The most-repeated insult in startup Twitter, and also the description of two companies worth a combined $40 billion. Here is the honest definition, real examples of both, and the test that tells you which one you are building.
An AI wrapper is a product that calls a foundation model's API, such as OpenAI's GPT or Anthropic's Claude, and adds an interface, prompt logic, or workflow on top, instead of training and hosting its own model. By that definition, nearly every consumer AI product today, including Cursor at a $29.3 billion valuation and Harvey at $11 billion, is technically a wrapper.
The word only became an insult because so many thin, undifferentiated versions of it launched at once, roughly 90% of which are projected to fail by the end of 2026. This page separates the definition from the insult and gives you a concrete way to tell which one your product actually is.
Every wrapper, defensible or disposable, is built from the same four layers. What differs is how much real value sits in layers two and four.
The chat box, form, or dashboard the user actually touches. Usually the newest, most polished part of the product and the easiest part for anyone to copy.
Prompt templates, business rules, formatting, and any proprietary data you inject before or after the model call. This is the layer that determines whether you have a real product or a thin skin.
A call to OpenAI, Anthropic, Google, or a similar provider that does the actual language understanding and generation. Every wrapper, defensible or not, depends on this layer.
How the raw model response gets validated, formatted, stored, or chained into further actions before the user sees it. Weak here means unreliable outputs; strong here is a real differentiator.
Not all wrappers are the same category. Defensibility rises as you move down this list.
A UI wrapped around a single prompt template calling a foundation model, with little to no logic in between. The lowest-defensibility category and the one most exposed if the model provider ships the same feature natively.
Example: Generic "AI writer" or "AI caption generator" tools with no proprietary data
Chains multiple model calls and business logic into a repeatable process, often integrating with other tools via API. More defensible than a chat wrapper because the workflow itself has value beyond any single prompt.
Example: AI meeting-notes tools that summarize, extract action items, and sync to a calendar automatically
Built on top of a foundation model but layered with domain-specific data, compliance requirements, and workflow depth for one industry. The category with the strongest moat, because the domain data and integrations are not something a model provider ships for free.
Example: Harvey (legal AI, $11B valuation, built on law firm workflow data)
Coordinates multiple tools, maintains memory or context over time, and takes multi-step actions rather than answering a single prompt. Often the most technically complex and, when done well, the hardest to replicate quickly.
Example: Cursor (codebase-aware coding agent, $29.3B valuation)
Useful for buyers, investors, and journalists, not just founders. Five tells that usually give it away.
Marketing leads with the model name
A landing page bragging "powered by GPT-4" before it explains any workflow or data advantage usually means there is no workflow or data advantage to lead with instead.
Free trial output is identical to pasting the same prompt into ChatGPT
If you can reproduce the entire output by copying the visible prompt into a generic chatbot yourself, the product added no meaningful logic on top of the model.
No visible data or history that improves with use
A dashboard that looks the same on day one and day one hundred signals nothing is being accumulated that would create switching cost.
Pricing is a flat markup on token usage with no outcome framing
Pure usage-based pricing with no outcome story usually means the company has not found (or does not have) a value driver beyond the raw API cost.
The team cannot describe a single feature the model provider could not ship next quarter
This is effectively the four-step replacement test asked out loud. If founders cannot answer it confidently, the product likely cannot either.
Both built on the same category of foundation model APIs everyone else can call.
AI coding assistant built on top of foundation model APIs, valued at $29.3 billion after a $2.3 billion funding round in November 2025, reportedly on more than $1 billion in annualized revenue.
Why it worked: The product indexes and understands an entire codebase, maintaining project-specific context that a raw model call to ChatGPT or Claude does not have. Switching away means losing that indexed context, which is a real switching cost.
Legal AI platform valued at $11 billion after a $200 million round in March 2026, reaching roughly $190 million in annual recurring revenue by January 2026, up from $100 million in August 2025. Used by more than 100,000 lawyers and the majority of the AmLaw 100.
Why it worked: Harvey layers legal-specific workflows, firm data, and compliance requirements on top of foundation models. A law firm cannot get the same output by pasting a prompt into a generic chatbot, because the value is in the domain integration, not the underlying model.
The same underlying models, an opposite outcome.
