The line that matters on every enterprise-AI call now is some version of "agents." The pitch is that AI moves from answering questions to doing work, and that the doing is what customers pay for. The useful question is not whether the pitch is exciting but what, concretely, is being sold — and the patents answer that more honestly than the keynote.
Salesforce's granted patents US12632442B2, "Artificial intelligence agent creation in a database system" (issued 2026-05-19), and US12645674B2, "Artificial intelligence agent runtime in a database system" (issued 2026-06-02), describe the unglamorous core: how an agent is defined, created, and executed against enterprise data. Note what they are about — creation and runtime inside a database system — not a frontier model. The defensible IP is the orchestration layer that sits on top of someone's model and connects it to the customer's data.
Management said, quote, that agents are the future of enterprise software — but read the guidance again, slowly. The monetizable thing is the platform that turns a model into a governed, data-connected agent that an enterprise will trust. That is what these grants protect, and it is a more durable competitive position than the model itself, which a vendor may not even own.
Tie it to the revenue line, with appropriate caution. There is no public filing that breaks out "agent revenue" as its own audited segment yet; the pitch is forward-looking and the disclosure lags the marketing. So treat the agent narrative as a strategy disclosed in patents and earnings commentary, not as a booked, segmented number you can model precisely.
The forensic point is about where value accrues. If the IP is the agent creation and runtime layer — not the underlying model — then the competitive moat is integration with proprietary enterprise data, governance, and execution, not raw model capability. That is a meaningfully different investment thesis than "owns the best model," and the patents are what let you tell the two apart.
The disciplined takeaway: when a software vendor monetizes "AI agents," read the patent estate to see which layer it actually owns. Salesforce's grants point at the orchestration-and-runtime layer over enterprise data — a real, defensible position — while leaving the model question open. The receipts are in the claims; the segmented revenue is not yet in any filing.