Is it too late to build AI apps?

Why does it feel like everyone already built the obvious AI apps?

Building an AI app today feels late because software supply has exploded, and the easy ideas are genuinely crowded.

RevenueCat’s latest study of more than 115,000 subscription apps shows how extreme the change has been. Roughly 2,000 new subscription apps were launching each month in early 2022. The figure recently passed 14,700 a month, an increase of more than 7x. RevenueCat says the steepest acceleration on iOS began in early 2025, around the same time AI-assisted coding tools became much more capable.

AI is already everywhere inside that flood. RevenueCat classifies 27% of subscription apps as AI-powered. The share reaches 61% in Photo & Video and 41% in Productivity. Sensor Tower separately counts more than 200,000 mobile apps whose store descriptions mention AI.

That has killed one of the easiest advantages of the first generative-AI wave. Connecting a good model to a clean interface once made a product feel unusual. These days, users expect AI features inside ordinary software.

Established companies have also caught up. Canva, Notion, Adobe, Microsoft, Google, Grammarly and dozens of other incumbents can put new AI features in front of an existing audience immediately.

So the feeling of being late has a real basis. The scarce thing is no longer the ability to build an AI feature. The question is whether enough new demand is appearing to absorb all this new supply.

Is demand for AI apps still growing today?

Yes. Demand for AI apps is still growing fast enough that calling the market mature would be premature.

Sensor Tower estimates that people spent more than $4 billion through in-app purchases on AI apps during the first half of 2026, 36% more than during the previous six months. Time spent inside generative-AI apps was projected to rise from 17.2 billion hours in the first half of 2025 to 36 billion hours one year later.

Enterprise spending has grown even faster. Menlo Ventures estimates that companies spent $37 billion on generative AI in 2025, versus $11.5 billion one year earlier. About $19 billion of that went to AI applications rather than models and infrastructure.

That last number matters more for builders than the overall AI spending headline. Application spending went from roughly $4.6 billion to $19 billion in a year, which works out to just over 4x growth.

We therefore have two trends happening together. Far more AI products are being built, but the amount of money flowing into AI software is also expanding at an unusual pace.

AI demand indicator	Earlier level	Latest reported level	Change
Consumer GenAI time spent, first half	17.2B hours	36B hours projected	~2.1x
AI app in-app spending	Previous half-year	>$4B	+36%
Enterprise GenAI spending	$11.5B	$37B	~3.2x
Enterprise AI application spending	~$4.6B	$19B	~4.1x

Is the AI app market already too crowded for a new startup?

The AI app market is brutally crowded, but the current numbers still show room for exceptional new companies.

RevenueCat gives us a useful way to separate those ideas. Median monthly recurring revenue growth across subscription apps was only 5.3% year over year. The top 10%, however, grew more than 306%.

Those outcomes barely resemble each other.

Competition has therefore produced a very wide gap between ordinary apps and the few products that really catch on. RevenueCat describes the market as “winner-take-more,” and the underlying numbers support that description: its top quartile grew more than 80% while its bottom quartile shrank by 33%.

Older apps still have a huge advantage. Products launched before 2020 generate 69% of the subscription revenue in RevenueCat’s dataset, while apps launched in 2025 or later generate just 3%.

Some of that gap is unavoidable because older apps have had years to collect subscribers, refine pricing and build distribution. Still, it tells us something important about the market today: launching is becoming almost trivial compared with building a customer base that compounds.

Crowding has made mediocrity easier to produce and harder to monetize.

Are people actually paying for standalone AI apps?

Yes. People currently pay more per customer for AI apps than they do for conventional subscription apps.

RevenueCat finds that an AI-powered app generates median realized lifetime value of $18.92 per payer after 30 days, compared with $13.59 for a non-AI app. After one year, the figures rise to $30.16 and $21.37.

That leaves AI apps with roughly 41% more revenue per payer after a year.

AI products also convert downloads into paying users slightly better at the median: 2.4% versus 2.0%. Trial-to-paid conversion shows an even bigger difference, at 8.5% for AI apps against 5.6% for non-AI apps.

This makes one popular version of the “AI apps are over” argument hard to defend. Customers have not collectively decided that AI should simply be free inside ChatGPT.

They are paying separately when an application solves a specific problem well enough.

The tougher issue appears after the payment.

Do people keep paying for AI apps after the novelty wears off?

AI app retention is currently weak, and this is one of the strongest reasons to be cautious about the market.

RevenueCat finds lower 12-month retention for AI apps at every major subscription duration. Among monthly subscriptions, only 6.1% of AI subscribers remain after one year, compared with 9.5% for non-AI apps. Annual subscriptions retain 21.1% versus 30.7%. Refund rates are also higher for AI apps, at 4.2% compared with 3.5%.

