"It's just a wrapper" is one of the quickest ways to dismiss an AI product. Sometimes the dismissal is fair. Sometimes it's lazy. The useful question isn't whether your product sits on top of a model someone else built, because almost every AI product does. It's what you own that the model provider can't copy next quarter.

This article defines the wrapper, explains why the word is used as a criticism, and lays out a spectrum from a thin prompt layer to a deep workflow product. For the wider definition of the category, start with what an AI-native startup is.

What Is an AI Wrapper?

An AI wrapper is a product whose main value is a user interface and a set of prompts placed over a third-party model. The customer types or uploads something, the product reshapes it into a prompt, calls the model, and formats the answer. The model does the thinking. The product does the packaging.

That's not automatically bad. Packaging is real work, and plenty of buyers will pay for something that's easier to use than a raw model. The problem is what happens next: anyone can build the same packaging in a weekend, and the model provider can add the same feature to its own interface.

Bessemer's State of AI 2025, published in August 2025, puts it bluntly when it describes the fastest-growing AI companies. Those applications are often "so close to the functionality of core foundation models that 'thin wrapper' labels could be thrown," in the report's words.

Why "Wrapper" Became an Insult

The label stuck because it points at three real weaknesses.

No differentiation. If two products call the same model with similar prompts, the customer sees two similar outputs. A16z made this point early, in a January 2023 analysis of the generative AI stack. It noted that many apps are "relatively undifferentiated, since they rely on similar underlying AI models and haven't discovered obvious network effects."

Fragile retention. Curiosity drives early usage, and curiosity fades. Sequoia's Generative AI's Act Two, published in September 2023, found that generative AI apps had a median of 14% daily-to-monthly active users, against 60% to 85% for strong consumer apps. Their conclusion: the biggest problem isn't demand or distribution, it's proving value.

Dependence on a supplier who is also a competitor. The provider that sells you the model can see what works and move into your space. We'll come back to that risk below.

The Wrapper Spectrum

It helps to stop treating "wrapper" as a yes-or-no label. Products sit on a spectrum, and the further right you go, the more the value comes from things a model provider can't ship by itself.

Stage	What the product is	Where the value sits	Typical exposure
1. Thin wrapper	A prompt behind a clean interface	Convenience and a pre-written prompt	Copied in days; absorbed by the model's own interface
2. Templated tool	Prompts plus templates, saved history, light formatting	Speed for one repeated task	Copied in weeks; low switching cost
3. Integrated tool	Connects to the customer's systems and pulls their context	Context and integrations	Harder to copy; needs real engineering
4. Workflow product	Owns a multi-step job from trigger to result, with review and approval	The workflow itself and the data it generates	Switching means retraining a team
5. System of record	Becomes where the work is stored, tracked and audited	Data, habit and process	Switching means migrating the business

A legal drafting tool that only rewrites clauses sits at stage 1. The same tool that pulls from a firm's past agreements, follows its approval chain, tracks redlines and logs who signed off is closer to stage 4. The model underneath might be identical. The business is not.

Bessemer's report points at the same shape in a different vocabulary. It argues that for application founders, "context and memory may be the new moats," and it ties vertical AI defensibility to "domain expertise: integrations, data moats, and multimodal interfaces built for vertical-specific needs." Notice that none of those items is the model.

What Makes a Product More Than a Wrapper

Four assets, alone or combined, move a product away from stage 1. None of them is a better prompt.

Proprietary data and feedback loops

If every customer interaction teaches the product something competitors can't access, the product improves in a way a copycat can't match. Corrections, approvals, rejected suggestions and outcomes are all signals. But be careful here. Sequoia's Act Two piece walked back the early optimism, saying the "data moats" are "on shaky ground" because the data application companies generate doesn't automatically create an insurmountable barrier. Data only helps when it's genuinely proprietary and genuinely improves results. If you want to see how retrieval changes this, read about RAG for AI agents, where a customer's own documents become the differentiator.

Workflow depth

The same Sequoia piece found something more encouraging: "workflows and user networks seem to be creating more durable sources of competitive advantage." A product that owns a whole job (intake, drafting, review, approval, handoff) is embedded in how a team works. Ripping it out means changing the process, not just swapping a tool.

Integrations

Connecting to the systems where work already lives is dull engineering, and that's exactly why it protects you. Each integration is a small barrier: authentication, permissions, edge cases, error handling. A model provider's general chat interface will rarely build and maintain those connections for one narrow industry.

Distribution

Distribution is the asset founders under-rate. Owning a channel, a partner network, a marketplace position or a trusted brand in a narrow segment means you reach buyers before anyone else does. A wedge strategy is the usual way in: win one painful use case, then expand.

Key Facts

Bessemer's August 2025 report separates fast-growing "Supernovas" from steadier "Shooting Stars." It reports Supernova gross margins of roughly 25%, "often negative," against about 60% for Shooting Stars (Bessemer, State of AI 2025).
The same report puts Supernova annual recurring revenue per full-time employee at about $1.13 million, which it describes as four to five times a typical SaaS benchmark, while Shooting Stars average roughly $164,000 (Bessemer, State of AI 2025).
Sequoia found a median of 14% DAU/MAU for generative AI apps, compared with 60% to 85% for established consumer apps (Sequoia, Generative AI's Act Two, September 2023).
A16z reported application gross margins "as high as 90% in a few cases but more often as low as 50-60%," driven largely by model inference costs (a16z, January 2023).
Enterprises purchased 76% of AI use cases in 2025 rather than building them, up from 53% in 2024. Enterprise AI spend reached $37 billion, with $19 billion (51%) going to applications (Menlo Ventures, December 2025).

