The Remove-the-AI Test
Does this need a model, or would rules do? Remove the AI entirely and look at what's left. If a simpler product still delivers most of the value, an AI-native architecture isn't justified, and we say so.
Take the model out. See what's left.
Remove the AI. Does the simpler product still deliver most of the value?
Build it with rules.
Forms, reminders, search and plain logic do the work. Add a model only where the rules run out, and say exactly where that is.
The model is the product.
The value leaves with the AI, so it earns its place. Build it properly: tested on real requests, with every step it takes traced.
Running the test
Say what it does, without the word AI
One sentence: what the product does for the person using it.
The sentence describes a job the user needs done, not a technology.
Remove the model
Describe the product that's left: the forms, rules, reminders and search that would still work.
A concrete description of the simpler product, specific enough to build.
Compare the value
Ask whether the simpler product still does most of the job. This is a judgement, so write down the reasoning.
You can point to the part of the job that only the model can do, or you can't.
Place the model
If the model stays, it goes only where the rules run out, and nowhere else.
The model's job is written down in one line.
The booking assistant, without the assistant.
Illustrative example A clinic booking assistant. Not client work.
Take the AI out of the clinic booking assistant and a lot still works. Reminders over WhatsApp and payment at the time of booking cut no-shows with no model at all. Most of the value survives.
What doesn't survive is reading a message like "anything Thursday after five?" and offering the right slots. That's where the rules run out, so that's the model's one job.
The test also settles a second question. The model's job is narrow and common, so an off-the-shelf model does it. Custom training gets a Don't build.
Where it goes wrong
- Misread
Reading it as anti-AI
Why The test isn't there to remove the model. It finds exactly where the model earns its place.
- Misread
Testing the demo, not the job
Why A model makes a demo impressive. The test asks what the user needs done on an ordinary day.
- Misread
Keeping the AI for the pitch
Why AI in the positioning isn't a reason for AI in the architecture.
In the Journal: Is it worth building? · AI prototype to production
Published 24 September 2026. Graylemon original, from our engagement method.