Artificial intelligence has stopped being the shiny thing in the innovation lab. It’s approving credit, screening job candidates, setting prices, flagging fraud, and in some organizations quietly performing core financial functions. Which means the question boards and regulators are starting to ask isn’t “are we using AI?” — it’s “who is checking the AI?”
Increasingly, that question lands on internal audit. And most audit functions aren’t ready for it. AuditBoard’s 2026 Focus on the Future survey found a stark disconnect: organizations have high expectations for internal audit to provide AI assurance, while auditors themselves report low readiness and continue to underestimate AI’s impact. That’s not just a skills gap. When the function responsible for independent assurance can’t see into the systems making consequential decisions, the gap itself becomes an enterprise risk.
AI Outgrew the IT Audit Bucket
For years, anything with a server attached got routed to IT audit. That model breaks down with AI for a simple reason: AI systems aren’t infrastructure, they’re decision engines. When a model influences who gets hired, who gets a loan, or what a customer pays, a flaw isn’t a technical incident — it’s a compliance violation, a discrimination claim, a regulatory disclosure problem, and a reputational event, all at once.
Three things follow from that. AI risk is enterprise risk, cutting across compliance, ethics, strategy, and operations rather than sitting in a technology silo. Ownership is diffuse — models get built by data science teams or vendors, deployed by business units, and governed loosely by everyone and therefore no one. And traditional controls don’t fully apply, because AI systems change over time; a control assessment that was accurate in January may be fiction by June. Static, point-in-time assurance was built for systems that stay put. AI doesn’t.
The Three Risks Audit Leaders Keep Underestimating
Ask most auditors what AI risk means and you’ll hear about model accuracy. Accuracy matters, but it’s one dimension and rarely the most dangerous one. The bigger exposures are systemic, and they compound weaknesses most organizations already have.
Data risk is the foundation. Every AI system inherits the flaws of its training data: bias, gaps, data collected without a lawful basis or used beyond its original purpose, drift over time, and missing lineage that makes outputs impossible to trace back to inputs. This is where AI governance and privacy compliance collide — an AI audit conducted on top of a weak data governance program produces incomplete assurance at best and false comfort at worst. Assess the data governance program first; the model second.
Governance gaps are the real exposure. The most common AI failure isn’t a broken model — it’s the absence of anyone who could have caught one. No centralized inventory of AI systems in use. No clear ownership. No policies governing development and deployment. No approval gates or ongoing monitoring. Third-party AI makes this worse: the vendor builds and maintains the model, the business consumes its outputs, IT wires it into systems, and risk and compliance may have almost no visibility into how it actually works. Nobody owns the whole picture — which is exactly the kind of cross-functional blind spot internal audit exists to find.
Regulatory risk is accelerating and fragmenting at the same time. There is no single comprehensive federal AI law in the U.S., yet enforceable AI obligations already exist across cities, states, and federal agencies. That combination creates a distinctly modern compliance problem: organizations can be non-compliant without knowing it, because the obligations are emerging in different jurisdictions, on different timelines, aimed at different use cases.
The Rules Already on the Books
New York City’s Local Law 144 is the clearest proof that algorithmic auditing has arrived. Employers using automated employment decision tools must obtain an annual independent bias audit, publicly post a summary of the results, and notify candidates before AI is used in hiring decisions. Read that list again from an internal auditor’s perspective: an annual audit requirement, public disclosure of findings, and consumer notice obligations — a full assurance regime, already in force, for a single AI use case.
Colorado’s SB21-169 targets insurers’ use of algorithms and predictive models, but it set a precedent that reaches well beyond insurance. It prohibits unfair discrimination from algorithmic systems, requires organizations to test models for bias and document the results, and mandates a formal risk management framework with ongoing monitoring. That’s the significant move: from transparency to demonstrable, documented governance — territory squarely within internal audit’s mandate. Colorado has since doubled down with its broader AI Act covering high-risk AI systems across industries, and other states are following the pattern.
At the federal level, the SEC isn’t waiting for AI-specific legislation. It’s applying existing disclosure and governance rules to AI: material AI risks must be disclosed to investors, boards are expected to oversee AI use, and the agency has sharpened its scrutiny of “AI washing” — companies overstating their AI capabilities or understating their AI risks. The practical effect is that AI governance is already a board-level issue, with disclosure exposure attached.
Underneath all of it sits the patchwork: federal enforcement through existing authorities, state momentum across privacy, employment, and insurance, and no single source of truth. But the fragments rhyme. Bias testing. Transparency. Documentation. Accountability. The specific statutes differ; the expectations are converging.
Move Before the Mandates Do
That convergence is the strategic insight. Organizations that chase each new law individually will always be behind, because there will always be a new law. The organizations getting ahead are building capabilities that anticipate where every jurisdiction is heading: an AI inventory before a regulator demands one, bias testing before it becomes mandatory in their sector, AI governance embedded into existing control frameworks rather than bolted on afterward.
For internal audit specifically, a practical starting sequence looks like this. Establish the inventory — you cannot audit systems you don’t know exist. Assess data governance as the prerequisite it is. Map which AI regulations already touch your jurisdictions and use cases. Then bring AI into the audit universe as a standing risk domain with recurring coverage, not a one-time special project.
AI regulatory risk stopped being a question of “if” some time ago. The only variable left is how quickly the expectations become enforceable where you operate — and whether your assurance function closes the gap before a regulator, a plaintiff, or a headline does it for you.
