Do people still pay for AI wrappers?

Why are people suddenly saying AI wrappers are dead?

AI wrappers look much easier to kill today because ChatGPT, Claude and Gemini keep absorbing jobs that once supported entire standalone products.

A few years ago, putting a friendly interface around a large language model could create a genuinely different experience. Separate products emerged for chatting with PDFs, searching the web, rewriting emails, summarizing meetings, generating images, researching topics and writing code.

That gap has narrowed dramatically. ChatGPT can work with files, browse the web, conduct longer research, generate and edit images, write and run code and connect to outside services. Claude has added connectors, Artifacts and increasingly capable coding workflows. Gemini combines research, Google data and an interactive workspace for creating documents, applications and other outputs.

So the concern behind the “AI wrappers are dead” argument is legitimate. A startup can launch around a missing model feature and wake up later to find that OpenAI, Anthropic or Google has added something close to it.

The strange part is that customers are spending more money on AI software at the same time. Recent consumer-app data shows record spending, while enterprise spending on the AI application layer has reached tens of billions of dollars.

Customers clearly still pay. The much harder question is what they are still willing to pay for.

What actually counts as an AI wrapper today?

The phrase “AI wrapper” has become so broad that it can describe both a disposable prompt interface and a company doing hundreds of millions of dollars in revenue.

At the thin end, an AI wrapper takes a user's request, sends it to OpenAI, Anthropic or another model provider, adds some instructions and displays the answer. Most of the useful intelligence comes from the model, and a reasonably skilled user can reproduce much of the result inside a general chatbot.

Then consider Gamma. Gamma also depends on foundation models, yet users work inside a dedicated system for creating and editing presentations, documents and websites. The product handles layouts, themes, visual structure, collaboration, publishing and sharing. Gamma said it crossed $100 million in ARR while profitable, and its careers site now says the product has passed 100 million users.

Harvey stretches the definition even further. Harvey uses foundation models inside legal and professional-services workflows, with enterprise controls, organizational knowledge, document systems and specialized tools around them. Harvey said recently that it added more than $100 million of ARR in a single quarter.

Calling all three products “wrappers” therefore tells us very little. The useful question is how much value the application adds around the underlying model.

Type of AI app	What the customer is really buying	Exposure to model providers
Thin prompt wrapper	Easier access to a model	Very high
Focused AI utility	A specific finished result	High to medium
Workflow product	Editing, context, integrations and repeated work	Medium
Vertical AI platform	Domain workflows, data, controls and expertise	Lower

Are consumers still paying for AI apps right now?

Yes. Consumers are currently spending more on generative-AI apps than at any previous point, and the increase is far too large to dismiss as a few successful chatbots.

Sensor Tower estimates that generative-AI mobile apps generated about $6.1 billion in in-app revenue over the twelve months ending in Q1 2026. That was 232% higher than the previous comparable period.

The longer trajectory is even clearer. Quarterly generative-AI app revenue rose from less than $60 million in Q1 2023 to about $1.9 billion in Q1 2026, an increase of more than 30 times in three years. Sensor Tower's broader State of AI report later estimated that AI-app in-app purchases would exceed $4 billion during the first half of 2026 alone.

Usage has continued to climb alongside spending. Global time spent in generative-AI apps was projected to rise from 17.2 billion hours in the first half of 2025 to around 36 billion hours one year later.

We can rule out the idea that consumers have stopped paying for AI software. The market is still expanding extremely quickly. What has changed is where that money is going.

Will people pay for specialized AI when ChatGPT can do something similar?

Yes. People still pay for specialized AI when the product gets them to the finished result faster than a general chatbot does.

Gamma is a good example because ChatGPT, Claude and Gemini can all help write a presentation. That has not stopped Gamma from building a business above $100 million ARR. The difference is the work between “write me a presentation” and actually having a presentation worth sending to someone. Gamma organizes the material, designs the slides, keeps them editable, hosts the result and lets users continue working inside the same product.

Photo AI shows the same behavior at a much smaller scale. General image models can already produce photorealistic portraits. Yet founder Pieter Levels disclosed in March 2026 that Photo AI was making about $105,000 per month in revenue and roughly $80,000 per month in profit.