An early, well-funded AI writing wrapper that reportedly lost more than 50% of its revenue after ChatGPT shipped native writing features that overlapped heavily with Jasper's core product.
Why it failed: Jasper's core value, prompt-templated long-form writing, became something users could get for free directly from the model provider. The wrapper had a UI advantage, not a data or workflow advantage, and that advantage evaporated in one product update from OpenAI.
A Y Combinator-backed startup (W23 batch, founded 2022) that originally pitched converting text stories into Webtoon-style visual content, then pivoted in 2024 to offer generic enterprise "AI wrapper" products across several business use cases. It quietly shut down in 2025.
Why it failed: Once GPT-4-class models and image generation tools shipped equivalent capability for free or near-free, Wuri's wrapper offerings had no differentiation left, no proprietary data, no workflow lock-in, and no distribution advantage, in a crowded field of similarly undifferentiated products.
Builder.ai is often mentioned alongside wrapper failures, but its collapse is a different story and deserves an accurate label. The London startup raised $445 million and reached a $1.5 billion valuation with backing from Microsoft, marketing an AI assistant called "Natasha" that could allegedly build apps with no code. It filed for bankruptcy protection in May 2025 after reporting revealed Natasha was largely a facade, with hundreds of human engineers manually assembling code behind the scenes, and after allegations of inflated revenue through round-tripping transactions. This is a story about overstating AI capability and misrepresenting revenue, not simply a wrapper that got out-featured by a better model. It belongs in the AI-washing conversation more than the "thin wrapper" conversation, and lumping the two together misreads what actually happened.
MediaFast finds the exact subreddits and communities where your niche already gathers, so a defensible product actually reaches the people it was built for.
Both sides of the most common insult in AI startup discourse, argued honestly.
"You are not building technology, you are renting it"
The core objection is that anyone can call the same OpenAI or Anthropic API you are calling, so the product has no technical moat, only a UI head start that a better-funded competitor can erase in weeks.
Margins are structurally worse than real software
Every model API call costs money on a per-token basis, which caps gross margins well below traditional software. Thin wrapper products commonly run 25 to 35% gross margins, compared to 70 to 85% for defensible SaaS.
The model provider is your biggest competitive threat
If your entire feature set fits in a system prompt, OpenAI, Anthropic, or Google can ship the same capability as a checkbox in their next release and remove your reason to exist overnight.
Nearly all modern software is "built on a platform" now
Shopify apps are built on Shopify. iOS apps are built on Apple's frameworks. Most SaaS is built on AWS. "Built on someone else's infrastructure" has never by itself been disqualifying, what matters is what you add on top.
The wrapper label ignores the data and workflow layer entirely
Harvey is technically "built on" foundation models. So is Cursor. The criticism collapses once you compare a real vertical or agent wrapper with genuine domain depth to a copy-paste prompt tool, they are not the same category even though both call an LLM API.
Distribution and trust are moats models cannot ship natively
A foundation model provider shipping a generic feature does not automatically capture your specific customer relationships, workflow lock-in, brand trust, or community. Those take years to build and are not solved by a model update.
Score your own product honestly against these three axes before anyone else does it for you.
| Axis | Defensible looks like | Disposable looks like |
|---|---|---|
| Proprietary data or context | Accumulates data over time (codebase state, case history, customer records) that a fresh model call does not have access to. | Every session starts from zero. The product adds nothing to the raw model beyond a prompt template. |
| Workflow depth | Embedded in a multi-step process with real switching costs: integrations, saved history, team permissions, compliance sign-off. | A single input-output loop the user could replicate by pasting the same prompt into ChatGPT directly. |
| Distribution advantage | Owns a community, brand, or channel independent of the model itself (a large existing user base, a partner network, a marketplace presence). | Relies entirely on paid ads or generic SEO to acquire users who have no reason to stay once a competitor undercuts on price. |
Run your product through these four questions in order. The more "yes" answers to the risk column, the more exposed you are.
If your entire feature set fits in a system prompt with no other logic, the answer is yes, and that is the single strongest predictor of a wrapper getting replaced.
If a brand-new user gets the exact same output as your longest-tenured customer, there is no compounding data advantage protecting you.
If a user could copy their last prompt into a competitor's tool and get an equivalent result in under five minutes, your switching cost is effectively zero.