This creates an unusually clear pattern. AI products attract payment well, charge successfully and generate more revenue per payer, but customers disappear faster.

That probably explains part of the gap between impressive AI launch numbers and the smaller number of companies that develop lasting revenue.

A user may happily pay $20 to try an AI headshot generator, research assistant or document tool. Keeping that subscription active for twelve months requires something more repetitive and valuable.

The strongest AI products now have to survive a second test after the initial “wow”: does this software become part of someone’s normal work?

Subscription metric	AI apps	Non-AI apps
Download-to-paid conversion	2.4%	2.0%
30-day revenue per payer	$18.92	$13.59
One-year revenue per payer	$30.16	$21.37
Monthly-plan retention after one year	6.1%	9.5%
Annual-plan retention after one year	21.1%	30.7%
Median refund rate	4.2%	3.5%

Will ChatGPT, Gemini and Claude eventually crush standalone AI apps?

ChatGPT, Gemini and Claude will probably crush plenty of generic AI utilities, but they are nowhere close to absorbing every useful application.

The pressure is real because the major assistants keep expanding their territory.

ChatGPT can search, research, create images, analyze files, write code, connect to outside services and run third-party apps. Google can put Gemini beside Gmail, Docs, Sheets, Search and the rest of its ecosystem. Claude has expanded deeply into coding and professional workflows. Microsoft can bring Copilot straight into software companies already use all day.

This creates an ugly position for products whose entire pitch is “ChatGPT, but for this one simple task.”

The platform companies have three advantages over those products: gigantic distribution, access to their own models and the ability to bundle new capabilities into subscriptions users already have.

Yet the application layer is expanding at the same time. Menlo Ventures estimates that more than half of enterprise generative-AI spending already goes to applications. Companies such as Glean, Harvey, Sierra and Lovable have continued growing while foundation models became dramatically better.

The boundary is becoming clearer. General intelligence gravitates toward large assistants. Deep workflows still leave a lot of space for specialized software.

Is building an AI wrapper still a viable business?

A basic AI wrapper is a bad place to build from today, although using somebody else’s model is completely normal.

The important distinction is how much work happens around the model call.

If an application receives a prompt, forwards it to a frontier model and reformats the response, competitors can reproduce most of the product quickly. The model provider can also add the capability itself.

Glean looks completely different despite using outside models. Its product connects company information, permissions, people, applications and workflows so an AI system can understand what is happening inside a particular organization. Glean recently passed $300 million in ARR after reaching $100 million only 15 months earlier.

Sierra also sits on top of foundation models, yet its agents connect to business systems and carry out customer-service work. The company entered its third year with more than $150 million in ARR and says more than 40% of the Fortune 50 now use its technology.

Lovable depends heavily on frontier AI too. Its value comes from turning natural-language instructions into deployable software through an increasingly complete creation environment. It passed $500 million in annualized revenue and recently said its platform hosts around 60 million projects.

Outside models are not the problem here. Interchangeable products are.

Can a new AI app still become a huge business quickly?

Yes. Some of the fastest revenue growth in software is currently happening in AI applications founded after the first ChatGPT boom.

Lovable is the clearest consumer/prosumer example. The company reported crossing $400 million in annualized revenue in February 2026 and $500 million a few months later. It also said users were creating around one million new projects each week. By August, Lovable had raised another $400 million at a $13.3 billion valuation.

Glean provides a more enterprise-heavy example. The company went from $100 million ARR to $300 million in about 15 months while nearly doubling its Fortune 500 customer count.

Harvey may be the most interesting recent vertical example. The legal-AI company had reported more than $100 million ARR in 2025. This year, Harvey said it added more than $100 million of ARR in a single quarter, and then another quarter produced the same milestone. In September 2026, the company raised $550 million at a $15.5 billion valuation. Harvey also says 80% of the Am Law 100 now use its software.

Legora, one of Harvey’s competitors, crossed $100 million ARR this year as well.

These are unusually large outcomes for companies operating in supposedly “late” AI application categories.

We should still be careful with annualized revenue figures, especially for usage-heavy AI businesses where one strong month can be multiplied by twelve. But ARR milestones from Glean, Harvey, Sierra and Legora show that the pattern goes beyond one unusually viral coding product.

AI app company	Recent reported scale	What customers are paying for
Lovable	>$500M annualized revenue	Software creation
Glean	>$300M ARR	Enterprise search, context and agents
Sierra	>$150M ARR	Customer-service agents
Harvey	Added >$100M ARR in a quarter	Legal and professional workflows
Legora	>$100M ARR	Legal AI workflows

Is enterprise AI a better opportunity than consumer AI right now?

