Salesforce’s AI problem is not code. It is the cost of walking away

The “SaaS-pocalypse” has a clean narrative: AI agents write code faster, custom software gets cheaper, and customers finally cut expensive subscriptions. But The Register’s analysis around Salesforce points to the less elegant reality of enterprise software. The hard part was never just writing an application. The hard part is escaping a system that already holds data, processes, sales habits, integrations, and operational accountability.

Salesforce is a useful test case because it is not merely a CRM form with a few fields. In many organizations it acts as the operating memory for sales, marketing, support, and reporting. The official Salesforce platform is therefore not something that can be replaced by one internal coding sprint. It combines data models, access rules, automations, dashboards, integrations, and a partner ecosystem. Even if AI agents speed up the replacement interface, someone still has to prove that migrated data, workflows, and business reporting will work as well as before.

That is where the distinction between the cost of programming and the cost of change matters. AI can reduce part of the bill tied to code generation, prototypes, and internal tools. But migration away from SaaS means auditing existing processes, exporting and cleaning data, rebuilding integrations, training users, checking security, and accepting the risk that something breaks while a sales team is trying to work. This is not a romantic software revolution. It is a long dependency spreadsheet.

The more likely near-term effect is pressure on negotiation, not the immediate collapse of SaaS giants. Customers may gain a stronger argument against price increases: if some functions can be built internally for less, the vendor has to justify the platform’s value more clearly. At the same time, Salesforce can use the same AI wave to defend its position by embedding automation inside the existing system, including its Agentforce approach for business-process agents.

That does not make large SaaS models untouchable. If AI substantially lowers development costs, some edge functions, internal portals, and specific workflows may move back in-house. Companies will look harder at which licenses they actually need, which modules survive only through habit, and where a standardized tool forces an expensive compromise. But central systems with deep data and operational or compliance consequences will not be switched off just because a prototype is now cheaper.

The real lesson is that SaaS is protected not only by software quality, but by exit friction. Technically, that layer includes APIs, security, identities, and integrations; Salesforce’s developer documentation shows how much can attach to the platform once it is deeply adopted. Commercially, the defense is even tougher: contracts, accountability, user habits, and fear of disruption.

So the SaaS-pocalypse can wait. AI agents are changing the economics of building software, but they have not yet erased the economics of replacing systems. For Salesforce and platforms like it, that difference remains the most important defensive line.