If every single user found you through a generic Google ad and has no community, integration, or brand reason to stay, your growth channel is as replicable as your product.
What the people funding and gatekeeping this category are actually saying, not anonymous commentary.
If you're really just counting on the back-end model to do all the work and you're almost white-labeling that model, the industry doesn't have a lot of patience for that anymore.
Darren Mowry
VP, Google for Startups
Reported by PYMNTS, 2026, a direct warning aimed at thin, undifferentiated AI wrapper startups.
The ultimate sign of conviction.
Pat Grady
Partner, Sequoia Capital
Describing Sequoia leading three consecutive funding rounds in Harvey, a vertical AI wrapper, as the firm's repeated bet on defensible domain-specific AI products.
The numbers that separate the theory from the balance sheet.
25 to 35%
Typical gross margin for thin AI wrapper products, versus 70 to 85% for defensible SaaS
~90%
Share of thin AI wrapper startups projected to fail by the end of 2026
60 to 70%
Share of AI wrapper startups reportedly generating zero revenue
~65%
Wrapper product churn within 90 days, versus roughly 35% for the average SaaS product
$11B
Harvey's valuation in March 2026, on roughly $190M ARR, as a vertical wrapper with real defensibility
$29.3B
Cursor's valuation in November 2025, on more than $1B in annualized revenue
Three different technology bases, three different risk profiles.
| Dimension | AI wrapper | Traditional SaaS | AI-native product |
|---|---|---|---|
| Core technology | Calls a third-party foundation model API for the core intelligence layer | Runs proprietary business logic and a database, no model dependency | Built AI-first, often fine-tunes or orchestrates models around proprietary data |
| Typical gross margin | 25 to 35% | 70 to 85% | 70 to 85% once past early scale |
| Biggest competitive threat | The model provider shipping the same feature natively | A better-funded direct competitor | A faster-moving AI-native competitor, rarely the model provider itself |
| Defensibility source | Usually none beyond UI, unless data or workflow is added | Data, integrations, switching costs built over years | Proprietary data plus AI-specific workflow depth |
Concrete moves, not just theory, for turning a thin wrapper into a defensible one.
Wrap the model call inside a multi-step process specific to your users' job, so the value is the process, not the single response.
Store history, preferences, or domain data per customer so day-100 output is measurably better than day-1 output. That compounding gap is what a fresh competitor cannot copy.
Harvey did not launch as "AI for every profession," it launched for lawyers specifically. Depth in one domain's compliance and workflow requirements is a moat breadth is not.
A community, an integration marketplace listing, or an existing audience on a platform like Reddit gives you a growth engine independent of whichever model you are calling underneath.
Wrappers priced per API call inherit the model provider's cost structure and margin ceiling. Products priced around a business outcome (hours saved, cases closed, revenue generated) can price above raw usage cost.
Step four, distribution, is where most technically-solid wrapper products actually die. A tool like MediaFast helps you find the exact subreddits and communities your specific niche already gathers in, so the moat you built on the product side actually reaches the people it was built for.
Six honest questions. The more you answer "no" to, the closer your product sits to the disposable end of the spectrum.
Could a competitor rebuild your core feature by copying your system prompt from a screenshot?
Does a returning user get a measurably better result than a brand-new user, or is every session identical?
Would switching to a competitor cost your average user more than five minutes of setup?
Do you have a distribution channel (community, marketplace, existing audience) that does not depend on paid ads?
Is your gross margin closer to 70%+ or closer to 30%, and do you know why?
If OpenAI, Anthropic, or Google announced your exact feature tomorrow, would your product still have a reason to exist?
Each one is fixable if caught before launch, expensive if caught after.
Consequence: Advertising the model name invites the direct comparison "why not just use ChatGPT," which is the exact question a defensible product should never raise.
Consequence: Spreading a thin logic layer across ten unrelated use cases means none of them get deep enough workflow integration to become a real moat.
Consequence: If the whole value proposition lives in a prompt a competitor could reverse-engineer from your own marketing screenshots, there is no defensibility to protect.
Consequence: Per-token API costs eat margin at scale in a way flat-rate SaaS infrastructure never did. Wrapper founders who do not model this early get surprised by margins in the 25 to 35% range instead of the 70%+ they expected.
