Your AI Governance Has a Kill Switch Problem Nobody Wants to Name

Most organizations have built AI governance the same way they built compliance programs in the 1990s: lots of documentation, clear ownership of the "yes," and almost no one owning the "stop." Ask your board who authorized your latest AI deployment. You'll get five hands. Ask who is empowered to shut it down if it causes harm. You'll get silence. That silence, as MIT Sloan Management Review puts it, is the most important answer you'll receive about your AI governance maturity in 2025.

Here is the self-contained finding that boards need to hear: despite widespread claims of active AI governance, MIT Sloan Management Review identifies the persistent absence of a named individual with authority to decommission a harmful AI model as the gap that transforms governance from a risk-management function into a reputational liability.

Why does AI governance fail even when policies exist?

The structure of most AI governance frameworks mirrors the org chart that built them: siloed by function, optimized for deployment velocity, and allergic to accountability for failure. As HR Executive recently observed, AI readiness requires HR leaders to actively resist letting existing org-chart logic dictate AI strategy — because the org chart was designed for a world where failure modes were slower and more visible. When an AI hiring filter quietly screens out qualified candidates for months, the org chart doesn't surface the problem; it distributes the blame until no one owns it.

The Google case is instructive here. HR Executive reported that Google insiders were circulating guidance advising job applicants on how to sidestep the company's own AI screening tools. That's not a technology malfunction. That's a governance failure that metastasized into a culture problem. When your own people are coaching candidates around your AI filters, you don't have an adoption problem; you have a legitimacy crisis.

AI governance, stripped to its function, is the set of policies, roles, and decision rights that determine how artificial intelligence systems are built, deployed, monitored, and — critically — stopped. The field-specific term to understand is "model accountability": the assignment of a named human decision-maker who holds ongoing responsibility for a deployed model's outputs and who retains the explicit authority to halt it. Without model accountability assigned to a specific role, governance is a document, not a function.

What should the 'stop' authority actually look like in practice?

The answer is not a committee. Committees are where accountability goes to die. The MIT Sloan Management Review analysis argues that the right question to force inside any organization is brutally specific: who, by name and title, can shut this model down today, without a committee vote, if it is causing harm? If that question produces hesitation longer than ten seconds, you have your answer about your governance posture.

This matters beyond the ethical dimension. As Ana White, chief people officer at Lumen, writes in HR Executive, scaling AI is less about deploying tools and more about building the conditions for people to use them well. That cuts both ways: the conditions for people to stop using AI badly are just as foundational as the conditions for adoption. A CPO who can champion deployment but cannot champion cessation is operating with one hand tied behind their back.

The operational risk compounds when you consider the hiring pipeline specifically. HR Executive's reporting on workforce data — drawing on analysis from Peter Miscovich, global future of work leader at JLL — points to mixed hiring numbers at AI-mature organizations as a signal of deliberate strategic choices about which roles AI augments versus which it replaces. Those choices require human judgment at every step. If the governance layer overseeing that judgment has no named "stop" authority, the strategy is being executed without a brake pedal.

The question for your next leadership meeting is not "Are we governing our AI?" Every organization's answer to that is yes. The question is: "Can the person responsible for stopping a harmful AI model in our organization answer an email by close of business today?" If you don't know who that person is, your governance framework is a liability dressed as a policy — and the board should know it.