Wall Street is convinced AI will kill SaaS. History and economics say something else

To many investors, the logic is simple: AI models, such as Anthropic’s Claude, will eliminate the need for specialized business software. Companies will just ask AI to perform the tasks currently handled by other vendors’ software, or they will ask AI coding agents from Anthropic, OpenAI, or Google, to “vibe code” bespoke software tailored to their needs. Either way, businesses will no longer need to buy expensive software from traditional vendors like Salesforce, ServiceNow, and Microsoft, or from the hundreds of smaller, niche firms offering legal tech, procurement systems, or financial planning software.

The doomsday view has sparked a heated debate, with critics insisting the selloff is overdone. Bank of America senior analyst Vivek Arya called investors’ souring on SaaS businesses “indiscriminate,” “overblown,” and “logically inconsistent.” Even as big an AI booster as Nvidia CEO Jensen Huang told CNBC that “the markets got it wrong.” Rather than ending the need for pre-built software, AI agents will use the software to accomplish tasks, he argued.

While the advent of AI is certainly a new development for investors to wrap their heads around, the fundamental question at the heart of the SaaSpocalypse debate is as old as innovation itself. Each new piece of technology opens new possibilities and leads to existing practices being reassessed. The factors that cause one innovation to supplant its predecessor while another technology evolves alongside an incumbent or follows a different path entirely, can vary and are hard to predict, according to experts in history and economic theory whom Fortune spoke to. In many cases, the expected outcomes—both the overly optimistic as well as the gloom-ridden—turn out to be widely off the mark.

Take video cameras, which came on the scene in the 1980s. Some predicted a proliferation of filmmaking, because video lowered production costs and the need for specialized knowledge. And there was an explosion of visual content. The total number of film titles released in the 1990s was double the figure for the 1970s, according to data from IMDB. But video for the most part barely impacted the production of box-office hits, which continued to be shot on film. The average cost of feature film production, rather than falling, tripled in constant dollars between 1980 and 2007, according to data from the Motion Picture Association of America. What actually happened is that video and film occupied almost completely different market niches: People turned to video for nontheatrical content, like educational and training films, but still looked to film and professional production for major studio releases.

Around the same time that neophyte directors were discovering video cameras, desktop publishing was also gaining traction—and leading to dire predictions of the death of commercial print shops. After all, companies could now do their own “prepress” design and layout work, and businesses could print their own documents using compact and ubiquitous color laser printers. There would be no need to hire a commercial printer.

And yet desktop publishing didn’t kill commercial printing. Instead, in the decade after the tech debuted, the number of people employed in printing and related support activities in the U.S. actually climbed, peaking at about 680,000 workers in 1998, while the number of commercial print shops rose to an all-time high of 62,000 in 1995, according to U.S. government statistics. Desktop publishing actually allowed more people to set up print shops. What it did eliminate was the need for specialized prepress roles like typesetters, pasteup artists, and color separation specialists. Those specific jobs did see marked declines throughout the 1990s. Meanwhile, the expanding number of print shops led to increased competition and declining margins. This eventually led to industry consolidation, as the only way to make any profit was through economies of scale.

Could something similar happen with software as AI lowers the barriers to writing code? Rather than killing off software vendors, it may be that many more companies will be formed to write specialized business applications because it will be less important to hire scarce, specialized coding talent. And, similar to printing, SaaS profit margins may shrink, forcing consolidation. But that won’t happen because AI ate SaaS. It will happen because AI fed SaaS.

Lorenz Ekerdt, an economist at the State University of New York at Stony Brook, says history over most of the past century has shown that firms have become ever more specialized and outsourced more ancillary functions. He does not see AI reversing this trend. “I think we will increasingly see niche software being used and produced. But if we ask, ‘Who’s going to make that software?’ It seems like it would still be the same firms,” he says. Why? Because software vendors already have all the processes and infrastructure set up to make and maintain software, the security protocols, the code reviews, etc. Companies outside the software business would have to reinvent these processes. And while AI might be able to automate some of that, there’s still a high probability non-software companies would mess this up.

To buy or to build?

The economic theory that determines why a company decides to buy something rather than build it was pioneered by British economist Ronald Coase, beginning in the late 1930s. Coase’s idea was that buy-vs.-build decisions were largely determined by “exchange costs,” or how easy it was to specify and enforce contractual terms for what the firm needed. For commoditized production inputs, exchange costs tend to be low, and so businesses tend to buy those things. For highly bespoke ones, or when the firm’s requirements are likely to change rapidly, suitable contracts are difficult to make at a reasonable cost, and so businesses tend to produce those inputs internally. Businesses also have to weigh the risk that a supplier will take advantage of them, using leverage to exact higher payments over time, against concerns that building the capability in-house can be done at the right quality, quantity, cost, and time frame.

In the case of AI and software, it is possible that companies will allow employees to “vibe code” small, artisanal productivity tools that make their jobs easier. But for critical-but-somewhat-commoditized systems where specialization is of limited competitive advantage (think accounting or human resources software), companies may find it more efficient to continue buying from established vendors. Such vendors can afford to employ the best coders and UX designers, and they have established processes for delivering product road maps and maintaining large codebases. In that scenario, the software market may bifurcate, with the most successful existing vendors continuing to thrive, while more niche application makers find themselves increasingly competing against in-house “vibe coded” solutions.

Oliver Hart, a Nobel Prize–winning Harvard University economist who researches the economic theory of the firm, says there are several exchange cost considerations in AI and software. On the one hand, AI ought to lower the cost for existing software vendors to customize their products for specific customers, making it less likely that companies will choose to build their own software.

However, Hart says, to the extent that software vendors profit in part by locking in their customers’ data, or by charging fees to move data off their platforms, those exchange costs may compel some companies to take over software development in order to gain more flexibility and control over future costs. Some companies might also experiment with building their own software with AI, if only to use the threat of internal software development to extract better contract terms from outside vendors. Different businesses may come to different conclusions. “It’s going to become very specific to a particular client,” Hart says. “I don’t think it’s completely obvious which way it goes.”

The conclusion to the AI-vs.-SaaS battle may also play out more slowly than many people expect. Companies almost always adopt any new technology gradually, in order to minimize risks, and they are unlikely to switch wholesale from buying SaaS products to writing their own code, says James Cortada, a professor at the University of Minnesota who specializes in the history of information technology and who worked for years at IBM. It has taken four decades for companies to switch from programming mainframes in COBOL to more modern programming languages, such as C++ and Java, Cortada says. And it took most of a decade for businesses to migrate from using on-site servers to the cloud, and even then many companies continued to keep their most vital data in servers they directly controlled.

When it comes to AI and software, Cortada thinks it far more likely that existing vendors will increasingly integrate AI into their software, and that their customers will use these features, rather than using AI to completely replace the off-the-shelf software. “Asking people to use AI to code something critical, like a payroll system or a procurement system, is like asking people to change flat tires on a car that continues to drive at 60 miles an hour,” Cortada says. “It’s too risky.”

This article appears in the April/May 2026 issue of  Fortune with the headline “The great software panic.”