Original practice question. The scoring rubric is our teaching guide. How we research and score answers
TL;DR: "Wrapper" describes most software: products are built on infrastructure they do not own. A thin wrapper (a prompt and an interface) is fragile because it is easy to copy and the model provider can absorb it. A deep application owns proprietary context, the workflow, distribution and a quality flywheel, and its margins hold up as model prices fall. The question is not whether you use someone else's model, but what you own on top of it.
How to approach it
Acknowledge the real risks of thin wrappers, then show what separates fragile from durable AI applications, and how a team can tell which one it is building.
A strong answer
Why the criticism has a point. A product that is a well-written prompt and a clean interface over a model API can be copied quickly. The model provider can add the same capability to its own product. Margins depend on someone else's pricing. If the model gets better at the task directly, users may not need the product.
Why "wrapper" alone is not a verdict. Much successful software runs on infrastructure owned by others: cloud providers, payment networks, app stores. What made those companies durable was what they built on top: workflow, data, distribution, trust.
Thin versus deep.
| Thin wrapper | Deep application | |
|---|---|---|
| What it owns | A prompt and an interface | Customer data connections, workflow, evals, distribution |
| Switching cost | Near zero | Meaningful: data, habits, integrations |
| Exposure to provider | High: provider can absorb the feature | Lower: provider lacks the context and workflow |
| Margin trajectory | Squeezed by competitors | Improves as model prices fall if pricing holds |
| Quality over time | Same as anyone using the model | Improves from proprietary evals and outcome data |
Read diagram description
Two products on the same model API. A thin wrapper owns only a prompt and an interface, which can be copied quickly and absorbed by the provider. A deep application adds customer data connections, the workflow where work happens, evals and outcome data, and distribution and trust, so switching costs are real and falling model prices can widen its margin.
Margins. Gross margin for an AI application is roughly (price minus inference, retrieval and infrastructure, and human fallback costs) divided by price. Model prices per unit of capability have fallen fast: Epoch AI measured declines of between 9 and 900 times per year depending on the capability level, while warning the fastest rates may not persist (LLM inference price trends, March 2025). That helps applications that keep their pricing anchored to customer value rather than to token costs.
The squeeze shows up on the other side when usage grows faster than prices fall. In mid-2025 Cursor moved its Pro plan from request-based pricing to $20 of included usage and apologised that the change "was not communicated clearly" (Cursor, July 2025); in April 2026 GitHub announced Copilot was moving to usage-based billing because "the current premium request model is no longer sustainable" (GitHub changelog). Both are company statements, and both are applications discovering that flat prices and agent-heavy usage do not mix. Owning the pricing model is part of owning the business.
Three questions to tell which you are building.
- If the model provider shipped your core feature tomorrow, what would customers keep using you for?
- What data or context does your product have that a competitor with the same model would not?
- Is your product where the work happens, or a tool people visit occasionally?
What interviewers probe next
"How do you move from thin to deep?" Connect to customers' data, own more of the workflow, and build evaluation and outcome data that improves quality in ways competitors cannot copy.
"Should we switch to building our own model to escape the label?" Only if it creates a lasting advantage on your tasks. The label matters less than what customers would lose by leaving.
How it's scored
Weak: Accepts or rejects the criticism wholesale.
Hire: Distinguishes thin wrappers from deeper products.
Strong Hire: Explains the provider absorption and copying risks, what makes applications durable, the margin formula and the effect of falling model prices, and gives a test for which one you are building.
Common mistakes
Treating model ownership as the only path to defensibility.
Ignoring the risk that the provider ships the same feature.
Pricing on token costs rather than customer value.
See the moat ladder and wrapper gross margin.