AI feature or AI wrapper: a decision test an investor can run

An AI wrapper is a product whose value is the model's value with a user interface on it. An AI feature is a product where the model does one job inside a workflow the company owns, with data, integrations and controls that a competitor, including the model provider, would have to rebuild. The difference is what an investor is paying for: in the first case, distribution; in the second, a product.

The word "wrapper" is used as an insult, and that is unhelpful, because a wrapper with distribution can be a fine business and a feature with no customers is worthless. This page gives a test that classifies the product from evidence and then asks the only question that matters for the price: what does the company own that the model provider does not?

Why does the distinction decide the price?

The distinction decides the price because the model providers are moving into applications, and a company whose product is the model's output with a screen on it is competing with its own supplier. Reveneau's diligence classifies every AI target on this axis before the margin work starts, because the classification decides which margin is being rebuilt: a wrapper's gross margin is the provider's price list, and a feature's is a design choice.

Foundation Capital's September 2026 piece on model providers and application startups makes the argument from the provider's side. Providers have a view across everything built on their APIs, can identify which use cases work, and have already shipped products that compete with companies built on them; the piece names OpenAI's agent products and Anthropic's Claude Code as examples and describes the "Windsurf saga" as a case of preferential treatment followed by competition [1]. Its conclusion is that the startups that survive own high-exception workflows, subjective success criteria and fragmented legacy systems, which is a description of the things a provider will not build.

Two buyer-side surveys show the other half. Menlo Ventures' December 2025 survey of enterprise decision-makers found 76 percent of enterprise AI solutions were bought rather than built internally, up from 53 percent in 2024 [2], which says buyers will pay for a product on top of a model. a16z's June 2025 survey of 100 CIOs found 37 percent running five or more models in production, up from 29 percent, and reported that differentiation by use case is the main reason buyers use several vendors [3]. Buyers are willing to pay for the layer above the model and are also comfortable swapping the model underneath. That combination is good for a feature and bad for a wrapper.

The decision test

The test takes an hour and uses three artefacts: the product itself, the trace from a live run, and the model-provider invoices. Run it after the claim test on how to verify an AI claim, because the trace from that session is the input here.

1. Remove the model in your head. Describe the product with the model call replaced by a black box that returns nothing. What is left? A data pipeline, integrations, a permissions model, a workflow, a customer's history, a review queue? Or a text box? Write the list.
2. Count the model calls per user action. From the trace. A wrapper makes one call and shows the result. A feature makes several, with retrieval, tool use and checks between them, and the user never sees most of them.
3. Ask where the input comes from. If the customer types it, the product is one step from a chat window. If the product assembles it from systems the customer already runs, the integration is the asset.
4. Ask where the output goes. If the user reads it and leaves, the product is a viewer. If the output writes into a system of record, triggers a workflow, or becomes training data for the next run, the product is embedded.
5. Read the invoice against revenue. From rebuilding inference gross margin from invoices. A wrapper's inference bill tracks revenue one to one because every unit of value is a model call. A feature's bill is one line among several.
6. Ask what changed when the model was swapped. Every company has swapped a model at least once in the last year; Anthropic and OpenAI both publish retirement schedules and both retired models in 2026. If the swap was a config change, the model is a component. If the swap broke the product, the model was the product.
7. Write the classification and the reason. Wrapper, feature, or in between, with the evidence from the six steps above. The classification goes in the report next to the claim table.

What does a wrapper look like in the evidence?

A wrapper looks the same in the trace, the invoice and the retention data, and the table below is the shape to look for.

Evidence	Wrapper	Feature
Model calls per user action	One, result shown directly	Several, most invisible to the user
Input source	Typed by the user	Assembled from customer systems
Output destination	Read on screen	Written to a system of record or a workflow
Inference bill against revenue	Tracks one to one	One cost line among several
Effect of a model swap	Product changed or broke	Config change, evals re-run
Prompt as share of the codebase	Most of the logic	A small part of a larger system
What a churned customer replaces it with	A chat window, or the provider's own app	A rebuild of the integrations

The last row is the one to test against the churn data. Ask where the churned customers went. If they went to the model provider's own product or to a general chat interface, the product was a wrapper in the customer's eyes whatever the deck said.

When is a wrapper still a good investment?

A wrapper is a good investment when it owns something the provider will not build: a distribution channel into a specific buyer, a workflow with enough exceptions to need a specialist, a data loop that improves with use, or a regulatory or trust position the provider does not want. The test for each is whether the ownership shows up in the numbers.

Bessemer's August 2025 State of AI describes the fastest-growing group of AI companies as having explosive adoption, low switching costs and margin compression, and gives its gross margin as 25 percent, often negative, against 60 percent for the steadier group [4]. Low switching cost is the wrapper's weakness. Growth at 25 percent gross margin with customers who can leave for a prompt is growth the provider can take back at will. Growth at 60 percent with customers whose integrations would take a quarter to rebuild is a business.

So the questions for a wrapper's investor are: is the retention consistent with switching cost, is the margin consistent with a price above the model's price, and is there anything in the product a provider's feature launch would not cover? Data moat claims covers the data-loop version of that question, and model dependency risk covers what happens to a wrapper on the day the provider changes the price.

ICONIQ's July 2026 survey adds one detail: the companies it surveyed run 3.3 models on average and nearly half use two or more model types [5]. A product that is genuinely a feature usually shows that pattern, a cheap model for routing, a strong one for the hard step, a different one for embeddings. A product that runs one model for everything has a simpler architecture and a simpler dependency.

What if the target is a vendor rather than a startup?

If the company you are checking is a services firm claiming to be AI-native, the questions are different, because there is no product to classify and the claim is about how the work is done. Ten questions that expose a fake AI-native claim is the list for that case, and we ask people to use it on us.

What to write in the report

Write the classification, the seven pieces of evidence, and the one-line consequence for the valuation. "Wrapper: one model call per action, inference bill tracks revenue at 0.4, model swap in March broke the product for a week, churned customers report moving to the provider's app. Price as a distribution business with a supplier who is also a competitor." Or: "Feature: five model calls per action, inference at 12 percent of revenue, model swap was a config change verified by the eval suite, churned customers cite price. Price as software."

The AI startup due diligence guide treats this classification as the link between the engineering read and the money read, because the answer decides which margin you are rebuilding and how much of the technical work is worth doing. If you are building an AI product yourself and want the feature side of this line, building AI products is the guide for that.