How to tell an AI wrapper from a real product

Every AI product sits somewhere on a line. At one end is a company that owns a model, a dataset, a workflow or a distribution channel that nobody else has. At the other is a text box, a system prompt and someone else’s API. Most of the argument about AI startups is really an argument about where on that line a given company sits — and it is more answerable than it looks.

What the score actually measures

Wrapper risk is a 0–100 estimate of how much of a product’s value comes from a general-purpose model it does not own. It goes up when the product is a prompt and an interface over a public API, and down when there is proprietary data, workflow depth, regulatory access, or a model the company trained itself. It is a statement about business model, not about quality: plenty of high-scoring products are well made, well loved and profitable. What the score says is that their defensibility depends on somebody else’s pricing decisions. The full definition sits on our methodology page.

328 live tools carry no score yet, mostly recent additions we have not assessed. They are excluded from every figure below rather than counted as zero.

The score is almost binary

The first thing the distribution shows is that there is barely a middle. Products cluster hard at both ends, with a near-empty gap between 40 and 59.

That shape is worth sitting with, because it contradicts how the debate usually goes. People argue about wrappers as though every company is somewhere on a gradient, gradually earning its way toward defensibility. The data says something blunter: a product either owns something material or it does not, and very few are halfway. Whatever a company is building toward, on any given day it is on one side of the line.

Practically, that means the question is answerable early. If you cannot name the asset — the data, the workflow, the licence, the channel — the product does not have one yet.

Which markets are thin

Wrapper risk is not evenly spread. Averaged across each market’s live tools, the range runs from 7.7 to 55 — a sevenfold spread inside the same catalog.

Writing leads at 55, and the reason is mechanical: producing text is exactly what a general-purpose model already does. A writing product has to add something on top of a capability its supplier ships for free, and most do not add enough. Social (45.1) and Productivity (44.7) sit just behind for the same reason — summarise, schedule, rewrite, reply are all things the underlying model does natively.

At the other end, AI Models averages 7.7 because those companies are the supplier. Security (8.6) and Healthcare (8.8) get there differently: the product has to reach inside a customer environment, or satisfy a clinical and regulatory bar, and neither of those is a prompt.

What the thick end looks like

These are among the lowest wrapper-risk scores in the catalog, each paired with a high opportunity score. Read them as examples of what owning something looks like in practice — a model trained in-house, a proprietary scientific pipeline, a clinical integration, a translation engine built over years.

Why thin markets get crowded

Plot wrapper risk against saturation and the two move together. This is the mechanism behind the thing everyone observes anecdotally: markets do not get crowded because they are attractive, they get crowded because they are easy to enter.

The axis label reads opportunity out of habit; here the vertical is average wrapper risk. The pattern is the same either way — the thin markets are the contested ones.

Four questions that settle it

You do not need our score to run this check on a product in front of you. Ask what happens if the underlying model provider ships the feature natively tomorrow. Ask what data the company holds that a competitor could not buy. Ask what breaks for the customer if they switch — if the answer is “nothing, they re-enter their prompt”, that is your answer. And ask who signed the contract: an integration into a regulated system takes a year to win and a year to replace, which is a moat even when the model underneath is commodity.

Three yeses and the score will be low. Three noes and it will be high, and you will have reached that conclusion faster than we did.

Every market, scored

Photo: Reinis Brūzītis / Pexels. Colour-graded for Falcoscan.

← All articles