Unsanctioned AI use outruns monitoring by better than two to one in Mindgard's survey

Add the two halves of Mindgard's finding together and 78% of respondents had either evidence or a suspicion that their own staff were using AI without approval ^[1]. Set that against the share whose organisations actively monitored usage, and reported suspicion runs about two and a half times ahead of measurement ^[2]. Sharaf's column does not describe how the sample was drawn, so the figures describe that survey's respondents and nothing wider ^[1].

Husein Sharaf wrote that staff use public tools to draft grant proposals, summarize constituent correspondence, review procurement language and sharpen negotiation memos ^[4]. The same habit reaches a curriculum refined over several years or research data held back ahead of publication ^[5]. The material at risk, in other words, is the material people think hardest about. Much of that activity sits outside the institution's contracts, identity controls and audit logs ^[14].

Public and consumer AI services can give providers broad rights to use prompts or outputs, depending on the product, the account type and the settings, while enterprise agreements commonly provide stronger confidentiality and data-use protections ^[6]. An employee working from a personal account may therefore place institutional information under terms leadership never reviewed ^[18]. A written policy leaves that gap open: it "cannot provide the tool employees need at 4:45 p.m., when a council briefing, grant submission or board packet is due," Sharaf wrote ^[16].

Sharaf leads Cloudforce, a firm in the DC metro area recently named Microsoft's global Education Partner of the Year ^[7]. What he recommends is broad access to approved AI services inside an environment covered by enterprise terms, with identity management, usage review and documented provider handling of prompts and retained data ^[8]. The commercial interest is plain, and the diagnostic is separable from it: anything that comes back missing from the four requests identifies a specific governance gap ^[10].

Blocking changes what security teams can see before it changes what staff do. Usage shifts into places where those teams have fewer logs and fewer practical controls, according to Sharaf ^[12]. He keeps a role for bans on specific high-risk services, and says a ban works best when employees already have a useful, approved option ^[13].

Provisioning early carries its own cost: the institution signs enterprise terms before the review that was meant to test them, and standardises on a provider picked against a deadline. Sharaf's claim is that the alternative cost is already being paid, because an institution without approved, enterprise-wide AI access "remains exposed while its evaluation continues" ^[15]. Of the staff doing it, he wrote: "They are doing their jobs with the tools available to them. Leadership decides which tools are available." ^[5]