Forty years ago I was a sheet metal apprentice climbing around commercial ductwork. Ten years ago I was explaining to a client why their Exchange server needed a reboot. Last month I built a locked, camera’d, leash-only enclosure for a piece of software.

Progress.

I’ve spent my whole career being the guy people call when something doesn’t work — the Answer Man. This year, for the first time, one of my “employees” doesn’t have a face, doesn’t sleep, and can read faster than my entire helpdesk combined. I’m talking about AI. Specifically, about the moment a client asked me to point one at their file server, and I had to decide whether that made me an innovator or an idiot.

The Ask: “Can the AI just… read our files? Answer questions about our documents?”

Sure. Absolutely. Also — over my dead body is it touching anything else on that network.

The Answer: You don’t let a new dog have the run of the house on day one. You get a crate, you pick one room, you watch it for a while. So that’s what we did. Instead of pointing an AI assistant at a live production network and hoping for the best, we built it its own house — a dedicated little server sitting alone in its own segment of the network, with exactly three doors out: one to the file share it’s allowed to read, one out to the internet for the AI itself, and one to check in with the domain so it knows who’s allowed to ask it things. Every other door in that building? Locked, alarmed, drywalled over. Default answer to “can this thing talk to X” is no, unless we explicitly said yes.

It only reads. It never writes, never deletes, never “helps out” by reorganizing anything. And nobody logs into it directly to fiddle around — it’s managed the boring, safe way, the same tools we use to watch over a hundred other servers we’ve never once had to remote-desktop into by hand.

The technical meat, for the nerds in the audience: dedicated DMZ VLAN on the firewall, default-deny egress (nothing gets out unless there’s a rule that says it can), least-privilege file share access, domain authentication without domain-wide trust, and remote management over an outbound-only tunnel instead of an open inbound port. It’s the same paranoid checklist I’d use hardening any server that talks to the internet — I just happen to be pointing it at a piece of software that can technically hold a conversation.

Here’s the thing nobody tells you about “diving into AI” when you’re 58 and better known for torque specs than transformer models: the tech isn’t the hard part. Claude and Grok will hand you working code before you finish your coffee. The hard part is the same as it’s always been in my business — trust, but verify, and never give anything more access than the job strictly requires. Forty years in the trades taught me that lesson with box cutters and bandsaws. Turns out it applies to robots too.

This is Part 1 of a series. Next up: the actual firewall rules that keep this thing in its crate, and then — because a client asked me this exact question after a scary headline — whether an AI like this could ever “escape.”

— Adam