Companies Push AI Use But Skip Training and Official Policy

* 43% of knowledge workers report their company has not provided formal AI training.
* 26% of knowledge workers do not have a written AI security policy.
* 53% of respondents have access to a company-managed or enterprise AI tool at work.
* 47% of respondents are encouraged or required to use an AI tool for certain tasks.
* Training provided included AI basics (writing prompts/tool use) at 33%, data privacy/cybersecurity at 29%, ethical/compliance guidelines at 29%, and role-specific training at 24%.
* Only 24% of workers report their company has a clear, communicated AI policy.
* Specific attacks like prompt injection are not recognized by most workers; only 26% correctly recognized the risk of exploiting public-facing chatbots.
* 31% of workers with a written policy spotted prompt injection, compared to a general average recognition rate of 26%.
* Only 25% of workers would check with management before copying client contracts into a personal AI account.
* 77% of workers have some level of awareness or understanding of AI security risks.
* Among workers with a clear written policy, 96% feel at least somewhat confident using AI safely, compared to 61% for those with no policy.
* 49% assume their AI vendor is responsible for company security.

The narrative reveals a critical gap between the existence of policy and the actual knowledge required to enforce it, suggesting that governance structures are insufficient without robust, targeted education. The most significant pattern is how confidence in safety is contingent on documentation; policies act as a reliable safeguard when present, suggesting that establishing clear written rules shifts the locus of responsibility away from individual, potentially flawed, judgment toward an organizational standard. This dynamic implies that the absence of formal training and policy creates a vacuum where assumptions about automated safeguards thrive, leaving employees vulnerable to sophisticated, low-effort attacks like prompt injection. The focus on "basics" training over specific vulnerability training reflects a systemic failure to address actual exploitation vectors in the current knowledge economy. A deeper implication is the distribution of risk: while high-level access exists for many, and some enterprise contracts exist, the reliance on unstated assumptions about vendor responsibility means that security liability defaults dangerously onto the individual worker when formal guardrails are absent. The missing link is translating abstract policy into actionable, specific defense skills, rather than simply distributing documents.
Bridge Questions: If policies are only effective when understood, what framework is needed to measure the *effectiveness* of AI training beyond completion rates? How can organizations shift the burden from employee vigilance to systemic architectural controls that inherently prevent high-risk actions? What mechanisms can be established to ensure accountability for vulnerabilities like prompt injection, irrespective of formal policy existence?