For years, enterprise software buyers had a workable threat: cut seats, run a competitive process, or replace the application with something narrower and cheaper. Generative AI is weakening that threat because the product being purchased is no longer just an application.

The new bundle contains models, governed company data, identity, permissions, workflow orchestration and the system where employees already do their work. Forrester’s read of the latest earnings cycle is blunt: AI and platform adoption are shifting negotiating power toward vendors Forrester. A review spanning more than 25 software earnings calls found executives repeatedly tying AI to monetization, packaging and expansion Q1 2026 SaaS Earnings Review, meaning buyers are hearing the same pitch across an entire procurement calendar rather than from one overeager account executive. You can see where this is going.

Cheap models; expensive context.

An enterprise agent does not become useful when it can write a decent paragraph; it becomes useful when it can read the customer record, check entitlements, retrieve the correct contract, obtain approval and modify the billing system without exposing payroll data to the intern who asked the original question. What controls that chain? Usually the incumbent vendor, because it already owns the permissions, metadata and workflow hooks.

That is the reversal. SaaS buyers spent the last cycle decomposing suites into best-of-breed tools, then using overlapping products as bargaining chips during renewals. AI rewards recombination: the more data and workflows a vendor can put behind one policy layer (and one consumption meter), the more valuable its agent appears.

Oracle’s latest fiscal-year release reported fourth-quarter GAAP earnings per share of $1.45, up 21%, while adjusted earnings per share reached $2.111, up 24% Oracle FY2026 Results, meaning shareholders received a considerably fatter slice from each share while customers were being pushed toward broader cloud commitments. Salesforce, meanwhile, said it had closed more than 4,000 paid Agentforce deals by its fiscal first-quarter report Salesforce Q1 FY2026 Results, meaning the sales motion had already moved beyond innovation-lab demos into thousands of procurement decisions.

Credit where it’s due—this part was well built.

The model itself is becoming the least defensible layer. Stanford found that inference costs for systems performing at roughly GPT-3.5 capability fell more than 280-fold over the measured period Stanford AI Index, meaning a model task once priced like a hotel room moved toward vending-machine economics. Vendors know this, which is why their earnings scripts linger on data clouds, agent platforms, governance and industry workflows rather than the underlying model weights.

Yet cheaper intelligence does not guarantee cheaper software. Microsoft lists Microsoft 365 Copilot at $30 per user per month Microsoft 365 Copilot Pricing, meaning a broad deployment can become a standalone software budget rather than a casual add-on. Salesforce has also marketed Agentforce around metered consumption, including conversation- and credit-based pricing Salesforce Agentforce Pricing, meaning the support leader no longer buys only seats; she pays as the machine performs work.

A tidy little tollbooth.

For investors, the important distinction is between AI attachment and AI dependency. An attached product can be removed at renewal; a dependency is wired into approvals, customer service and financial operations, where replacement carries migration risk and the possibility that an automated process will quietly do the wrong thing at machine speed. The latter supports higher contract values, lower churn and more credible expansion, though it also shifts cloud-compute expense onto vendors whose pricing may prove less clever than their demos.

But reported “AI ARR” deserves suspicion when the product was bundled into a suite renewal, discounted against shelfware or funded through expiring credits. Ask how much usage is production traffic, who absorbs inference costs, whether gross margin improves as models get cheaper, and whether customers can route workloads to another model without rebuilding policy and evaluation layers. Not great, if management cannot answer.

Buyers still have defenses. They can demand model portability, exportable audit logs, hard consumption caps and separation between orchestration logic and the vendor’s preferred inference service; the NIST AI Risk Management Framework offers a useful baseline for governance and measurement NIST AI RMF. This is a mistake if procurement treats those provisions as security boilerplate, because they are now economic rights disguised as technical clauses.

The next great software lock-in will not look like a database contract—it will look like an employee asking an agent to get something done.