I got pitched an AI platform last week that was going to "revolutionize network operations." The deck had forty-one slides. Slide nine had a photo of a glowing brain made out of fiber optic strands. Slide twenty-two promised a 340% efficiency gain. Slide thirty-eight was the pricing, and it was not a small number.
I asked one question. What happens when a backhoe cuts a line in Pacoima at 2 a.m.?
The guy smiled and said the platform would "surface the incident with predictive context."
Cool. So it will tell me the thing my alarms already told me forty seconds after it happened, and then somebody still has to get in a truck.
I'm Not an AI Skeptic. I'm an AI Realist.
Let me get this out of the way, because people hear one complaint and decide you're a Luddite yelling at a server rack. I use this stuff every day. It writes code with me. It drafts things I don't want to draft. It reads contracts faster than I do and it is genuinely, stupidly good at the boring middle of almost any task.
That's not the argument. The argument is that we've entered a phase where "we added AI" has replaced "we fixed the thing," and those are not the same sentence. One takes engineering. The other takes a press release.
Telecom is a beautiful place to watch this happen, because telecom is physical. You cannot vibe your way through a conduit. The fiber is in the ground or it isn't. The radio has power or it doesn't. The splice is clean or the light doesn't make it to the other end. Physics does not care about your roadmap.
You can automate the ticket. You cannot automate the shovel.
The Part Everyone Skips.
Every AI pitch I've sat through assumes the data is clean. That's the whole trick. The model is only as good as what you feed it, and what most companies have to feed it is fifteen years of inventory records maintained by four different teams using three different naming conventions, two of which involved somebody's initials.
I've seen a circuit labeled three ways in the same system. I've seen a site marked decommissioned that was actively carrying traffic. I've seen an address field with a sticky-note's worth of prose in it. Point a language model at that and you don't get insight. You get a confident, well-written lie.
That's the thing that worries me more than anything else — not that AI is wrong, but that it's wrong in complete sentences. A bad spreadsheet looks bad. A bad AI answer looks like the best answer in the room.
If your data is garbage, AI doesn't clean it. It launders it.
The Real Work Is Unsexy and Nobody Funds It.
Want to know what actually moved the needle for us? Not a model. A naming standard. A single source of truth for what we own and where it is. Somebody sitting down and reconciling records by hand until they matched reality.
No one gets a keynote for that. There's no glowing brain slide for "we finally agreed what to call things." But it is the entire foundation of everything the vendors want to sell you on top of it, and they'd rather you skip it, because selling the foundation means admitting the fun part comes later.
I'd take a boring, correct database over a brilliant model any day of the week. The database compounds. The model on top of bad data just makes your mistakes faster and harder to trace.
What AI Is Genuinely Great At.
I'm not going to do the whole grumpy-CEO routine without being fair, so here's my actual list.
It's great at the first draft of anything. It's great at reading a 60-page interconnection agreement and telling me which clause is going to bite me. It's great at pattern-matching across thousands of alarms to say hey, these nine things are probably one thing. It's great at writing the script I need once and never again. It's great at being the coworker you can ask a dumb question at 11 p.m.
Notice what those have in common. Every one of them is leverage on a human who already knows what they're doing. Not a replacement for one. The value isn't the model — it's the person holding it who can tell when the answer smells wrong.
AI is a power tool. Power tools in the hands of somebody who doesn't know what they're building just get you to the mistake faster.
The Bubble Isn't the Technology.
People keep asking me if it's a bubble. Sure, parts of it. But the bubble isn't the technology — the technology is real and it's not going anywhere. The bubble is the layer of companies that are a thin wrapper around somebody else's model, charging enterprise prices for a feature that'll be a checkbox in eighteen months.
That's the part that pops. And when it does, everybody's going to act like the whole thing was a fad, which will be just as dumb as pretending the forty-one-slide deck was a strategy.
We did this with the cloud. We did it with 5G — I watched an entire industry promise holograms and deliver, mostly, slightly better download speeds. We did it with blockchain, which somehow got proposed to me for tracking truck rolls. Every cycle, the same shape: wild promise, quiet disappointment, then a slow boring adoption where the actual value shows up years after everyone stopped writing about it.
So Here's Where I Land.
Show me the failure mode. That's my only real test now. Don't show me the happy path in a controlled demo with your cleanest data — show me what happens when the input is wrong, when the API times out, when the field tech types the address in wrong. Tell me who's accountable when it confidently makes something up.
If you can answer that, I'm interested and I'll pay for it. If your answer is a 340% efficiency gain and a glowing brain, I'm going to let you finish the deck out of politeness and then go back to fixing my data.
Because the backhoe is still coming. It always is. And when it does, the only thing between a customer and an outage is a human being in a truck who knows exactly where the line is.
Get that part right first. Then bring me the brain.