She's Retiring in March

Tribal knowledge is undocumented infrastructure.

Every property has a Martha. Maybe she's the night auditor who's closed the books since 2009. Maybe he's the reservations manager who survived three PMS migrations and two flag changes. The title doesn't matter. What matters is what's in their head:

UDF23 is the guest's ETA, except on group blocks, where it's the bus number. Rate code SPCL19 looks abandoned but you can't delete it, because the comp routing for the Henderson account breaks if you do. None of this is in the SOP binder. All of it is load-bearing.

We call this tribal knowledge, which makes it sound quaint. Folklore. Something for the holiday party.

Call it what it is: an undocumented production dependency.

If your property had a server that only one person knew how to reboot, you'd flag it as a critical risk. You'd document the procedure, cross-train a backup, maybe lose some sleep. But the same situation in human form (a system whose meaning exists in exactly one place, complete with a pulse and a pension date) gets a laugh and an "oh, just ask Martha."

Now scale that up to the corporate level. Because at the brand, the server only one person can reboot isn't a metaphor. It's a real platform, it's been running for twenty years, and the person who knows how it's wired is named Terry.

Terry owns the integration layer: the tangle of interfaces, middleware, and shims that moves data between the PMS, the CRS, and forty other things with three-letter acronyms. He didn't build all of it. He's just the only one who still remembers why the nightly job has to run before the rate upload, why the channel-manager mapping carries a hand-coded patch from a 2014 schema change nobody ever cleaned up, and why you never, ever restart the posting service between 2 and 4 a.m. unless you enjoy explaining double-charged folios to a regional VP.

None of that is documented either. All of it is load-bearing too.

So now you have two of them, and probably way more. Martha knows what the data means at the edge. Terry knows how it moves at the core. Together, they are the whole picture.

Before you file Martha and Terry under technical debt: don't. This is just how the industry was built. Hospitality grows by acquisition and carries its technology history with it. Every deal bolts on another PMS, another flag, another tangle nobody reconciles.

Except it was reconciled, just strictly in the minds of a few key people. They became the integration budget the merger never funded, the migration doc nobody wrote, and the data model nobody drew. It's a human business, and it has worked this way for a hundred years.

Now watch what happens next.

Somewhere at HITEC this year, a vendor shows you a slide. Feed your systems to their model, it learns your operation, it answers the questions Martha answers now. The demo is clean. The data on stage is perfectly labeled. For about forty-five minutes, you feel solved.

The model never met Martha or Terry.

The vendor is selling the illusion that Martha's knowledge is a fixed asset you can simply download and cache. It isn't. Martha doesn't just store exceptions; she actively negotiates them against a reality that changes every shift. The tool on the slide is great at applying static rules to clean data. What it can't do is replicate the real-time human judgment required when the Henderson account shows up with an unannounced bus at 11 p.m.

But in the feeding frenzy of AI, nobody remembers the exceptions.

So you buy it, because the slide was clean and the quarter is loud. And it works. Or, it looks like it works. The answers are mostly right, and the wrong ones are wrong quietly, and nobody downstream knows to check.

You find out in the fall. A folio routes wrong, or a block prices wrong, or a number that fed a number that fed a board deck turns out to have been an AI hallucination since spring.

This was never an AI story.

You go to ask Martha.

Martha retired in March. Terry is taking a sabbatical to learn AI.