Enterprise AI currently gives founders more ways to become difficult to replace, although consumer AI can still grow much faster at the beginning.

Menlo Ventures estimates that companies spent $19 billion on generative-AI applications in 2025. Horizontal applications took roughly $8.4 billion, departmental applications $7.3 billion and vertical applications $3.5 billion.

Coding alone represented around $4 billion of departmental spending. Customer support, legal, healthcare, sales, marketing, IT and HR have also developed meaningful AI software markets.

Enterprise products gain protection from all the messy things that surround the model: permissions, proprietary documents, integrations, audit requirements, company-specific workflows, security reviews and actions inside existing systems.

Harvey demonstrates how powerful that combination can become in one profession. The company says 80% of the Am Law 100 use Harvey, up sharply from 42% reported around a year earlier. Its customers now run tens of thousands of custom agents, and Harvey’s own customer research found that 92% of surveyed in-house legal teams reported faster turnaround after adopting the product.

Consumer applications have fewer integration barriers and can spread globally almost overnight. Their weakness is that users can also leave almost overnight.

That helps explain why consumer AI often produces the biggest viral charts while enterprise AI produces some of the strongest retention and revenue stories.

Which AI app ideas are getting commoditized fastest?

Generic writing, summarizing, chatting, image generation and other tasks that frontier models already perform directly are becoming difficult standalone categories.

The pattern keeps repeating. A startup finds an attractive model capability, turns the capability into a product and gains users. A few model releases later, the same action appears directly inside ChatGPT, Gemini, Claude or an incumbent application.

Image generation is a good example. Standalone products once benefited enormously from model scarcity. Today, users can generate and edit images inside major assistants, design applications and social products without deliberately seeking out a specialist AI image tool.

Office productivity is heading the same way. Google can put Gemini into Docs and Sheets. Microsoft can put Copilot into Word and Excel. Anthropic has pushed Claude toward spreadsheets and professional documents. OpenAI has also been moving ChatGPT deeper into work applications and connected data.

A product that simply writes an Excel formula is therefore standing directly in the path of Microsoft, Google, OpenAI and Anthropic.

More complex spreadsheet workflows can still become valuable companies. Reconciliation, forecasting, audit trails, workflow automation and domain-specific financial work require more than generating a formula.

We can use that distinction almost everywhere. Features are being commoditized quickly. Complete jobs are much harder to absorb.

What can an AI app actually own if everyone uses the same models?

The strongest AI apps increasingly own context, workflow, customer access or accumulated work rather than the raw intelligence underneath.

Glean owns connections into an organization’s information and permissions. Harvey accumulates legal workflows, customer-specific knowledge and integrations around professional work. Lovable stores projects that users continue building and publishing. Sierra connects AI agents to the systems required to resolve actual customer requests.

Those assets survive model improvements.

A better Claude or GPT model may even strengthen an application like these because the product can adopt the better model without asking customers to abandon the workflow they already use.

This is why proprietary training data is sometimes overrated in discussions about AI moats. A startup does not automatically become defensible because it owns a large dataset.

Knowing what a specific customer is trying to accomplish, what happened previously, which systems contain the relevant information, what actions are allowed and what has to happen next can be much more useful.

The applications that own that context can keep upgrading the intelligence behind them.

Is distribution now harder than building the AI app itself?

For most new AI apps, getting attention is currently harder than producing the first credible version of the product.

The 7x increase in subscription-app launches makes that problem visible. Coding agents and vibe-coding products reduce the amount of engineering required to reach the market, but every competitor receives essentially the same benefit.

That shifts the bottleneck toward discovery.

The first wave of AI startups benefited from enormous curiosity around anything that looked like generative AI. That source of free attention has weakened. “Powered by AI” barely differentiates a product these days.

New distribution surfaces are appearing, though. ChatGPT now allows developers to build applications that run directly inside conversations through its Apps SDK. OpenAI says those apps can connect to an existing backend, and the architecture uses MCP rather than requiring the whole product to live inside ChatGPT.

That could create a meaningful acquisition channel for software that naturally appears during a conversation. Claude and other AI platforms are also expanding their connector ecosystems.

Still, building entirely around one platform would create the same dependency problem that developers have faced with app stores and social networks for years.

The better position is to own the customer relationship somewhere and use AI platforms as additional ways to reach that customer.

Can a solo founder or tiny team still compete in AI apps?

Small teams have more leverage today than at almost any previous point in software, but that advantage mostly helps with production rather than distribution.

AI coding tools have reduced the amount of engineering needed to build, debug and maintain a product. The same tools can help with design, customer support, research, documentation, marketing and internal operations.

