Your Staff Are Already Using AI at Work. The Question Is Whether You Know Which AI

Long before most organisations finished writing an AI policy, their employees had adopted AI. The tools are free or cheap, a browser tab away, and genuinely useful for the daily grind of drafting, summarising, translating and untangling code, so people used them — quietly, individually, without asking. This is shadow AI: capable systems in daily use across a company that the company has not approved, does not monitor, and often does not know about. It arrived the way shadow IT arrived before it, from the bottom up, because the sanctioned path was slower than the unsanctioned one. The risk is not the use; it is what gets pasted in. To get a useful answer, an employee feeds the tool the real material — a customer list, a contract, unreleased financials, a chunk of proprietary source code, a patient record — and in doing so sends confidential information to an outside service whose terms, retention and training practices nobody in the company has read. The data may be stored, may be used to improve the provider's systems, and is in any case now outside the organisation's control. A single well-meaning paste can put regulated or competitively sensitive information somewhere it was never allowed to go, and because it happened in a personal tab, there is no log and no way to know it occurred. The instinctive response — block everything — reliably fails, and fails in the worst way. A hard ban does not remove the underlying need that drove people to the tools; it removes the visibility. Usage moves to personal phones and home accounts where the company can neither see nor protect it, and the organisation trades a manageable problem for an invisible one. The productivity the tools provide is real enough that people will route around a prohibition, so a prohibition mostly guarantees that the sensitive pastes keep happening where no one is watching. What actually works is to make the safe path the easy path. That means providing a sanctioned tool — one with an enterprise agreement that contractually forbids training on your data and controls retention — and making it at least as convenient as the shadow alternative, because convenience is what drove the shadow use in the first place. It means a clear, short, human rule about what may and may not be put into any AI tool, taught rather than buried in a policy document. And it means accepting that the goal is not zero AI use, which is neither achievable nor desirable, but visible, governed AI use. The companies handling this well are not the ones with the strictest ban. They are the ones who gave their people a good tool and a clear line, so the shadow version stopped being worth the risk.