I handed a coach I work with the entire context file behind my menu engineering reports. The whole thing, in one markdown file, so he could give it to his own AI.
People in my line of work don't usually do that. The file is the thing. If the client has it, what are they paying for? And the answer, which took me a while to be comfortable with, is that they're paying for the file. That was always the deal.
So if you're a coach or consultant talking to an AI automation agency, or a freelancer, or someone like me, there is exactly 1 question that separates the people who'll leave you with a system from the people who'll leave you with a subscription. What do I keep if you leave?
TLDR
- The question is "what do I keep if you leave?" Ask it on the first call.
- I handed a client my whole file because the file was always what he was paying for.
- 4 answers should worry you, and 1 of them sounds reassuring.
- A good answer names 4 specific things, and none of them is "the prompts."
- Behind the question is a bigger one, about whether you're buying output or production.
The Question: What Do I Keep If You Leave?
Most conversations with an AI person go the same way. They show you outputs. The outputs look good, because outputs always look good in a demo. You ask about price and timeline. You don't ask about the day it ends.
Ask about the day it ends. Not because you're planning for it, but because the answer reveals the shape of what you're buying. Somewhere between $160 billion and a lot has been spent on AI that delivered very little, and my read on why is that most of it bought outputs and none of it bought the thing that produces them. I've written about why AI consultants fail. Consultants deliver decks. Operators deliver systems that run. The question tells you which one you're talking to.
Why I Handed a Client the Whole File
Back to the markdown file, because I want to be clear about what was in it and what wasn't.
It held the context behind those reports: the job, the standard, the reference examples, the checks that run before anything ships. My QA checklists run, I think, 20 to 40 filter points, and they were in there. One thing I deliberately left out was a lever library I'd built recently. It was working across the last 3 reports. It was also new and not battle-tested, so I told him it wasn't in the notes yet and needed more runs before I'd stand behind it.
46 minutes later I shipped the report that library was built for. 2 and a half hours after that, his reply: "This is fantastic."
Here's the thing. The part that took work was never the prompt. It was writing my own judgment down in a form a machine can follow, then adding the checks that catch it when it drifts. That's why I could hand the file over without losing anything. The file is the asset, and I can point it at whatever model I'm using next month. So can he. That's the Ownership Clause, and it's not generosity. It's the only honest version of the deal.
4 Answers That Should Worry You
When you ask the question, here's what worries me:
- "You keep all the outputs." Of course you do. You paid for them. This answer means there's nothing else.
- "You keep the prompts." Prompts without the reference library and the checks are templates, and templates fail at the edges. The prompts are the least valuable 10%.
- "It's all in the platform." Meaning it's in a tool they control, on a plan they pay for, and the day the invoice stops the system stops. That's a lease with extra steps.
- "Don't worry, we're not going anywhere." This is the one that sounds reassuring. It's also not an answer. Everyone goes somewhere eventually, including me.
The pattern in all 4 is the same. The judgment, the part that makes the outputs good, stays with them. What you keep is the residue.
What a Good Answer Sounds Like
A good answer names things. Specific files you could open on your own laptop the day after they leave.
- A Reference Library. Your approved work, your rejected work, the reasons for both. Built from your material, not theirs.
- A rules file with the reasoning in it. Not "no server scripts" but "no server scripts, because the owner can't control what staff say after we're gone." The why is what lets you change the rule later without breaking it.
- The verification questions. The fixed checks every draft gets read through. If they can't show you the questions, the quality was living in their head.
- Documentation a normal person can run. Not a developer. You, or your VA, in an afternoon. If it needs them to operate, you don't own it.
Notice that none of those is a model or a platform. The model changes. Mine has changed under me more than once, and the files carried over each time. A good AI person is selling you the files and renting you their time to build them. A bad one is renting you the files and calling it a service.
The Question Behind the Question
The real thing you're asking is: am I buying output, or am I buying production?
Output is what arrives every week. It's fine. Plenty of businesses run on it. But it doesn't accumulate, and the day it stops, you're back where you started with a folder of PDFs.
Production is a standard, written down, with checks, that keeps producing whether or not the person who built it is in the room. It's slower to set up and it costs more in the first month. It's also the only version where month 6 is better than month 1. I've watched a client move to a new person with a 1-week wobble because the files were there. I've watched another start from 0 because they weren't.
So ask the question. If the answer makes them uncomfortable, that's information. If the answer is a list of files, that's a person you can work with, whether or not that person is me.
If you've read this far and you're about to sign with someone who answered "you keep the outputs," I can't help you.
Just kidding. I probably can, and the first step is free either way. Send me 1 piece of your work. You get a finished asset back in your voice within 48 hours, and if we go further, you keep every file. That's written down too.
FAQs about hiring AI help
What should I ask an AI automation agency before hiring them?
Ask what you keep if they leave. A good answer names specific files: a reference library built from your material, a rules file with the reasoning included, the verification questions every draft is checked against, and documentation a non-developer can run. If the answer is outputs, prompts, or a platform they control, you are buying a lease.
Are the prompts the valuable part of an AI system?
No. Prompts are the smallest layer, and without the reference library and the checks around them they behave like templates. The judgment lives in the examples and the verification loop. That is why I handed a client my full context file and lost nothing by doing it.
What is the difference between buying AI output and buying AI production?
Output arrives weekly and stops when the invoice stops. Production is a written standard with checks that keeps running and improving whether or not the builder is present. Output is fine for a month. Production is the only version where month 6 beats month 1.
Can an AI system I own survive a change of model?
Yes, if it is built as files rather than as a platform configuration. The reference library, rules, and verification questions are plain text and point at whatever model you run next. Mine have moved across model changes more than once with the files intact.