The formal AI adoption process at most large organisations is careful and slow. A committee, a security review, a legal opinion on the terms, a pilot with a defined scope, a procurement cycle. It is a reasonable process and it takes the better part of a year.

The informal adoption took about a fortnight and happened two years ago. Staff who found a tool useful started using it on personal accounts, on personal devices where necessary, and told nobody — not out of defiance but because the tool made a Tuesday easier and asking would have produced a form.

A ban relocates the problem and removes the evidence

The instinct is prohibition, and organisations that tried it report the same outcome. Usage does not stop; it moves off the corporate network and onto a phone, where there is no logging, no retention policy and no ability to establish afterwards what was shared. The company trades a visible risk for an invisible one and records the change as a control.

What was actually being shared is the part worth establishing before deciding anything. In most audits it is not the dramatic material. It is a customer email pasted in to be rewritten more politely, a contract clause someone wanted explained, a spreadsheet fragment, a paragraph of code with an internal hostname in it. Individually trivial, and collectively a fairly complete picture of how the business works, sitting in an account the company does not control and cannot subpoena.

This is where residency stops being abstract. A company that has spent real money establishing where its data sits as an architecture problem has that whole effort undone by an employee pasting a customer record into a consumer service, and the contract clause promising regional processing is not false so much as irrelevant to what actually happened.

The response that works is unglamorous and follows the pattern of every other shadow IT wave: provide a sanctioned option good enough that the unsanctioned one is not worth the friction, then make usage visible rather than forbidden. Visibility is the harder half, and it is the reason the browser has become the place security actually happens — the paste is the event worth seeing, and it is only observable at the tab.

The procurement side is catching up, slowly. Corporate buyers pushing for standard terms in model contracts are asking exactly the right questions about training use, retention and indemnity. They are asking them about the tool the company is evaluating, while the staff are using a different one, and closing that gap is mostly a matter of admitting it exists.

Topics aigovernancesecurityworkplace

Technology Correspondent

Priya Natarajan

Priya Natarajan reports on artificial intelligence, enterprise software and the infrastructure behind the modern internet. Her work focuses on how technical decisions become business decisions.