The question I get more than any other is some version of is it safe for us to use AI.
I don’t have a good answer because the question can’t be answered as asked. It’s like asking whether a truck is safe. Safe doing what, carrying what, driven by who?
So, here’s the version that does have an answer. Stop asking whether AI is safe for your business. Start asking which of your information it’s allowed to touch. I use colors because people remember colors. I touched on these a few issues back, and they’re what people ask me about most, so let’s do it properly.
Green is public. Anything already out in the world, or anything you wouldn’t care if it was. Website copy. A job posting. Marketing material. The first draft of a proposal built from your own standard language. Real work, real hours, and it can move today.
Most businesses have far more green work than they think, and they’re frozen on it all because they’re worried about the red stuff. That’s the real cost of never sorting this out. It isn’t a breach. It’s a year of standing still while a competitor gets faster.
Amber is internal. Yours, but not dangerous. Standard operating procedures. Project notes. Meeting recaps. Operational numbers with nobody’s name attached. If amber leaked, it would be embarrassing, not a phone call to your attorney. Before AI touches amber, you want a written policy people have actually read, one approved tool instead of whatever everybody found on their own, and somebody who can see what’s being used.
Red is where I need you to slow down, because most people treat red as one bucket, and it’s really two.
The first kind is sensitive. Financials. Employee files. Signed contracts. Internal strategy. Information that would hurt you if it got out, but it’s yours, and you get to decide. Sensitive can go into AI if you build the protections first. Controlled access, so not everyone in the building can reach it. A documented approval, so there’s a record of who said yes. And a person reviewing every output before it goes near a customer.
The second kind is restricted, and restricted doesn’t go in at all. Client production data. Health information. Payment card numbers. System credentials and passwords. Anything covered by a compliance obligation you’ve signed, and anything a client handed you under an NDA. Our own written policy says restricted data is not permitted in any AI tool, and that’s the rule I’d give you. Not be careful with it. It does not go in.
The difference matters because the fix is different. Sensitive is a control problem, and you can go build controls. Restricted is a contract or a regulator, and being careful doesn’t get you out of it.
When you can’t decide which class something falls in, take the stricter one. You lose a day. The other mistake costs you a client.
One thing cuts across all three. The account matters as much as the data. Nothing belonging to your business goes into a personal consumer account. Not the free tier, not somebody’s personal subscription. Those are consumer products on consumer terms. No administrator can see what’s in them, and when that person leaves, everything they put in there goes with them.
Business plans work differently. Your content isn’t used to train the model by default; you have a real administrator, and you can see what’s happening. That’s the line between a tool you can put company work into and one you can’t. We build on Claude, and I’ll say that plainly so you know where my bias sits. The team is the floor for a business, and regulated shops go to Enterprise for audit logs. Pick something else if it fits you better. Just get the business version.
Now the part almost everybody gets wrong.
Don’t classify your company. Classify the workflow.
Owners keep trying to answer whether they’re a green business or a red business, and there’s no answer to that either. Everybody is mixed. Your marketing is green, and your HR files are red, same building, same week. A contractor’s boilerplate bid language is green, and the estimate with the client’s numbers in it is sensitive. A wealth manager’s market commentary is green, and anything with a client name on it is restricted.
So don’t think in departments. Write down your ten most repetitive workflows, the ones that eat time every week, and put a color next to each. That’s the whole exercise.
What comes out usually surprises people. Half the list is green and can start Monday. A few are amber, waiting on a policy you don’t have yet. One or two are red, and now you know which you can build toward and which is never going in at all.
One last rule, and it’s the one underneath all of this. The security work sets the speed limit. It is not something you catch up on later.
I’ve watched it go the other way plenty of times. A tool goes out to the whole team, everybody starts pasting whatever’s on their screen into it, and six months later, a client asks what your AI policy is. Now you don’t have a policy problem. You have a discovery problem, because nobody can tell you what’s already in there.
The old me would have said get the tools working and sort the rules out after. I don’t say that anymore. Move fast on green, take your time on amber, build carefully toward sensitive, and leave restricted alone. That isn’t slowing you down. It’s the thing that lets you keep going.
So that’s your hour this week. Ten workflows, three colors. Nothing to install, nothing to sign.
My guess is you’ll find you’re allowed to start on more than you thought.
That’s all. Just something to sit with.
