The $285B SaaS Reckoning: When AI Makes Switching Costs Meaningless

The most important thing about SaaS lock-in was never the contracts. It was the cost of replacement: the migration effort, the retraining, the integration rebuild, the six months of productivity loss. That cost is collapsing, and the software industry's pricing power is collapsing with it.

SaaS companies built their business models on one structural advantage: the pain of leaving. ServiceNow, to take the canonical example, reportedly holds renewal rates near 98%, which has nothing to do with customer delight and everything to do with the fact that ripping out a workflow platform costs more than the annual licence. That calculus held for twenty years. It's breaking now, and the catalyst is something most CFOs haven't fully priced into their renewal decisions yet: AI-assisted development has made building your own tools dramatically cheaper than it used to be.

Anthropic's Claude, GitHub Copilot, and a generation of agentic coding tools have compressed the skilled-developer hours required to ship a functional internal application. The conversation accelerated publicly when Anthropic's Claude Cowork plugin started enabling teams to build and iterate on custom workflow tooling through natural language prompts, sparking a broader argument in product and engineering circles: why pay per-seat for a tool that does 60% of what you need when you can build the 100% version in a week?

That argument used to be easily dismissed. Custom software means maintenance burden, security risk, and dependency on whoever wrote it. Those concerns are real. But they've been partially neutralised by the same AI layer that's doing the building. Copilots that generate code also explain it, document it, and help the next developer understand it. The "bus factor" problem of bespoke tooling is smaller than it was.

The scale of the repricing is striking. Roughly $285 billion in SaaS market value has evaporated since early 2026, with analysts at S&P Global noting in their credit trends assessment that AI risks vary significantly across software subsectors, with those most exposed to point-solution displacement at greatest risk. Morningstar, in a more optimistic read, argues the incumbents aren't finished because deep integrations and data gravity still hold. Both analyses can be true simultaneously. The strongest SaaS platforms will survive. The mid-tier point solutions, the ones that charge $40/seat for something a team could ship in a sprint, are in genuine trouble.

The economic mechanism here is straightforward. SaaS switching costs have three components: the direct cost of migration, the opportunity cost of disruption, and the risk premium of uncertainty. AI tooling attacks all three. Migration tooling is getting smarter. Disruption is lower when the new thing is built incrementally and tested in parallel. And the risk premium drops when your engineers can move faster and explain what they've built more clearly. The "great unbundling" thesis argues this pressure on SaaS isn't new, but the speed of the current shift is qualitatively different.

What's dying is the lock-in safety net that let average products command premium prices. The platforms with genuine network effects, with proprietary data that improves through collective use, with integrations so deep that extraction would take years rather than months, those retain their pricing power. Salesforce's CRM data layer isn't easily replicated in a sprint. A project management tool with basic Kanban, a few automations, and a reporting dashboard? That's a different conversation.

The practical implication for any team managing a software budget is that the make-vs-buy analysis needs revisiting with updated assumptions. The standard calculation assumed that "build" was expensive, risky, and slow. For complex, mission-critical infrastructure, that's still mostly true. For the layer of operational tooling that sits above your core systems, the tools your team uses daily to move work through a process, the assumptions may now favour building, especially if your workflows are non-standard enough that off-the-shelf software forces awkward workarounds.

The honest caveat: this shift benefits teams with engineering capacity. A 10-person ops team without developers still depends on SaaS, because the AI coding advantage requires someone who can direct it intelligently, review the output, and maintain what gets built. The "citizen developer" narrative oversells how much non-technical users can realistically own. The more accurate picture is that one mid-level engineer, equipped with modern AI tooling, can now do what previously required a small team. That changes the economics for companies with any engineering resource, not for everyone.

If you're on the buyer side, the immediate move is to audit your current stack for what your team actually uses versus what you're paying for. Any tool where utilisation is below 40% of licensed seats, or where your team has built shadow workarounds because the product doesn't quite fit, is a candidate for a build conversation. Run the real numbers: total seat cost over three years against a rough engineering estimate using AI-assisted development. You may be surprised how close the figures are, and at that margin, control over your own roadmap tips the decision.

If you're on the vendor side, the strategic response is to deepen the data layer, not add features. Features can be replicated. A model trained on your customers' collective behaviour, an integration that becomes load-bearing, a data asset that compounds: those are the moats that still hold.