Running a disclosure amnesty

An amnesty is the first promise put to work. You tell everyone what you're asking and why, you ask everyone at once, and you check what they tell you against two kinds of records. Done well, it gives you the first version of your AI list within a week, and it tells you something no survey can: whether people believe the promises.

People decide before the first question

Whether someone tells you the truth is settled before they see the survey. They decide from the announcement, from who makes it, and from what happened the last time someone admitted a shortcut. The reasons to stay quiet are real, and they're not only about rules.

The last number is the one to remember when you write the announcement. Even the people who write the rules do it. An amnesty isn't about catching the careless few. It's about finding out what nearly everyone is doing.

Announce it under the promises

The announcement does four things.

• It says why. You can't improve work you can't see, and the people furthest ahead know things everyone else needs.
• It restates the promises, in the words your sponsor signed in module 1.
• It says where the answers go: to the disclosure custodian, outside the management chain, with leaders seeing totals only.
• It says what happens next: anything worth adopting is adopted, and its maker gets the credit.

Make it the sponsor's announcement. An amnesty announced by the people who would have enforced a ban reads like a trap, however well it's worded.

Ask everyone, in the room

The survey runs at the baseline session, the one-hour session for everyone that opens the loop. Its last ten minutes are for the survey, filled in there and then. That matters more than it sounds. A survey sent by email gets answered by the people with least to hide. A survey filled in by the whole room, straight after the aim and the promises, gets answered by nearly everyone.

Keep it short. Four core questions find what the AI list needs: which tools, what for, what data goes in, and whether anything someone made is now relied on by others. Add a question only if you'll act on the answer.

Three sources, because one isn't enough

The survey is one source. The pilot kit starts the AI list from three, and each catches what the others miss.

1. The amnesty survey: what people tell you. It's the only source that finds free tools, which leave no record anywhere.
2. SSO and expense records: what people sign in to with work accounts, and what's been paid for on cards and expense claims. A personal subscription claimed back every month is one of the most common finds.
3. A check of software already in use: AI features vendors have switched on inside tools you already pay for, often in an update nobody read. Video calls, email, the CRM, the help desk and document tools are the usual places. This is AI that nobody chose at all.

Who sees what

Individual answers go to the disclosure custodian you named on your aim page. Leaders see totals: how many tools, how many people use AI for drafting, how many have put confidential data into something. Never who.

Monitoring follows the same rule. It counts tools, not people: fourteen people sign in to this note-taker, not a list of who they are. The framework borrows this firewall from aviation, where a confidential reporting system has protected pilots who report their own mistakes for fifty years, and the reports keep coming because of it.

There is one exception. If a disclosure or the records show that regulated data has been exposed, someone has to follow up with the person: to find out what went where, and to contain it. That follow-up runs through stop, tell, fix, learn, under the firewall, as containment and coaching. The custodian handles it, not the person's manager, and it's still never used in reviews or decisions about who leaves.

Disclosure against the records

With all three sources in, you can ask the question that tells you whether the promises are working: of the AI tools the records found, how many did people also disclose?

The pilot kit takes this measure on day 5 and again on day 90. If disclosure covers less than half of what the records found, the readout says, in its first line, that the promises didn't make disclosure safe. The aim is for the two to match, and wherever they don't, the gap shows you where trust is weakest.

Small organisations often have no monitoring at all. That's fine. Check the list against expense and single sign-on records, which every organisation has in some form.