Most AI policies are a list of banned tools and a sentence telling people not to paste confidential information. Both parts fail, and the research says how.
Why the usual policy does not work
The overwhelming majority of pasting into AI tools already happens on personal accounts. A policy that blocks corporate access does not touch it — it only removes your ability to see it.82% — the share of those pastes made from unmanaged, personal accounts — outside any enterprise oversight. LayerX, Enterprise AI & SaaS Data Security Report
And it is not occasional. At this frequency you are not writing a rule about an exception; you are writing a rule about how people now work.14 a day — average pastes into non-corporate accounts per employee, of which at least 3 contain sensitive data. LayerX, Enterprise AI & SaaS Data Security Report
The second failure is subtler. “Do not paste confidential information” asks every member of staff to make a legal judgement about a document, in a hurry, with no training. Most people know a customer name is sensitive. Fewer stop at a case reference, a postcode, or the combination of three ordinary facts that identifies somebody.
A rule that makes the work harder does not stop the work.
The four decisions
A policy is not wording. It is four decisions, and the wording follows from them.
- •What may go in. Not which tools — which categories of information. “Nothing that identifies a client, a patient or an employee” is enforceable and understandable. A list of tool names is out of date by the next quarter.
- •Which accounts. A sanctioned tool on a corporate account is visible and governed; the same tool on a personal one is neither. If you approve nothing, you have chosen the second.
- •What happens to an honest mistake. If disclosing one is a disciplinary matter, nobody will disclose one, and you will find out from the customer instead.
- •Who decides the edge cases, and how fast. A policy with no fast route to an answer is a policy people resolve for themselves.
What to write instead
Two clauses do more than a page of prohibitions.
- •A sanctioned tool that is genuinely good enough. The alternative to a good approved tool is not abstinence, it is a personal account. This clause is a budget decision wearing a policy's clothes.
- •A requirement that identifying detail is removed before anything is sent — and a tool that does it, rather than an instruction to do it by hand. Asking people to redact manually is asking them to be reliable at the fifth account number on page nine.
Test it before you publish it
Take a real piece of work — a complaint, a payroll query, a case note — and try to do it under the policy you have written. If the compliant route is slower than the non-compliant one, you have written the thing people will ignore, and you have learned that cheaply.
Try Secure AI free
Frontier AI chat — encrypted by default, with your identity stripped before any request reaches the provider.
Common questions
Should we ban ChatGPT at work?
Blocking it on corporate devices moves the behaviour to personal ones rather than ending it — 82% of pasting into AI tools already happens on unmanaged accounts. The ban costs you visibility and buys less than it appears to. Sanctioning something adequate works better than prohibiting everything.
What should an AI policy actually say?
Which categories of information may never be sent, which accounts may be used, what happens when somebody makes an honest mistake, and who answers edge cases quickly. Naming tools dates the policy within a quarter; naming categories does not.
Is 'do not paste confidential information' enough?
No, because it delegates a legal judgement to everybody in the organisation, made in a hurry. People reliably spot a name and unreliably spot a case reference, a postcode or a combination of ordinary details that identifies somebody. The instruction needs a tool behind it.