Perhaps in 2024 and 2025 we were all a bit too gullible. The predictions came quickly: by 2026, enterprise AI would be integrated into every workflow, productivity would increase by double digits, and every CIO had a “land and expand” strategy that included generative AI. The figures presented to investors eighteen months later reveal a far more subdued narrative. According to The Economist’s widely cited analysis from last year, only roughly 10% of businesses are using AI in a way that actually changes output. The Federal Reserve’s own monitoring recently confirmed this statistic. Pilots frequently find themselves caught between the IT ticket queue and the proof-of-concept deck.
Software stocks are currently experiencing a loud bleed from that disappointment. In the first quarter of 2026, the iShares Expanded Tech-Software ETF dropped more than 24%, marking the index’s steepest three-month decline since late 2008. Despite ServiceNow’s 21% increase in subscription revenue on January 29, the stock fell by nearly 10%. By any reasonable measure, Microsoft’s Azure grew by 39% that same day, and it was also penalized. The earnings didn’t matter to the tape. Something deeper was being repriced.
These sessions give the impression that investors finally calculated AI margins and recoiled. Because the marginal cost of an additional user was nearly zero, the SaaS model was elegant. In contrast, each AI query uses actual GPU processing power that must be paid for. Approximately 70% of software vendors now acknowledge that AI features are lowering gross margin in surveys, phone conversations, and occasionally in whispers to analysts. Microsoft, Oracle, Adobe, and Salesforce are all in correction territory. After years of discussion, the shift to usage-based pricing is now essential for survival.
However, the bear case seems overly neat. The reason the adoption curve is flattening is evident to anyone who has actually attempted to implement an AI agent within a Fortune 500 company. The technology isn’t the problem. Middle management, legal review, HR anxiety, data stored in four separate CRMs, and a compliance team that demands that every prompt be recorded before anything is put into production are all involved. In a piece published last year, Brian Solis referred to it as “authority friction”—a helpful term. It’s remarkable how recognizable the pattern is as you watch this play out. ERP systems, PCs, tractors, and the cloud itself all lagged years behind their hype cycles.
According to Goldman Sachs, the market has overshot. Analyst Matthew Martino described the sell-off in a note dated April 11 as a “rapid shift in sentiment rather than a sudden deterioration in fundamentals.” His six-factor framework, which includes orchestration risk, monetization exposure, system-of-record ownership, data-integration moat, AI execution, and budget alignment, is designed to distinguish between businesses that are genuinely at risk of displacement and those that are caught in the crossfire. The company claimed that the repricing has been implemented “broadly rather than selectively” and released a basket of four buy-rated names, including Salesforce. For 2028, market-implied software revenue growth fell from about 15% to just 5%. That would be a gift, even half of it.
The more somber voice of Morningstar isn’t quite as comforting. Although the incumbents’ retention rates and RPO metrics appear to be in good shape, analyst Dan Romanoff reduced the moat windows on a number of software names from twenty years to ten. He noted that it is more difficult to defend the likelihood of sustained excess returns beyond the next ten years. Romanoff still believes Salesforce and ServiceNow have a structural advantage for agentic AI, so it’s not a surrender, but it’s a significant hedge. Ten years of self-assurance turn into ten. There has been a significant shift in the market’s assessment of compound machines.
For the time being, Goldman is probably closer to the practical read than the bears. Businesses have never ripped and replaced software stacks in two years. By 2030, there won’t be enough inference capacity on the planet to significantly replace knowledge workers. Additionally, because the customer is already present, integrated, and renewing, companies with distribution—such as Salesforces and ServiceNows—continue to win agentic rollouts. The question of whether the old multiples will ever return is less certain. It’s difficult to ignore the fact that each previous software cycle concluded with a number of names that never fully recovered their previous valuations in addition to winners. A person is going to discover which side of that line they are on.