The customer is paying to skip several steps. Instead of choosing a model, learning how to prompt it, handling bad generations, maintaining consistency and finding another tool to finish the job, the customer buys the outcome.

This has become more important as access to raw model intelligence gets cheaper. Convenience still has real monetary value when it removes enough work.

What killed the first wave of AI wrappers?

The first wave got into trouble when the expensive-looking feature they were selling became a standard feature of general AI assistants.

Jasper is the classic example. Before ChatGPT, Jasper gave ordinary users a much easier way to generate marketing copy with large language models. The company grew to roughly $80 million in ARR and raised money at a $1.5 billion valuation.

Then conversational AI became widely available through ChatGPT. Jasper suddenly had a direct competitor offering general-purpose writing inside the product that millions of people were already using.

The Information subsequently reported that Jasper lowered revenue expectations and cut its internal share valuation. The company later moved much harder toward enterprise marketing workflows rather than relying on broad AI writing as its core pitch.

The lesson was harsher than “wrappers are risky.” If the scarce thing you sell is access to a model capability, the advantage can disappear almost overnight once that capability becomes standard.

That danger is still here today, and the major model providers are moving much faster than they did during the first generative-AI boom.

Are ChatGPT, Claude and Gemini still swallowing AI wrapper features?

Yes. General AI assistants are currently moving deeper into workflows that used to belong to separate apps.

The first wave covered obvious features such as writing, summarization, document chat, coding and image generation. The newer features go further.

OpenAI has expanded ChatGPT into web research, connected data, coding, files, apps and increasingly agent-like tasks. Anthropic has pushed Claude further into software development, interactive Artifacts, integrations and computer-based work. Google has connected Gemini more deeply with Search, Gmail, Drive and other parts of its ecosystem.

This makes a certain kind of wrapper particularly fragile. Imagine a product whose entire experience is “upload a file, type a question, receive generated text.” Every major assistant can move closer to that workflow without changing its fundamental product.

The risk grows with each step the wrapper asks users to perform outside its own environment. If the useful part of the experience begins and ends with model inference, a model provider can eventually offer a close substitute.

Products that keep the user's actual work inside the application have more room to survive.

Does cheaper AI help or hurt AI wrapper businesses?

Cheaper AI helps established wrappers make more money, while making the next competitor cheaper to build.

OpenAI's GPT-4.1 pricing illustrates how quickly inference economics moved. GPT-4.1 launched at $2 per million input tokens and $8 per million output tokens, with GPT-4.1 mini at $0.40 and $1.60 respectively. OpenAI said the standard model was 26% cheaper than GPT-4o for a median query, and cached input became cheaper again.

For an application charging $20 or $30 per month, lower inference costs can improve gross margins substantially once usage becomes predictable. A company can also route easy requests to cheaper models and reserve expensive models for tasks where quality really matters.

The same economics lower the barrier for challengers. A competent founder can access strong language, image, speech and coding models without spending years training them. AI coding tools then reduce the cost of building the rest of the application.

Competition gets pushed toward the things model pricing does not commoditize so easily: customers, brand, workflow, proprietary context, integrations and distribution.

Do AI apps actually keep people paying?

AI apps currently have a retention problem: they are unusually good at getting the first payment and unusually weak at keeping the subscriber for a full year.

RevenueCat's 2026 subscription study covers more than 115,000 apps representing over $16 billion in revenue, so there is finally enough data to see the pattern rather than guess from individual startups.

AI apps produced 41% more first-year realized revenue per payer at the median: $30.16 versus $21.37 for non-AI apps. Their median download-to-paid conversion was also higher, at 2.4% versus 2.0%.

Retention tells a different story. Only 6.1% of monthly AI subscriptions remained after twelve months, compared with 9.5% for non-AI subscriptions. Annual AI subscriptions retained 21.1%, versus 30.7% for non-AI products. RevenueCat summarizes the gap by saying AI apps churn about 30% faster.

Refunds are also somewhat higher. The median AI-app refund rate was 4.2%, compared with 3.5% for non-AI apps.

That combination explains a lot of what founders see today. AI makes a strong demo extremely easy to sell. Turning the demo into a habit is much harder.

