AI tokens dropped 80%. Your AI bill rose 320%. Meet per-agent pricing.

In 2025 AI token prices fell 80% year over year. In the same window enterprise AI spend grew 320%. Magic? No: your SaaS stopped charging per user and started charging per "agent action" at $0.10 each. Salesforce Agentforce went from $200M to $800M ARR in a year. Let's do the math on who pays for the party.

The new trick is called "per-agent pricing"

For a decade your SaaS charged per user per month. Abusive, but at least predictable. In 2026 they switched models: now they charge per "agent action". Salesforce Agentforce bills 20 Flex Credits per standard action and 30 per voice action; at $500 USD per 100,000 credits, that's $0.10 per action and $0.15 per voice interaction. The alternative? $2 USD per conversation, or per-agent licenses from $125 USD/month. Three different models, all with the same effect: your bill no longer depends on how many people you have, but on how much the software works. And software works 24/7.

Jevons paradox, applied to your budget

Economist William Jevons observed in 1865 that when coal became more efficient, England didn't use less coal: it used vastly more. The same is happening with AI. Token prices fell 80% in 2025, but enterprise AI spend grew 320% the same year (SaaS & AI Pricing Report 2026). Every time an input gets cheaper, we find ten new uses. That's not the problem — it's healthy. The problem is WHO captures the margin between the real cost (pennies per million tokens) and what your SaaS bills you ($0.10 per action, which may consume just a few thousand tokens costing a fraction of a cent).

Who keeps the difference?

Salesforce's numbers tell the whole story: Agentforce went from roughly $200 million USD ARR in Q1 to $800 million by the end of fiscal 2026, a 169% increase (Salesforce public results). That money doesn't come from new users — it comes from charging you for every action an employee used to do for free inside a license you already paid for. The "per-agent" model turns your team's productivity into a billable metric. The more efficient you get, the more you pay. It's the perfectly inverted incentive: the vendor wins when you work more, not when you earn more.

The rule for 2026: own the agent, own the margin

If you're going to deploy AI agents across your operation — and you should, they're the biggest productivity lever of the decade — the question isn't "which SaaS rents them", but "who owns the agent's logic". With custom software, your agents run on your own API keys (OpenAI, Anthropic, Google), you pay the real token cost, and action number one million costs the same as the first. No meter running, no expiring credits, no end-of-contract true-up. You capture the margin you give away today. At TMS we build custom AI agents integrated into your processes, with fixed pricing and no per-action fee. If you're already paying for "actions" or "credits", ask us for a custom AI software quote: in most cases the savings pay back the build in under 10 months.