The wrapper risk index

Wrapper risk is the single most argued-about property of an AI company and the least measured. This report puts a number on it for all twenty-nine markets we track, shows how it distributes, and tests whether it predicts the two things people assume it predicts: saturation, and death.

Definition and coverage

Wrapper risk is a 0–100 estimate of how much of a product’s value derives from a general-purpose model it does not own. It rises with thin interfaces over public APIs and falls with proprietary data, workflow depth, regulatory access, or a model trained in-house. It is a statement about defensibility, not about quality. The complete definition is on our methodology page.

6,113 of 6,441 live tools carry a score. The remaining 328 are recent additions we have not assessed, and are excluded from every figure here rather than counted as zero. Market averages are reported only for markets with at least 40 scored tools, which is all 29.

A second, stricter measure appears alongside it: the count of tools we have flagged outright as wrappers, 405 across the live catalog. The score is a spectrum; the flag is a judgement. Both are shown below because they disagree in useful places.

Finding one: the score is bimodal

3,962 tools score under 20 and 1,182 score 60 or above. Between 40 and 59 there are 53 — under one percent of the scored population. A measure intended as a continuum is behaving as a classification.

We take that as a property of the thing being measured rather than of the measurement. There is no gradual way to half-own a model or half-hold a regulated integration. A company either has an asset the model provider does not, or its product is an arrangement of prompts around one that does. The middle of this distribution is empty because the middle does not exist in practice.

Finding two: a sevenfold spread by market

Writing averages 55.0 and AI Models 7.7. The ordering is close to a ranking of how directly each market’s core task overlaps with what a general-purpose model does unaided. Writing, social posting, summarising, scheduling and tutoring sit at the top. Training models, defending networks, and operating inside clinical or financial systems sit at the bottom.

Finding three: it tracks saturation

The six thinnest markets average 42.1 on wrapper risk and 48.5 on saturation. The six thickest average 10.2 and 25.5. This is the most useful relationship in our market-level data, and it supplies a causal story that the raw saturation number does not: markets do not become crowded because they are lucrative, they become crowded because the barrier to entering them is an API key.

It also explains why saturation is difficult to reverse. A market whose barrier is low stays low. Every departure is replaced, because replacement costs a weekend.

Finding four: it does not predict death

The intuitive next step is that thin products should fail more often. They do not, at least not measurably. Writing, the thinnest market in the catalog, has a failure rate of 8.6% — essentially the catalog average of 8.4%. Social, second thinnest, is at 6.7%, better than average. Gaming, which ranks 20th of 29 on wrapper risk at 15.3, has the worst failure rate in the catalog at 32.9%.

The explanation is that wrapper risk and mortality describe different failure modes. A thin product usually does not shut down; it gets squeezed. Its pricing power erodes, its free tier is matched by the model provider, and it persists as a smaller business than it planned to be. What kills companies, according to our mortality report, is selling a generated artefact to a churning consumer — which is a revenue problem, not a defensibility one.

For an investor that distinction matters. Wrapper risk is a margin question. Category and buyer are the survival questions. A diligence process that only asks the first one is asking about the wrong risk.

What this measure cannot do

Three limits worth stating. It is assessed at a point in time, and a company that ships a proprietary dataset next quarter will not move until we reassess it. It says nothing about revenue: several high-scoring products are among the most commercially successful in their markets. And it is a judgement about business model made from outside the company, which means it will be wrong about specific firms in both directions even where the market-level pattern holds.

The market averages are the durable part of this report. Use the per-tool scores on each product page as a prompt for a question, not as an answer to one.

Photo: Lucas George Wendt / Pexels. Colour-graded for Falcoscan.

← All reports