AI is killing the app economy and building its successor

Artificial intelligence has made it easier than ever to build an app. New releases on Apple Inc.’s App Store surged by about 80% earlier this year as developers vibe coded their way to new software widgets for markets such as wellness and productivity.

But follow the money in Silicon Valley and something strange is happening: Mobile apps are disappearing from some of tech’s most promising startups. AI is the reason for that, too.

San Francisco-based Y Combinator is the world’s most prestigious startup accelerator — a short program that helps new firms grow quickly by providing money, guidance and connections across the tech industry — with an estimated 10,000 teams applying every three months from all over the world. Typically, only about 1% get in. Alumni include the founders of Airbnb Inc., Stripe LLC, Coinbase Global Inc., DoorDash Inc. and Reddit Inc.; OpenAI Chief Executive Officer Sam Altman was one of its first participants. 

Of the 195 startups in the program’s last spring batch, only eight had a native mobile app as their main product, according to a Y Combinator spokeswoman. Go back to 2013 and around 15% of Y Combinator startups were in the app business, with the rest building marketplaces, crowdfunding platforms or on-demand delivery services. By 2016, that proportion had declined to 8%, before hovering around 4% between 2019 and 2022. In the last three years the average has dropped to between 0.5% and 1%. 

What gives? "Apps and SaaS dead," one venture capitalist, referring to software-as-a-service, texted me recently when I asked. "AI and agentic hot." 

In other words, the flood of entrants into app stores belies startups’ expectations for the most lucrative sources of future revenue: Building skills and agent workflows on top of AI models. About three years ago, these types of software were known as GPT-wrappers, because they "wrapped around" a model like OpenAI’s to do hyper-specific things like writing marketing text or acting as a companion. Many products were crushed as soon as the AI firms added the same features, echoing the experience of a bunch of companies building apps for Apple’s iPhone.

One term for the latest iteration of such products is a "harness," which doesn’t just pass text back and forth from an AI model but also manages its behavior as it carries out tasks. This year, that kind of work is driving at least one in six of the companies coming out of Y Combinator — building tools that help AI models like Claude or ChatGPT manage money or order food, while testing their work or preventing them from making mistakes or being hacked. Startup Salus, for instance, can stop an AI agent from accidentally issuing a refund; another firm called Allowance gives the software a temporary credit card number to order food or buy other goods without exposing a user’s real banking digits. Both are based in San Francisco.

Chaz Englander is a British entrepreneur who’s had three startups go through the Y Combinator program. While his first two were mobile apps (one for renting items from people nearby, the other for buying groceries), his latest guides AI models to process financial paperwork or do research. It can switch between several models to which a bank subscribes, using them, for example, to create a PowerPoint presentation by delegating different parts of the process such as extracting data or building a chart to the platform that’s best at each task. 

Building such cross-model software makes sense, Englander says. "I do see a world where these tools will become much more centralized into a single system," he says. "Everything is being integrated." Instead of navigating between different apps to do different things, people will increasingly use an AI assistant with access to multiple models — hence the race to augment those agents with specific capabilities. 

Mark Zuckerberg, for instance, is trying to create his own gateway for doing everything with Muse, a "personal AI agent" that Meta Platforms Inc. launched this week in the U.S. You can chat to it from inside WhatsApp to book travel, send email or buy things online. "Muse doesn’t just answer questions," Zuckerberg said in a video announcement. "It gets things done for you." The Meta CEO said he used it to organize baking projects with his kids and to monitor his mixed martial arts training.  

This centralization of control looks different to what you’ve probably experienced for more than a decade with mobile operating systems: it’s meant to replace a dizzying array of apps you never use, scattered across several pages on your smartphone. The vision of funneling more capabilities into Open AI’s ChatGPT, Anthropic PBC’s Claude or Google’s Gemini means less swiping and more efficiency — but it also points to a future where the controllers of these platforms have unprecedented power. 

That will play out in how companies like Englander’s latest startup Model ML figure out how to make money. One venture capitalist tells me the road to monetization for AI skills and agent workflows is still uncertain. It’s not as simple as the old app method of charging a subscription fee or making money from ads. Neither is it like enterprise software, where businesses charge "per seat," as a monthly fee for every employee with a login.    

Instead, many startups are adopting a "usage-based" approach, where customers like banks and law firms pay per task performed by their technology, while the startup acts as the middleman buying tokens wholesale from OpenAI or Anthropic and then billing its corporate clients for the finished work. 

The trick is to perform those tasks better than anyone else. "We believe you’ll get better intelligence from us for $100 versus the nearest competitor and the labs, and that’s where our margin comes from," says Englander. 

This new breed of startups has its work cut out. Previously, software businesses were beholden to app stores managed by Apple and Alphabet Inc.’s Google. Now, they’re becoming dependent on the decisions of OpenAI, Anthropic and (once again) Google. Gatekeeping is moving from controlling apps to overseeing AI services — but it isn’t going away. 

This column reflects the personal views of the author and does not necessarily reflect the opinion of the editorial board or Bloomberg LP and its owners. Parmy Olson is a Bloomberg Opinion columnist covering technology. A former reporter for The Wall Street Journal and Forbes, she is author of "Supremacy: AI, ChatGPT and the Race That Will Change the World."