The Risks That Come With Being a Wrapper

The provider moves up the stack

A model provider that sees millions of prompts knows which use cases are popular. When a use case is simple enough, the provider can add it as a built-in feature. Your product then competes with a free or bundled version from the company that supplies your engine. This is a structural risk, not bad luck, and it's the main reason a thin wrapper rarely survives. The difference between building on a model and competing with the model's owner is covered in application layer vs model layer.

Commoditisation of the output

When many products produce similar results from similar models, buyers have little reason to prefer one and tend to compare on price. That pressure lands on margins, which is why undifferentiated apps struggle to hold pricing.

Thin margins from inference

Because every request costs money to serve, a wrapper's gross margin depends on how much it pays the model provider per use. Bessemer's roughly 25% average margin for its fastest-growing group shows how far some companies trade profit for speed. That can be a deliberate bet on growth. It's a dangerous one if retention is weak.

Growth that looks like loyalty but isn't

Bessemer warns that fast adoption can either belie low switching costs or signal massive novelty. In both cases the revenue "may appear vulnerable." High growth with high churn is not a business yet.

The Opposite Mistake: Assuming "Wrapper" Means Doomed

It would be too neat to say every product on a third-party model is a dead end. Three counterpoints matter.

First, the model is a component, not the product. A company that buys cloud servers isn't dismissed as "a server wrapper." What counts is the value added on top.

Second, a thin product can be a fine starting point. Many strong products began as a narrow tool that got people in the door, then built integrations and workflow depth over time. The wrapper label describes where you are, not where you must stay.

Third, buyers don't care about architecture. They care about outcomes. Menlo Ventures' December 2025 report found that 76% of enterprise AI use cases were bought rather than built, which tells you buyers are willing to pay for finished solutions. A product that finishes a job well doesn't need to apologise for its supply chain.

The real test is whether a customer would be worse off, in a way they can measure, if your product disappeared and they went back to using the raw model. If the answer is "not really," you have a wrapper. If the answer involves lost context, broken integrations and a retrained team, you have something more.

A Quick Self-Check

Run any AI product, yours or one you're evaluating, through these questions.

Replacement test. Could a customer reproduce 80% of the value with the model's own interface and a saved prompt?
Context test. Does the product use data the customer would have to rebuild elsewhere?
Workflow test. Does it own a complete job, or only one step?
Integration test. Does it connect to the systems where the work already happens?
Distribution test. Does it reach its buyers through a channel competitors can't easily copy?
Switching test. What would a customer lose by leaving?

Two or fewer "yes" answers means you're closer to the wrapper end. Four or more means you're building the kind of economic moat that survives a model upgrade. The related question of whether to focus on one industry or serve many is covered in vertical AI vs horizontal AI, and how these choices show up in pricing and revenue is in AI startup business models.

What This Means for Founders and Owner-CEOs

If you're building a product, plan the path along the spectrum before you plan the launch. Ask what you'll own at stage 3, 4 and 5, and what you'll build first to get there. A thin start is acceptable if the route out is concrete.

If you're buying AI tools for your own team, the same logic helps you pick suppliers. A tool that lives inside your systems and carries your context is harder to leave, which cuts both ways: it's more useful, and it's more costly to abandon. Weigh that before signing. A buy vs build framework is a good companion here.

And if you're comparing a product's claims, remember that network effects, where each new user makes the product better for the next, are rare among AI applications. A16z noted that applications "haven't discovered obvious network effects," and where they do appear they're worth far more than any prompt.

Frequently Asked Questions about AI Wrappers and AI-Native Products

What is an AI wrapper?

An AI wrapper is a product whose main value is an interface and a set of prompts on top of a third-party model. The model does the core work, and the product mostly packages it. The term is often used as a criticism because that packaging is easy to copy.

Is every product built on a third-party model a wrapper?

No. Almost every AI product uses a model it didn't build. What matters is what the product adds: proprietary data, workflow depth, integrations and distribution. A product that owns a whole job inside a customer's systems is far more than a wrapper, even if the model underneath is the same.

Why do investors worry about thin wrappers?

Bessemer's State of AI 2025 notes that some of the fastest-growing applications sit so close to foundation model functionality that the thin-wrapper label could apply. Their revenue may be vulnerable because fast adoption can signal low switching costs or novelty. Weak retention and thin margins make that growth hard to sustain.

Are data moats enough to defend an AI product?

Not on their own. Sequoia's Generative AI's Act Two argued that the data application companies generate doesn't automatically create an insurmountable moat, and that workflows and user networks looked more durable. Data helps most when it's genuinely proprietary and measurably improves results.

Can a wrapper become a real product?

Yes, and many do. A narrow tool can win a painful use case, then add integrations, workflow ownership and stored customer context. The wrapper label describes a starting point. The risk is staying there while a model provider adds the same feature.

What's the fastest way to test whether a product is a wrapper?

Ask whether a customer could get about 80% of the value from the model's own interface and a saved prompt. If yes, the product is thin. If leaving it would mean losing context, rebuilding integrations and retraining a team, it has real depth.

About the author

Brian Tr

Co-Founder & COO

Brian Tr is Co-Founder and COO of Rework, with 12+ years in B2B go-to-market and operations. Brian scaled Rework from 0 to 10,000+ B2B customers across CRM and productivity tools. Brian writes for founders and owner-CEOs: startup fundamentals, founder-led and family businesses, partnerships, and how SaaS, marketplace, AI and EdTech companies grow.

View full profile LinkedIn