Metric	AI apps	Non-AI apps
Download-to-paid conversion	2.4%	2.0%
First-year revenue per payer	$30.16	$21.37
Monthly-plan retention after 1 year	6.1%	9.5%
Annual-plan retention after 1 year	21.1%	30.7%
Median refund rate	4.2%	3.5%

Why are companies buying AI apps if they can use the same APIs themselves?

Companies are buying more ready-made AI software, even though building an internal AI tool has become much easier.

Menlo Ventures estimates that enterprise generative-AI spending reached $37 billion in 2025, up from $11.5 billion one year earlier. About $19 billion went to applications: the software products sitting above models and infrastructure.

The build-versus-buy numbers are even more useful for our question. Menlo's survey found that 53% of enterprise AI use cases were purchased in 2024, while 47% were built internally. By 2025, the purchased share had jumped to 76%.

Companies therefore moved toward buying at the same time APIs, coding agents and open-source tools were making internal development cheaper.

The reason becomes clearer when we look at what companies actually need. Connecting an API is easy. Building permissions, integrations, monitoring, security, evaluation, user interfaces, organizational context, maintenance and support around that API is a different project.

The enterprise market has effectively answered the wrapper question with money. Companies are happy to pay another company to package AI into a reliable piece of work.

Enterprise AI measure	2024	2025
Total generative-AI spending	$11.5B	$37B
AI solutions purchased	53%	76%
AI solutions built internally	47%	24%
Spending on AI applications	—	$19B

What are businesses really paying for in an AI wrapper?

Businesses usually pay for the workflow around the model, especially when mistakes, permissions or fragmented data make the DIY version annoying.

Harvey makes that visible in legal services. A lawyer can already ask a frontier model to draft text or summarize a document. Large firms need much more around that interaction: access to internal knowledge, document workflows, controls, security and tools designed for actual legal work.

Harvey recently said Q2 2026 was its first quarter with more than $100 million of additional ARR. The company has also expanded beyond law firms into asset management, where it says more than 125 firms use Harvey for work including due diligence, data-room analysis and document review.

Coding tells us something similar at larger scale. Lovable takes models that can already write code and turns them into an environment where a user can create and modify an application through natural language. In June 2026, Lovable told TechCrunch it had passed $500 million in annualized revenue and was seeing roughly one million new projects created each week. Two months later, the company raised at a $13.3 billion valuation and said Lovable projects were attracting around 900 million visits per month.

Users in both cases have access to powerful general models. They still choose the specialized application because the application handles more of the job.

What actually makes an AI wrapper defensible now?

The strongest AI wrappers own enough of the user's work that swapping the underlying model is easier than swapping the application.

That can come from several places. Gamma keeps the presentation itself, its structure, design, editing history and collaboration inside Gamma. Harvey sits inside professional workflows and organizational knowledge. Coding products hold context about projects, files and development work. Sales products can accumulate information about prospects, campaigns and customer interactions.

A wrapper becomes much safer when the model behaves like one component inside that environment. The company can switch models, combine several providers or use different models for different tasks without making the customer rebuild the workflow.

Proprietary data helps, although it is hardly mandatory. Photo AI can make money without owning a unique corporate database. Gamma reached serious scale through product design, workflow and distribution. Enterprise products naturally have more opportunities to build around customer context because repeated use connects them to documents, systems and historical work.

The practical test is simple: imagine replacing the application's AI model tomorrow. If most of the customer value survives, the company owns something useful. If the whole product collapses, the model provider still owns most of the value.

Can a solo founder still make serious money with an AI wrapper?

Yes. A narrow AI wrapper can still be an excellent small business even when venture investors would find the moat unimpressive.

Photo AI gives us unusually transparent numbers. Pieter Levels disclosed $105,000 in monthly revenue and approximately $80,000 in monthly profit in March 2026. That is roughly $1.26 million in annualized revenue from a product built around external AI capabilities.

The economics are attractive precisely because a solo founder does not need to win an entire software category. A product doing $30,000, $50,000 or $100,000 a month with a tiny team can produce an exceptional outcome for its owner.

This is where a lot of online discussion gets confused. People use “bad startup” and “bad business” as though they mean the same thing.

A thin AI product with $1 million of annual revenue, high churn and limited defensibility may be difficult to finance at a huge venture valuation. The same product can give a founder enormous cash flow.

