AI may not be killing legacy software. It may keep it alive. Bad news for founders. And contrary to what the SaaSapocalypse supposedly was meant to be.
On Tuesday 4 February 2026, roughly 300 billion dollars came off software in a single day. Salesforce, ServiceNow, Adobe and Workday each fell about 7 percent. The trigger was an AI product release. The iShares Expanded Tech-Software Sector ETF (IGV) index was already about a third below its late September 2025 peak, forward multiples had been squeezed, down from around 39x to 21x. Short sellers cleared more than 20 billion dollars in 2026 betting against legacy SaaS.
We AI-founders read that as our moment. The incumbent’s product is terrible, and now the market knows it.
Fast-forward to last month: ServiceNow renewed at 98 percent, in the same quarter its agent deployments grew ninefold in nine months, its AI contract value passed one billion, and their subscription revenue rose 24.5 percent. By early June the index had gained 40 percent off its April low.
Same arena, opposite verdict on the play given, just five months apart. The swings aren’t the story. The story is what buyers did while investors panicked and then apologised. They renewed.
Three reasons that matters more to and founder than it does to the incumbent.
First, AI is WD40 and Play-Doh at the same time. WD40 lubricates what grinds, a chat layer over an interface nobody has liked in a decade, an agent that fills the form nobody wanted to fill, a model that reads data the schema never supported. Play-Doh fills the missing shape, the integration that was never built, the report nobody could get out. Legacy roadmaps are frozen; the requested feature will now sit on the roadmap forever because a language model covers the gap well enough. The approach does not fix the product underneath. But it raises the floor of what a buyer will put up with when it comes to replacement costs.
Second, the friction was the wedge. Hardly anybody in a large organization rips out a working system because a better one exists. They rip it out because the current one hurts every Monday, every Tuesday and so on. Or until someone senior gets the wrong number in a board pack. That’s when the replacement budget magically appears. Displacement runs most of the time on annoyance, not on architecture. Take the annoyance away and the replacement conversation gets harder, by a lot.
Third, AI attacks lock-in and defends it at the same time, and the defence is the half that reaches the buyer. On 23 February 2026, Anthropic published a post on using Claude Code for the analysis, dependency mapping and documentation work that made COBOL modernisation expensive. IBM fell 13.2 percent the next day, its worst day since October 2000, and closed February down 27 percent, its biggest one-month loss since the sixties. That market priced the death of switching costs. Yet five months later IBM’s CFO said there was no evidence of clients moving off mainframe, and that AI was pulling new workloads onto it…
Both things are true. Writing the replacement got much cheaper. Switching did not. Undocumented business logic from three decades ago, data structures welded to the old environment, regulatory sign off that moves at its own pace, the original authors retired years ago. Cheaper code is not a cheaper switch.
There is a second reading of the same data that I find harder to argue with. Test Double looked at what AI changes about legacy systems and separated three signals: user friction, engineer momentum, and the cost of carrying the thing. AI improves the last two. Friction stays where it was, or gets worse, because the same tools make it easy to bolt on features faster than anyone removes them. Their line is that AI is a maintenance tool, not a prevention tool. Put that next to my argument and it gets worse, not better. If AI preserves old bad software and helps produce new bad software faster, product quality decays as a differentiator from both ends at once.
So what is left as a wedge? Not the better interface, that is a weekend of work away. Not the missing feature, Play-Doh covers it. What is left is what cannot be generated. A process end to end that the incumbent does not own and cannot reach from where it sits. Data nobody else holds, which is the part everyone skips because curation is boring and that is exactly where the value sits. An audit trail designed in from the start, because governance does not retrofit onto a system built to be trusted rather than inspected.
All this puts single-purpose AI-vendors into a bad spot. Or to quote Terra Higginson, principal research director at Info-Tech Research Group: “Those guys are in trouble right now. They don’t own a lot of the workflow, they don’t own a lot of the transaction, they don’t own a lot of the data and they’re easy to replicate through vibe coding”.
That is not a reason to build a better AI first product. It is a reason to stop treating the incumbent’s badness as your go-to-market plan, because badness is now a solved problem.
