The OpenAI key rotated on a Thursday morning. Nobody at the dental practice noticed until Monday, when a patient called to ask why nobody had followed up about her root canal consult.
That was the moment the owner learned his office manager had built an AI tool.
Six months earlier, Vanessa, the office manager, had wired up an intake summarizer. She’d hated typing the same morning huddle note three times a day, so she fed the intake forms into ChatGPT, told it to write the summary, and stapled the workflow to the front desk’s morning routine. It worked. The front desk knew about it. She’d never told the owner, because it wasn’t his problem to solve and she’d already solved it.
Yeah. That’s the piece.
For six months the practice ran a process the owner couldn’t see, couldn’t fix, couldn’t replace, and didn’t know existed. When the key rotated, the front desk reverted to handwritten triage and just forgot to mention it. The follow-up calls stopped going out. The patient who didn’t get hers complained to the front desk, who told the office manager, who told the owner, who learned three things in a single conversation: the summarizer existed, it had been load-bearing, and it was now down.
He didn’t have a missing follow-up problem. He had a prototype graveyard problem.
That phrase isn’t mine. The way Nate B Jones put it last week, every operator has a shadow stack of half-real tools sitting in employees’ personal accounts, doing real work that nobody officially knows about. The question used to be should we build it. That question is over. Building is cheap. Every employee with a credit card and an afternoon has built something. The new question is should we keep it, promote it, or delete it, and nobody on your team has a rule for any of those.
Here’s the thing.
That’s not a tools problem. That’s a Station Plan that’s been growing without a Chef.
The Station Plan is how we describe an AI-native business when it’s working. The Chef is the human at the Hub, the one making the calls. The stations on the Line are the agents and automations doing the labor. When all of it sits in the right place, the kitchen cooks. When a station shows up on the Line without the Chef ever putting it there, that station is still on the Line. The dish still ships. The Chef just doesn’t know it exists.
That’s where every prototype graveyard lives. On the Line. Doing real work. Not on anyone’s org chart.
You’ve been letting every employee-built workflow accumulate as we use it sometimes. That isn’t a tool. That’s a load-bearing process with no owner. When it breaks, the business breaks, and nobody knows where to start looking, because nobody ever put it on a list. The office manager wasn’t hiding the summarizer. She didn’t think it counted. ChatGPT and a copy-paste habit don’t feel like infrastructure. They feel like a personal trick that saved her twenty minutes.
But she shared the trick with the front desk. And the front desk built their morning around it. The moment something is shared and depended on, it stops being a personal trick. It’s a station. It just doesn’t know it yet.
Four states, written down
Here’s where it gets practical. The way to drag these shadow tools out of the graveyard and onto a list is to classify each one, in writing, into one of four states. This isn’t our framework. It’s Nate’s, and we’re keeping it intact.
Personal tool. One person uses it for one job. The receptionist has a ChatGPT prompt she pastes in every morning to summarize the inbox. If she’s gone, nothing breaks. If the tool dies, only she notices. This is the right place for a lot of AI. The trouble starts when it stops being personal.
Team beta. A few people use it. No owner. No SLA. No fallback when it breaks. The intake summarizer lives here. The front desk depends on it. Nobody monitors it. Nobody knows what to do when the key rotates.
Supported internal product. Someone owns it. It has a fallback. When it breaks, there’s a name on the wall. The summarizer becomes this when the owner says yes, this is a thing we run, Vanessa owns it, the budget is mine, here’s what we do when it goes down. Same tool. Totally different posture.
Customer-facing product. It touches the work the patient sees. Now it isn’t a tool, it’s part of the practice. The bar is higher. If it dies, it can’t die quietly.
Right?
Every employee-built workflow in your shop sits in exactly one of those four states today, whether you’ve labeled it or not. The reason the office manager’s tool blew up the practice isn’t that she built it. It’s that the practice never moved it from team beta to supported internal product. It earned the upgrade and never got it.
Most owners hear this and assume the move is to lock down employee building. Don’t. That’s the wrong instinct. The shadow tools aren’t the problem. The lack of a classification rule is the problem. Kill the building and you’ve killed the cheapest R&D you have.
The ops manager already knows where they are
Here’s the thing the ops manager already knows. If you’re the one running the practice from the inside, you already know where every one of these tools is. You don’t need to find them. You need permission to surface them. The owner has been pretending none of this is happening, and you’ve been pretending you don’t know which automations are now load-bearing. Both of you have been right about the discomfort and wrong about the cost. Every shadow tool that survives one quarter unannounced graduates to team beta by accident. Wait two quarters and you’ve built a second practice inside the first one, breaking on a different schedule and answering to no one.
The first move isn’t governance. It’s a list.
Write down every AI tool or automation that has been built or installed in the last six months. Every prompt the front desk pastes into ChatGPT. Every Zapier flow somebody’s been running on a personal account. Every Make.com scenario, every custom GPT, every Claude project. Literally every one. Most practices land somewhere between twelve and forty items on this list, and that number, by itself, is the diagnosis. You did not know you had that many. Now you do.
Then classify each one. Personal tool. Team beta. Supported internal product. Customer-facing product. For everything in the team beta column, name the rule that would promote it and the rule that would demote it. If you can’t name either rule, it’s a personal tool you’ve been mistaking for infrastructure. Demote it before something breaks.
This is what Nate called the demotion audit. It’s also Failure Mode #3 in the kitchen, the Open Door Problem, the one where the kitchen accepts every ingredient delivery without ever checking who sent it. You’ve been accepting tools onto the Line because they made someone’s morning twenty minutes shorter. That was the right call when there were two of them. There are not two of them anymore.
Tools accumulate. Decisions don’t.
The graveyard isn’t full of bad tools. The graveyard is full of tools you never decided about. Until somebody decides whether each one stays, moves up, moves down, or gets deleted, the practice runs on a list of stations you can’t see, doing work you depend on, owned by nobody.
Does that make sense?
The Monday Move
Pick one AI tool or automation your team built or installed in the last six months. Just one. The one you can name fastest. Classify it. Personal tool. Team beta. Supported internal product. Customer-facing product. Then write the two rules. The one that would promote it to the next tier. The one that would demote it.
If you can’t write either rule in two sentences, it doesn’t belong where it is. Demote it today.
Do that with one tool this week. Do it with the next one next week. Inside two months you’ll have a list of every station on your Line, who owns it, what tier it’s in, and what happens when it breaks. That’s not governance overhead. That’s the first time you have a Station Plan instead of a graveyard.
The intake summarizer at the dental practice is back up. The office manager still owns it. The owner knows it exists. The bigger thing they walked out with isn’t a fix to the summarizer. It’s a list, on a whiteboard in the back office, of the twenty-three other tools that turned out to be running the practice. Eight got demoted to personal tools. Three got deleted. Twelve are in the team beta column with promotion rules now written next to them. The Chef just learned the names of his own stations.
Find your list. The graveyard isn’t quiet because everything’s fine. It’s quiet because nobody’s looked.
Source influences: Nate B Jones on the prototype graveyard and the demotion audit. Distilled and operator-translated.
Framework spine: The Station Plan. Failure Mode callback: The Open Door Problem.
~ source material · Source influences: Nate B Jones on the prototype graveyard and the demotion audit. Distilled and operator-translated.
