What I saw about different approaches to AI rules this fortnight
Earlier this fortnight I ran a Claude training session for an executive team, and one setting stopped the room.
Earlier this fortnight I ran a Claude training session for an executive team, and one setting stopped the room.
I was showing them how memory works. Claude can remember how you like things done, and you can open it up, see what it has saved and edit it. The head of technology put his hand up. Memory from chat history had been switched off in their account as a security control. He could see the benefit of having it on, and he wondered why they had been advised to turn it off.
I said they should ask their cyber team for the reasoning. He told me where it came from. A senior leader had asked Claude what should be switched on and off from a security point of view, and the setting went off on that advice. By the end of the conversation he said he would probably switch it back on.
I kept thinking about that moment, because I had seen a few other organisations’ AI rules up close in the same two weeks. They were drawn in different places. I’ve started sorting them into three kinds.
Ban, toggle or data rule
A ban puts the line at the tool. “We don’t use AI for anything sensitive.” “Nobody puts client information into Claude.” It’s a rule about the product.
A toggle puts the line at a setting. Memory, connectors, file uploads or chat history get switched off as a precaution. It’s a rule about a button.
A data rule puts the line at the data. “No clinical or personal records go in. Everything else can.” It’s a rule about what the information is, whichever tool it ends up in.
All three feel like governance. Only a data rule can be tested against real work. Nobody can check a ban, because the work either happens outside it or doesn’t happen. A toggle can’t be checked either if nobody wrote down which risk it was switched off to remove. With a data rule, someone can look at a piece of information and say which side it’s on.
The toggle nobody could explain
The memory setting was a toggle. Someone switched it off as a precaution, and the person responsible for the technology couldn’t tell me what it was protecting against. That’s how toggles tend to go. The person who made the call moves on, or the advice came from somewhere nobody can trace, and everyone after them treats it as a fact about the product.
I’ve seen the ban version at a webinar I ran earlier this year for around 320 people. Confidentiality questions kept coming back in different forms. People weren’t asking for permission to do something risky. They were asking where the line was, because nobody had told them.
The rule people could use
The same organisation also has a data rule. Before it rolled Claude out to a small group, its head of technology ran a security test on the platform. The test found that the data hosting didn’t meet the organisation’s legal obligations for clinical and personal records.
The easy response would have been to ban the tool. They wrote a rule about the data instead. No clinical or personal records go into Claude. Everything else is allowed.
That one sentence changed what the rollout was. People on the account could check before they typed. An acquisition memo, a position description, a summary of external provider proposals pulled from SharePoint: all of it sat on the right side of the line, and nobody had to ask. In the session, people also worked out how to de-identify what sat on the wrong side, like a generic review of a fall or a template warning letter.
A large consumer brand did a version of this with us the same fortnight. It re-reviewed one of our tools, and its questions were about the data. What does the tool hold? It holds a first name, a mobile number and an optional business name. What does it connect to? Nothing inside the brand’s own systems. Who provides each part? The review passed with small changes, including two-factor sign-in.
Where I’m still working it out
A data rule takes work. Someone has to list what the organisation holds, group it and decide which groups are off limits. Most organisations haven’t done that, and a ban or a toggle is quicker to write.
I also don’t know yet how well a data rule holds up. The rule I described is a few weeks old, and the edges are where it gets tested. In the same executive session, a clinical governance lead asked Claude to summarise some reports that held no personal information. Claude declined to summarise part of one, because a home and a date together could identify someone. She pushed back and asked how it knew that if it hadn’t read it, and Claude walked part of it back.
Both of them were right in a way. A date and a home name aren’t personal records on their own, and together they can point to a person. I’d like to see what happens when that kind of document turns up a few hundred times, and whether anyone reviews the rule when the tools change.