I Meant to Redact It

I just didn’t do it.

A few days ago I put my own company’s financial information, account numbers and all, into an AI tool without redacting it.

Not a client’s information though, I want to be clear about that, because the difference matters and I would never risk my clients’ information this way.

We were testing an updated reconciliation system we use on our own books. It was deliberate work, on my own account documents, in my own environment. The kind of use I’d tell anybody is completely fine. I had every intention of stripping the real numbers out first.

I just didn’t do it. I ran the test, and everything went up with it. I was in a hurry and excited to see our new system in action. I figured, one test, no biggie.

Then, I sat with that for a minute. I started thinking, why, why did I do that? I tell business owners to strip out account numbers and anything that identifies a person or business before it goes anywhere near a tool. I’ve told them the question was never whether their data goes into AI, because it does. The question is whose AI it goes into, and whether the person doing it made that decision or was just being unaware.

That’s the part I want to talk about, because it explains something the careful version of this conversation never gets to.

When people picture someone leaking company data into AI, they picture carelessness. Somebody who doesn’t take it seriously. Somebody who has no idea what the risk even is. Somebody who needs a policy.

That’s almost never what it is.

It’s somebody who meant to be careful and had one step of it living in their own head. It’s the person who was going to strip the client name out and got interrupted. It’s the person who used the free tool on their phone because getting to the approved one meant a login they couldn’t remember at nine at night. The intention was there, but intention isn’t enough unless it leads to action.

**Intending to be careful isn’t a safeguard. It’s a hope with good manners.**

Which is why the fix I landed on wasn’t a rule.

I could have written myself a note. Redact before testing. I’d have followed it most of the time, and most of the time is exactly the failure mode I’m describing.

So I decided to build a private LLM server in-house instead. Not for everything. Just for financial data and private personal information, which is the small category where I actually care.

The thing that makes it work isn’t the model. It’s that it won’t have internet access. It’ll live on our own network, on an encrypted server, with no route out.

The idea is that the data can’t go anywhere, because there’s nowhere for it to go. Nobody has to remember or think about it. Nobody has to be having a good day or read the rule book again. The risky version of the action stops being available, which is the only kind of safeguard that survives a busy afternoon.

I should say plainly that I’ve decided to build this and haven’t built it yet. When it’s up and running I’ll write about what it actually took, including the parts that were more annoying than I expected and the parts that went well.

A server is one way to handle this, but it’s not the only way, and it’s not the point. The point is moving the safeguard out of somebody’s memory and into the environment.

**Five things you can do this week instead**

1. **Name the one category of data that must never leave.** Just one. Financials, client contracts, health information, whatever it is for you. A list of nine gets ignored.

2. **Ask your team what AI tools they’ve used and what they put into them.** It’s true that most owners have never asked it once or even thought about it for that matter.

3. **Make the safe path the easier route.** If the approved tool takes three extra steps, people will keep using the one already open. Make it easy to follow the rules and follow up on compliance.

4. **Put redaction in the tool, not in your head.** A saved template with the fields already blanked beats remembering every time. Spreadsheets can have identifying data easily stripped away with a small script. PDFs need a real redaction tool, because just using a blocking box usually leaves the text sitting underneath, and it can still be found that way.

5. **Decide now what happens when somebody gets it wrong.** Because you know they will, and the only thing worse than the mistake is a culture where nobody mentions it. Make everyone aware of your wishes and guidelines.

I did it wrong on my own data, that’s on me, but changing the setup instead of writing myself a rule is the whole story.