawyer, it is vital to view AI audit readiness as a business capability rather than a compliance chore. Organizations that adopt this mindset are able to move faster and innovate with greater confidence, he said.
For CIOs, that's where the opportunity lies. Kornutick argues that CIOs are the key link in the AI auditing chain — and their involvement makes or breaks an initiative. "This is really the CIO's chance to step up and take ownership of it," she said. "They should lean into it."