Saaspocalypse

When the SaaSpocalypse hit in early 2026, markets sold off software companies as if AI made them all replaceable. That instinct is wrong, but it’s also not entirely wrong. Some SaaS really is becoming easy to rebuild internally. The trick is knowing which.

A useful frame comes from Christopher Stanton at Harvard Business School, writing in HBR. Place your software on two axes. The first is the nature of the task. Deterministic tools answer questions where the answer already exists somewhere (which part fits this unit, what did we invoice last month). Predictive tools have to infer something that isn’t in any single record (what’s likely wrong with this machine, which lead is worth chasing).

The second axis is the source of the context. Internal context means the tool draws on your own data. Pooled context means it learns from patterns across thousands of customers, jobs and edge cases.

That gives four quadrants, and two of them are genuinely exposed. Deterministic tools running on your own data (CRM search, dashboards, expense approvals) are the most vulnerable to internal builds. AI coding tools have dropped the cost of replicating them dramatically. Predictive tools built only on your own records (sales forecasts, churn alerts) are not much better off: foundation models from Anthropic, OpenAI and others can do this work without the vendor’s wrapper.

The right column is where defensibility lives. Shared-knowledge utilities like parts catalogues and compliance libraries hold up if the underlying data is genuinely proprietary. Operational intelligence (fraud detection, field-service diagnostics, threat intelligence) is the stickiest of all, because the value comes from pooled signal across many customers that no single firm can recreate.

Stanton uses Bluon, an HVAC AI assistant, as the worked example. It’s trained on around 135,000 tech support calls answered by veteran technicians. Crucially, 37% of the issues it sees are outliers that don’t cluster into recognisable patterns. That long tail is exactly where a single shop’s own records are useless, and where pooled context earns its keep.

The practical takeaway for buyers: don’t evaluate vendors on the polished demo. Demos run on the dense, common part of the data. Real value, and real moat, sits in the edge cases. Identify the queries that previously required a skilled human, and test the software specifically on those. If it handles them, the vendor is earning its contract. If it doesn’t, you’ve learned something useful either way.

AI doesn’t make all SaaS less valuable. It makes weakly differentiated SaaS easier to replace.

Where in your own stack does that line actually fall?