SaaS Apocalypse?

Until two years ago, I had access to Tableau Professional through my work. It was my favorite data visualization software and one that I loudly championed from 2016 to about 2023— even after they were purchased by Salesforce. Then I switched jobs and there was a long delay in getting a license— longer than my deadline. I knew that matplotlib could do it with work, but out of curiosity I asked Claude, "Can you make this data viz in HTML and JavaScript and load in this CSV?" Sure enough, it worked. The cost was about a penny and about 45 seconds of time.

I sat there, staring at the data viz for a while, thinking about what it meant— for myself and for Tableau as a company. After about 30 minutes of prodding Claude on my second monitor, I had a small suite of data visualization functions that could handle almost all of our future requests and did so in a way that was higher performance and hewed closer to our style guide than Tableau ever had. We didn't have 100 middling functions. We had about 5 excellent ones— and it cost us practically nothing.

That experience crystallized something I've been watching for almost a decade: people want software that is focused on their personal preferences and requirements, not software that tries to be everything to everyone. With every new Claude specialization, a wave of software company stocks are sent spiraling downward. The Citrini blog post may feel like the beginning of this trend to most of Wall Street, but from my perspective it is simply the acceleration of something that's been building for years. Large language models paired with an agent harness like Opencode, Codex, or Claude Code have reduced the barrier to creating small software projects that solve your specific problems. This may become a threat, even an existential one, to many software-as-a-service companies.

The Great Feature Bloat, 2010–Present

It's important to look back at why this appetite for small software exists. Following the Great Recession, the tech industry benefited greatly from zero percent interest rates. Hiring swelled, teams grew, and with all that talent gathering in the Bay Area, Seattle, Austin, and New York, the aspirations of tech companies heightened. Software that did one thing very well seemed obsolete compared to software platforms that could do many different things pretty well. Silicon Valley had an insatiable appetite for features.

This makes sense. If you want to maximize your market share, you need to satisfy as many feature requests as possible. With tech companies operating at a net loss on the assumption that they would one day dominate market share, the focus on features was of utmost importance. You couldn't let your competition say, "We have X more features than your product," because customers would think, "What if I someday need one of those features?"

Feature-rich software is more complex to maintain, more likely to break, more expensive to run, and more difficult to use. The gap between what most users actually need and what the platform offers grows wider every year. That gap is exactly where AI-generated tooling now lives.

The Switching Cost Trap

If SaaS products offer middling, cumbersome solutions, why do they keep getting picked by IT departments and CTOs? Consider a CTO who heads a 200-person organization with at least six departments that all perform different functions— engineering, marketing, sales, finance, HR, customer support. Each department could probably find a best-in-class tool for their specific needs, but that CTO would then be negotiating half a dozen licenses, providing training on half a dozen systems, and maintaining enough internal knowledge across all of them for the help desk to quickly resolve issues. Feature-rich SaaS products allow you to negotiate one license, maintain knowledge of one system, and offer training on that one system.

It's not that these leaders aren't aware of superior tools— it's that the total cost of ownership for supporting multiple specialized tools exceeds the benefits those tools offer. But that cost-benefit analysis is being shifted by tools like Claude Code, starting with low-risk replacements. This recalculation is happening everywhere, of course, but it is particularly threatening to the big SaaS business model— because their entire value proposition rests on being "good enough" across a wide surface area.

Claude Will Eat Elephants

The large, feature-bloated SaaS platforms are the elephants here— and they won't be taken down in a single bite. I don't expect the transition to company-specific tools to happen all at once, or even for companies to realize how much of their SaaS spend is replaceable. A very small percentage of software engineers are currently familiar with tools like Claude Code, Codex, and Opencode. An even smaller percentage of companies have adopted them.

What I suspect is that it will begin with individual engineers trying to optimize their own work and sharing the tools that come of it with their team members, percolating upward through existing companies. This is the "slow" route— and it mirrors my own Tableau experience. No one issued a mandate to stop using Tableau. I just stopped needing it.

Meanwhile, new companies will likely start from the perspective of, "Do we really need this whole service?" New companies are often cashflow negative, or even revenue-less, and the possibility that they could get the minimal functionality they need for a few dollars will be so tempting that they will try it— and be the first to discover what's possible. This is the "fast" route.

Both routes lead to the same outcome: churn for the SaaS providers and increased revenue for the AI companies and inference providers. Unlike the Citrini blog, I don't think this plays out in the next year or two. Existing companies, particularly larger ones, will not change their path unless they themselves are disrupted.

But the dynamic that killed my Tableau license is the same one that will chip away at any SaaS product that it can chip away at. The features they built to win enterprise deals are the same features most users never touch— and now, for the first time, the cost of building just the features you actually need is potentially near zero. The SaaS companies that survive will be the ones that do something AI genuinely can't replicate. For the rest, the question isn't whether this disruption arrives. It's when.