On the stock ticker, the verdict on enterprise software looks settled. Shares in the likes of UiPath, Appian, Atlassian and Dynatrace surged this week, part of a broad rally built on a single reassuring idea: that generative AI is a catalyst for the software business, not a threat to it. The exhibit everyone points to is Salesforce, up around 20 percent, whose Agentforce product has reached $1.5 billion in annual recurring revenue and whose Slackbot assistant became the fastest-adopted product in company history. Okta says its new AI identity tools drove roughly 30 percent of fresh bookings. The story writes itself: buyers want AI, incumbents are selling it, everyone wins.
A quieter number complicates the picture. In a new McKinsey survey of more than 1,700 employees, nearly one in three respondents said their organisation had decided against buying at least one software feature because it could now build that capability in-house. Agentic coding tools, the report's authors note, have become a genuinely viable option for pulling software development back inside the company. Spend on off-the-shelf AI assistants is increasingly treated the way firms treat the phone bill, a cost of doing business, while the more interesting work migrates to homemade systems.
Both things are true at once, and the tension between them is the real story. AI is expanding the enterprise software market and dissolving parts of it in the same motion. For every feature a company happily rents from Salesforce, there may be another it used to license and now simply assembles itself over a weekend. The vendors winning today are the ones whose products are hard to rebuild. The ones that sold a thin wrapper around a workflow are discovering that the wrapper is exactly what a capable model can now generate on demand.
Vishal Sikka, the former Infosys chief now raising $32 million for an enterprise AI startup, frames the shift as redefinition rather than destruction. AI is not replacing software-as-a-service, he argues, but changing what it is: instead of static products, software becomes continuously personalised and adaptive, with companies increasingly knowing what they want built while the AI supplies the how. The line between a product and a service, on that reading, starts to blur. That is good news for flexible platforms and bad news for anyone whose moat was simply that building the thing used to be too much trouble.
Underneath both narratives sits a stubborn fact that neither the bulls nor the build-it-yourself crowd have solved. Four in five workers in the McKinsey data say AI has improved their productivity, yet the share reporting actual cost savings has not budged from a year ago. Morgan Stanley pegs average net productivity gains at 8 to 12 percent, while cautioning that governance is lagging badly. And Domino Data Lab's Thomas Robinson notes that 57 percent of organisations still cannot generate returns that outrun their AI spending, with a similar share scaling agents faster than they can govern them. His warning is that value is moving away from the models themselves and toward the unglamorous work of integration, controls and workflow, the last mile where money is actually made or lost.
Put together, the week's signals describe an industry sorting itself into two piles. Software that embeds AI into a mission-critical process, with the data plumbing and guardrails to make it trustworthy, is commanding real premiums. Software that merely stood between a user and a task the user can now automate is quietly being canceled and rebuilt in-house. The rally is real. So is the disruption. They are just landing on different companies.