So a solo founder should care much more about acquisition cost, gross margin, churn and operational simplicity than whether the company sounds sufficiently defensible in a venture-capital pitch.

Is distribution now more important than the AI model?

For many new AI apps, yes. Getting attention has become harder than getting access to capable AI.

RevenueCat's latest data shows that roughly 27% of subscription apps are already classified as AI-powered. The share reaches 61.4% in Photo & Video and 41.1% in Productivity.

The broader app market is flooding with new products too. Around 2,000 new subscription apps launched each month in early 2022. RevenueCat says the figure had climbed above 14,700 per month by early 2026, roughly seven times higher.

Sensor Tower provides another measure of the same crowding: more than 200,000 apps now mention AI in their store descriptions.

AI-assisted development is making supply grow far faster than human attention. A founder can build the product in days and still spend months trying to find customers.

Products with built-in distribution have a huge advantage these days. Gamma documents and presentations get shared with other people. Lovable applications are published onto the web. Workplace tools can spread from one employee to a team. Specialized products can dominate narrow search queries where ChatGPT itself is less likely to be the first destination.

That is becoming one of the dividing lines between AI demos and AI businesses.

Which AI wrappers are hardest to charge for today?

Generic AI wrappers are becoming extremely difficult to monetize when the customer can reproduce the main result with one obvious prompt in ChatGPT, Claude or Gemini.

Generic AI writers sit in this danger zone because writing is already a basic capability of every major language model. Simple PDF chat tools face the same problem as file handling becomes standard. Basic research wrappers now compete with research modes inside general assistants. Plain image generators face increasingly strong multimodal products from the model companies themselves.

These products can still make money through SEO, mobile distribution, localization, better onboarding or a strong brand. The underlying position is simply much less comfortable than it was a few years ago.

The weakest products usually have another problem: the customer finishes the task and leaves. Resume generation, one-off headshots, document conversion and similar utilities can be profitable, yet forcing them into a monthly subscription often creates churn because the user has no monthly problem left to solve.

Current subscription data fits that behavior. AI apps sell very effectively at the beginning, then lose subscribers faster than traditional apps.

Founders building in these categories have to be unusually good at distribution or design the business around transactions, credits and repeat use rather than assuming every AI feature deserves a subscription.

Which AI wrappers are people happiest to pay for?

People are paying most willingly when an AI product removes a whole chunk of work rather than adding a prettier place to type a prompt.

Presentation software such as Gamma turns rough material into something a user can edit, present and share. Coding products such as Lovable turn natural-language instructions into working software and keep the project alive for further changes. Legal platforms such as Harvey connect AI with the documents, controls and professional workflows surrounding legal work.

Consumer products can win through a different route. Photo AI packages image generation around a clear result. Companion apps sell a persistent experience rather than a generic answer. Other specialized tools earn money by knowing exactly what the user wants before the user writes a detailed prompt.

The common feature is specificity. Good AI products make the customer think about the task rather than the model.

As AI models keep improving, that becomes a stronger requirement. Raw intelligence is increasingly abundant. Knowing what to do with the intelligence, where to place it and how to turn it into a finished result still has plenty of economic value.

So, do people still pay for AI wrappers?

Yes. People still pay a lot of money for AI wrappers today, but simple access to somebody else's model has lost most of its pricing power.

The evidence is unusually strong. Consumer spending on generative-AI apps is still climbing quickly. Subscription data shows AI apps generating more revenue per payer than non-AI apps. Enterprises are shifting toward purchased AI products rather than internal builds. Companies such as Gamma, Harvey and Lovable have reached revenue levels that would have sounded extraordinary for an “AI wrapper” only a few years ago. Solo products such as Photo AI show that the opportunity also exists far below venture scale.

The weak part of the market is much easier to identify now. A product that sends a prompt to a model and gives the answer back has almost no room for error. General assistants keep adding those capabilities themselves, model costs keep falling and competitors can recreate simple interfaces very quickly.

The businesses holding up best go further into the customer's job. They own the editor, the files, the integrations, the professional workflow, the data, the publishing step, the collaboration or the final outcome. Strong distribution helps even more.

So the answer has become sharper over time. People are still paying for AI wrappers in huge numbers. They are increasingly reluctant to pay simply for the wrapping.