You’ve probably been losing sleep over AI hallucinations or rogue agents making disastrous decisions. You’re focused on the wrong threat. The real existential risk to your enterprise isn’t model accuracy—it’s the invisible toll booths being erected right in the middle of your tech stack.
Consider this: one company recently discovered it had 900 AI agents running wild across its systems for a year before anyone thought to build a committee to oversee them. Sixty percent of enterprises are building AI workflows completely outside of IT’s purview. You have a shadow workforce of autonomous agents, and no one is tracking the meter.
The legacy vendors saw this chaos and recognized a massive opportunity. SAP and Salesforce appear to be taking opposite paths, but they are driving toward the exact same destination. SAP played the heavy, locking down its APIs in April. External AI agents can no longer directly interface with SAP; they must route through SAP’s proprietary assistant, Joule. If you try to bypass it using a proxy server or custom code, they block you at the technical level. Meanwhile, their integration gateway started charging per API call.
Salesforce took the “open” route, decomposing its entire platform into 60+ tools and 30+ skills, allowing agents to execute CRM tasks entirely through Slack. But don’t confuse openness with charity.
They aren’t selling you software anymore; they’re charging you rent to cross a bridge they built around your own data.
ServiceNow is charging per “assist” as agents pass through its Control Tower. Workday published a point-based menu: checking a payroll anomaly costs 5 points, screening resumes costs 6 points, and finding a candidate costs 750 points. Google is literally charging for “semantic governance strategies” starting this August. Salesforce is billing roughly ten cents per action. Atlassian is charging per point when you feed their context to your AI.
Notice the pattern? The rules for managing agents and the price sheet for charging them are the exact same mechanism. The gatekeeper is also the tax collector.
The pricing model is the governance model. Whoever controls the meter controls the market.
Most buyers think they are paying for outcomes or usage. You are actually paying the vendor to unilaterally define what an “action” counts as, who gets authorized, and what gets recorded. And that meter is deliberately opaque.
The bills are already blindsiding executives. Zendesk customers opened their invoices to find they had been quietly charged for overages they didn’t know existed. On GitHub, the cost of a model jumped from 7.5x to 27x in two months; users found out from their bills, not from an announcement. By 2028, the cost of using AI to write code will exceed the salary of the developer it’s assisting. The old software model charged you for a seat before you used it. The new AI model charges you for every step the software takes—and the vendor decides how many steps it takes.
Buyers are starting to fight back, demanding outcome-based pricing. Intercom charges 99 cents only when a ticket is resolved. Sierra hit $200 million in revenue doing the same. But even this is a trap. When a vendor handles 76% of your support tickets automatically, the volume of “resolutions” skyrockets. You think you’re saving money on human labor, but the vendor is logging every single successful action. The more capable the AI, the more you depend on it, and the higher the bill scales.
You think you’re paying for AI outcomes. You’re actually paying a vendor to define what an ‘action’ is, who gets authorized, and what gets recorded.
The next two years will be a brutal tug-of-war. The vendors have installed the meters and started the clocks. The buyers are still trying to figure out where the meters are hidden. Your CRM, ERP, and HR systems are quietly becoming tax collectors for your digital workforce. If you don’t define what an “action” is worth before you deploy, the vendors will gladly do it for you—at their own premium rate.
FAQ
Q: Isn't usage-based pricing fairer than flat enterprise licenses?
A: It's fair only if the meter is transparent. Right now, vendors unilaterally define what an 'action' is and change the multipliers without notice. You're paying for their arbitrary definitions of work.
Q: How do we avoid surprise AI bills from vendors?
A: Demand outcome-based pricing (pay only for resolved tickets or successful transactions) and build internal observability tools. Never accept a contract where the vendor controls the definition of a billable 'action'.
Q: Is this just a passing phase before open-source AI dominates?
A: No. Open-source models don't solve the governance and identity problem. Enterprises will always need a 'bouncer' for their proprietary data, and vendors know this. The toll road is the permanent endgame.