Take the AI Out and See What Is Left
An AI tool that is only a screen and a few instructions around someone else's model has nothing to defend. Before you buy or build one, take the model out: the value lives in the workflow, the processing and the data around it.
If the company that makes the AI model can ship your tool as a free feature, the tool has nothing to defend. Before you buy or build one, take the model out and see what is left.
The quickest AI product to build is a wrapper: a screen and a few clever instructions around someone else’s model. It takes a weekend to build and a weekend to copy, and it loses its value when the model’s maker adds the same feature, when competitors copy the screen, or when model prices fall.
That matters whether you are building an AI tool or paying for one. The value that lasts is what the model cannot supply: the processing around it, the workflow it fits, and the data built up over time.
How hard each layer is to copy
A screen over the model
A good interface on someone else’s model. Anyone can copy it, and when the model’s maker ships its own screen, it is worth nothing.
Clever instructions
Well-written prompts and examples. Useful, but anyone with the same model can discover, share and improve them.
The workflow around it
The AI sits inside a process people already follow: preview and approve, who may see what. Harder to copy, because it takes knowing how the users work.
Processing the model does not do
Work before and after the model call: statistics, date parsing, layered checks. Take the model away and this layer still has value.
Your own data, and trust
Data the model does not have: what users supplied, their domain’s categories, a record of how accurate the tool has been, and the trust users build on it. Changing model provider cannot copy it.
The deeper the layer, the harder it is to copy. A tool that lives only in the top two depends on its model supplier never doing the same thing.
The test: take the model out
An empty text box and a loading spinner with nothing to load. Every feature depended on the model.
Date parsing still works, six statistical checks still find patterns, and every timeline view still renders. The product is slower to use and still works: the model was the narrator.
Chronologiq is an AI product our founder, Shan Peiris, built before neesh Inc. The engineering below is from it, and the numbers are his rather than a client’s.
Chronologiq passes the test. Without the model, users lose automatic parsing of pasted text and keep everything else, including manual entry. The model makes the product faster to use; the product exists without it.
How a product moves down the layers
From a screen to a workflow
The first step past a wrapper is fitting the AI into a specific workflow and owning what happens before and after it.
Chronologiq’s upload goes: upload, choose the timeline, let the AI process the text with that timeline’s context, preview each event, approve and save. The AI is one step of five.
From a workflow to processing
The second step is processing that does not depend on the model, which makes results more accurate and cheaper and cannot be copied by switching model providers.
In Chronologiq, a date parser reads the dates first, a smaller model summarizes long documents, and six statistical checks find patterns without any AI. The main model gets smaller, checked input, and its output improves.
From processing to your own data
The deepest layer is data the product builds up as it is used.
In Chronologiq, that includes the parsing guidance stored for each timeline, which teaches the model that domain’s language. None of it exists in the model; it exists in the product and grows with use.
Three questions to ask of any AI tool
What would be left of your AI tool if you took the model out?
Book Free Assessment