Some current companies show how far revenue per employee can stretch. Lovable passed hundreds of millions of dollars in annualized revenue with a headcount that would have looked absurdly small for a traditional software company at the same scale.

Open-source projects can also reach enormous audiences very quickly when they solve the right problem. OpenClaw, originally built by Peter Steinberger as a side project, exploded across GitHub before OpenAI acquired the project.

Yet the small-team advantage has an obvious limit: everybody else has the same AI leverage.

A solo founder with strong access to a niche market can be dangerous today. A solo founder building the fiftieth generic AI utility without any idea how to reach customers is simply able to fail faster.

Do you need your own AI model to build a defensible AI app?

No. Most new AI apps would probably waste time and money trying to build a foundation model when the valuable part of the business sits elsewhere.

Glean supports multiple underlying models. Sierra builds customer-service agents on top of frontier AI. Lovable uses outside models while also adding its own specialized model work where that improves the product.

Even Harvey, which has recently started investing more deeply in post-training and specialized models, became a major legal-AI company before owning a foundation model.

The common pattern is pragmatic. These companies buy general intelligence where the market already provides it, then invest internally where specialization gives them an advantage.

For a new startup, that usually means using the best available model first and finding out whether customers genuinely care about the product.

Model ownership becomes worth discussing later if it materially improves cost, latency, reliability, privacy or task performance.

Training a model simply so the company can claim to own AI is an expensive distraction.

Where are the best AI app opportunities now?

The strongest AI app opportunities are increasingly found in expensive, repetitive jobs where existing software still leaves people doing too much work manually.

That can mean legal work, customer service, healthcare administration, financial operations, sales research, procurement, compliance, logistics, specialized engineering or countless smaller professional niches.

The current app data also suggests we should be careful about assuming that every category has already absorbed AI equally. RevenueCat finds that 61% of Photo & Video subscription apps are AI-powered and 41% of Productivity apps are. The proportions fall to 19% in Business, 12% in Travel and only 6% in Gaming.

Lower penetration does not automatically mean better economics. Some jobs simply benefit less from generative AI.

But it does show how uneven the market remains.

The interesting founder question today is therefore rarely “What new thing can an LLM do?” Frontier labs answer that question every few months.

A better question is “Where are people still wasting hours or paying somebody a lot of money because the existing workflow is bad?”

That gives an AI startup something more durable to attack than novelty.

Would waiting another year make building an AI app easier?

Waiting will probably give founders better models and better tooling, while also giving competitors another year to build customer relationships and distribution.

There is little reason to expect the underlying technology to stop improving. Models will get cheaper, faster and more capable. Agent tooling will improve. Building an application will probably become easier again.

That sounds like a reason to wait until we consider who receives those improvements.

Every competitor receives them too.

Meanwhile, the companies that find real product-market fit today can spend the next year collecting users, integrations, workflow data, brand recognition and recurring revenue.

The distribution landscape is also still being created. ChatGPT’s app ecosystem remains young. MCP is spreading across AI products. Agent interfaces are still changing. New product categories such as vibe coding reached large-scale adoption remarkably quickly after the original chatbot wave.

There will almost certainly be better AI technology one year from now.

There is much less reason to believe there will be more uncontested customer problems.

So, is it too late to build AI apps?

No. It is still a very good time to build the right AI app, while the window for generic AI utilities is closing quickly.

The evidence is stronger than the pessimistic narrative suggests.

People are spending billions of dollars on AI apps. Consumer usage is still climbing quickly. Enterprise application spending has grown to roughly $19 billion. New companies founded during the supposed AI-app gold rush have subsequently crossed $100 million in recurring revenue, and several have gone far beyond that.

At the same time, competition has changed dramatically. Roughly 15,000 new subscription apps can appear in a month. More than a quarter of subscription apps already use AI. Frontier-model companies keep adding functionality that previously supported standalone startups. And, as seen above, AI subscriptions currently lose customers faster than conventional software subscriptions.

That makes the dividing line fairly clear.

Simply connecting a frontier model to a familiar task is becoming a weak business. Plenty of those products will still make money, but they will live under constant pressure from copycats, model improvements and platform bundling.

The bigger opportunity is moving into the actual workflow: understanding the customer’s context, connecting the necessary systems, taking useful actions, remembering what happened before and becoming part of how the work gets done.

Glean, Harvey, Sierra and Lovable all reached serious commercial scale without having to become the next OpenAI. Their products use foundation models as one part of a much larger system.

That is probably the most important change for anyone starting now.

The first AI-app wave rewarded founders simply for getting powerful models into people’s hands.

Today’s market asks for much more. A new AI app needs a reason for customers to choose it after AI itself has become ordinary.

There are still an enormous number of those businesses left to build.