Why AI governance is failing — and what actually works

AI adoption has outpaced AI governance, with 65% of organizations reporting at least one AI agent-related incident in the past year, according to Cloud Security Alliance research. Nearly half of organizations have tied confirmed or suspected data leaks to unauthorized generative AI use, per EY's Technology Pulse Poll. The gap is not a lack of policies but a lack of enforceable controls, with 82% of enterprises discovering shadow AI agents despite high confidence in visibility.

AI adoption has outrun AI governance, and the consequences are impacting the business. According to Cloud Security Alliance CSA research https://cloudsecurityalliance.org/press-releases/2026/04/21/new-cloud-security-alliance-survey-reveals-82-of-enterprises-have-unknown-ai-agents-in-their-environments , 65% of organizations report having experienced at least one AI agent-related incident in the past year. Nearly half have tied confirmed or suspected data leaks to unauthorized gen AI use, per EY’s Technology Pulse Poll https://www.ey.com/en us/newsroom/2026/03/ey-survey-autonomous-ai-adoption-surges-at-tech-companies-as-oversight-falls-behind . The gap https://www.cio.com/article/4028154/ai-governance-gaps-why-enterprise-readiness-still-lags-behind-innovation.html is no longer a lack of policies; it’s a lack of controls that work. And the fix isn’t more documentation; it’s minimum viable governance focused on what actually matters. “Many companies have AI activity, some have AI principles, fewer have enforceable AI controls, and fewer still have evidence that those controls work,” says Sara Jodka https://www.dickinson-wright.com/our-people/sara-h-jodka , an attorney at Dickinson Wright who advises clients on AI governance. Gartner analyst Lauren Kornutick https://www.linkedin.com/in/lkornutick/ sees this frequently with her clients, warning that retrofitting governance is harder than building it in. “The biggest issue I am observing with governance after deployments is that it is really hard to walk back previous decisions,” she says. “If an organization was previously very relaxed in their AI use and an incident occurs, it’s much more challenging to decommission tools or models that early adopters were accustomed to using.” You can’t govern what you can’t see. And most organizations can’t see as much of their AI portfolio https://www.csoonline.com/article/4157486/cisos-tackle-the-ai-visibility-gap.html as they think they can. More than two-thirds 68% of CSA survey respondents expressed high confidence in their visibility into AI agents, but 82% also reported discovering shadow AI agents in the previous year. Hillary Baron https://cloudsecurityalliance.org/about/csa-staff , AVP of research at CSA, calls this a blind spot masquerading as self-assurance. “Organizations have strong visibility into the agents they know about, and it’s easy to mistake that for seeing everything,” Baron says. “Good insight into the known set quietly becomes a belief that there’s nothing else out there.” Meanwhile, 78% of technology leaders say AI adoption is outpacing their ability to audit or monitor systems, according to EY. And 52% say department-level AI initiatives operate without formal approval or oversight. Shadow AI https://www.cio.com/article/4178359/why-your-most-ai-savvy-employees-are-driving-shadow-ai.html has evolved beyond employees signing up for ChatGPT. “It now includes browser extensions, embedded SaaS features, code assistants, meeting transcription tools, copilots, and agentic workflows,” Jodka notes. At payments and data company Deluxe, CTDO Yogaraj Jayaprakasam https://www.linkedin.com/in/yogaraj/ took a different approach: Flood the zone with sanctioned tools before shadow AI could take root. “Shadow AI isn’t a technology failure; it’s a governance vacuum,” he says. “So, we deployed broadly on purpose, to open a sanctioned lane before the unsanctioned ones hardened.” Even when AI is visible, it’s often unclear who’s responsible for governing it. Roughly a quarter of CISOs fully own AI governance, while more than half co-lead with another function, according to Kornutick. “CISOs have the technology and technical expertise to address the runtime inspection and policy enforcement requirements for AI governance,” she says, “but they frequently lack the expertise to apply guardrails with context or make the decisions about what AI should be used and for what purpose.” Fragmented ownership https://www.cio.com/article/4162949/cios-struggle-to-find-clarity-in-their-organizations-ai-strategies.html is a problem, according to Dickinson Wright’s Jodka. Legal-only ownership can set standards but can’t enforce controls in code. IT-only ownership fails because AI governance spans privacy, discrimination, IP, cybersecurity, and regulatory risk. Committee-only ownership “often becomes slow and performative,” Jodka says. She recommends a three-line model: builders own the systems they deploy; legal, security, and compliance set standards and review high-risk uses; and internal audit tests whether the program actually works. The coordination challenge is real, agrees Aslam Rawoof https://www.beneschlaw.com/people/aslam-a-rawoof.html , a partner at Benesch Law. “AI cuts across all facets of the organization. It can’t be owned by tech or legal alone,” he says. “You literally need a committee that meets regularly — legal, tech, finance — and reports up to the CEO if not the board.” Governance tends to focus on the front end: approving tools and writing acceptable-use policies. But the back end — knowing when to turn something off — gets far less attention. In fact, per CSA, only 21% of organizations have formal decommissioning processes for AI agents, and the associated risks compound over time. “An un-decommissioned agent isn’t dormant,” CSA’s Baron warns. “It is still holding live credentials and standing access, but now no one is watching it anymore. That’s what makes it more dangerous than a flaw at launch.” CSA calls this “retirement debt,” and it accumulates invisibly. Every agent stood up and never taken down quietly widens the attack surface. Agentic AI raises the stakes further. When AI was a chat window, the governance question was which tool to buy. When AI becomes an agent taking actions inside systems, the question becomes who is accountable for what it does https://www.cio.com/article/4183249/cios-plagued-by-a-growing-ai-accountability-gap.html . “That crosses IT, the business unit, legal, and risk all at once,” Jayaprakasam says. “Most org charts weren’t drawn for that.” Most governance models were designed for a slower velocity. “Even where governance exists, the survey shows it’s built around periodic review and human approval at decision points rather than continuous, real-time enforcement,” Baron says. “This is a model designed for human-paced work applied to systems that act autonomously and at machine speed.” Perfection isn’t the goal — prioritization is. “CIOs don’t have to excel in every aspect of governance, just excel at governing non-negotiables like high-risk use cases, transformative use cases, or security and privacy,” Gartner’s Kornutick says. CSA’s Baron agrees: “The agents that combine broad access with high autonomy are where your strongest controls and attention should go.” Beyond prioritization, a few things distinguish AI governance that delivers: The aim, Kornutick adds, is alignment — “everyone rowing the boat in the same direction” — where governance is viewed as a business enabler, not a stop gate. Gartner calls this “Return on Integrity”: making investments in governance that enable the business rather than slow it down.