Consequence: A product with no owned channel competes purely on paid acquisition against every other wrapper doing the same thing, driving CAC up for the whole category.
The terms this whole debate depends on, defined in one sentence each.
AI wrapper
A product built primarily by calling a foundation model's API (OpenAI, Anthropic, Google) and adding a layer of interface, prompts, or light logic on top, rather than training or hosting its own model.
Foundation model
A large, general-purpose AI model (GPT, Claude, Gemini) trained by a major lab and exposed to developers through an API, which most wrapper products are built on top of.
Thin wrapper
A wrapper with minimal proprietary logic beyond the prompt itself, the category with the weakest defensibility and the highest projected failure rate.
Moat
A durable advantage (proprietary data, workflow depth, distribution) that protects a product from being replicated by a competitor or absorbed by the underlying model provider.
Vertical AI
An AI product built for one specific industry's workflow and compliance requirements, generally the most defensible wrapper category because the domain depth is hard to replicate quickly.
AI washing
Overstating or misrepresenting a product's actual AI capability, sometimes to the point of using undisclosed human labor behind an "AI" interface, as alleged in the Builder.ai collapse.
The primary sources behind the definitions, numbers, and examples on this page.
Calling a product "just a wrapper" is technically true of almost everything built on AI today, including the two companies in this piece worth a combined $40 billion. The word only tells you what layer the product is built on, not whether it deserves to exist.
The real question is the one this page keeps returning to: does your product accumulate proprietary data, sit inside a real workflow, and own a distribution channel independent of the model underneath it. Answer honestly, run the four-step replacement test, and you will know which side of the 90% failure line you are actually on.
Definition: interface plus logic on an API
An AI wrapper calls a foundation model's API and adds interface, prompts, or logic on top rather than training its own model.
Defensibility: data, workflow, distribution
The three axes that separate a moat from a coat of paint over someone else's model.
Proof: same models, opposite outcomes
Cursor and Harvey are worth a combined $40B+ built on the same APIs that Jasper and Wuri lost to.
The six questions people ask most when trying to figure out if their product is defensible or disposable.
An AI wrapper is a product that calls a foundation model's API, such as OpenAI's GPT or Anthropic's Claude, and adds a layer of interface, prompt engineering, or business logic on top, instead of training and hosting its own AI model. Nearly every consumer-facing AI product today is technically a wrapper in this sense, the real question is not whether a product is a wrapper, it is whether the layer on top of the model is defensible or disposable.
Not inherently. Cursor is valued at $29.3 billion and Harvey at $11 billion, and both are built on the same foundation model APIs anyone can call. What determines whether a wrapper survives is whether it adds proprietary data, real workflow depth, and a distribution advantage on top of the model, or whether it is just a prompt template with a UI, which is the category with roughly 90% projected failure by the end of 2026.
Jasper AI, an early AI writing tool, reportedly lost more than half its revenue once ChatGPT shipped native writing features that matched its core capability for free. Wuri, a Y Combinator-backed startup, pivoted into generic enterprise AI wrapper offerings in 2024 and shut down in 2025 once larger platforms shipped equivalent generative features natively, with no proprietary data or workflow to fall back on.
Three things: proprietary data or context that compounds over time instead of resetting every session, workflow depth with real switching costs like integrations and saved history, and a distribution advantage such as a community or existing audience that does not depend on the underlying model. Products missing all three are the ones most exposed if a foundation model provider ships an equivalent feature natively.
Cursor indexes and understands a user's entire codebase, maintaining project-specific context a raw call to ChatGPT or Claude does not have, which creates real switching cost. Failed wrappers like Jasper or Wuri offered a UI layer over a prompt with no comparable data or workflow depth, so once the underlying model shipped the same capability, there was nothing left to differentiate them.
Not exactly. Builder.ai's collapse involved allegations that its "Natasha" AI assistant was largely powered by hundreds of human engineers rather than genuine AI, plus allegedly inflated revenue figures, leading to a $1.5 billion valuation and bankruptcy in May 2025. That is a story about misrepresenting AI capability and revenue, closer to what is now called AI washing, rather than a defensible-versus-disposable wrapper story like Jasper's or Wuri's.
