Oracle wants HR teams to think about AI the way they think about hiring a contractor: pay for the job done, not the hours. Its contracts tell a different story. The company has unveiled eight HR apps running on Fusion Claw, its new engine for AI agents that carry out tasks rather than just answer questions, yet its service terms charge for Claw by the token, at 10 or 20 times the standard rate.
That gap between the pitch and the paperwork is the one HR buyers most need to understand before they sign, and it isn't the only one. Oracle's own product documentation so far names just one of the eight apps, and most of them stop short of acting on their own.
The Outcome Pitch Meets the Token Meter
Oracle has been clear about how it wants Claw to be bought. Customers "pay for trusted execution of enterprise work, described in outcome terms rather than component or token math," the company says in its Fusion Claw technical brief, and it repeats the point in its guidance on AI billing: "Oracle bills for the execution of enterprise work, not for AI components."
The contract works differently. Oracle sets out in its service terms, dated Oct 20, that AI usage is measured by adding up the input and output tokens an action consumes (tokens being the small chunks of text an AI model reads and writes), dividing by 10,000, rounding up and multiplying by a "factor." Fusion Claw, listed for autonomous execution and planning, carries a factor of 20x on a frontier model, the most capable class of AI model, or 10x if the customer brings its own. Oracle's cheapest Basic model isn't an option for Claw at all.
For most buyers, the 20x rate is the one to plan for. "Each Claw Outcome begins with a frontier model," Oracle said when it launched Claw on Sep 29. However the bill is described, the meter underneath counts tokens.
The currency for all this is the AI Unit, which Oracle says is worth one cent. Each Fusion production environment gets 30,000 AI Units a month, worth about $300, and customers can buy pooled units on top, including through the Agentic Applications subscription. That transparency is useful, because it gives buyers real numbers to model and negotiate with.
When Oracle introduced its HR agents in April, it said it was delivering them at no additional cost. The billing guidance limits free use to "general actions" such as Q&A and summarization on Oracle's basic model. Since Claw has no zero-rated option in the factor table, every Claw run draws down AI Units.
What's Actually Shipping, and When
The eight apps arrived in two waves. Workforce Staffing, which "helps HR leaders create executable staffing plans," was among 25 Claw apps Oracle described as "available today" on Sep 29. Oracle then added seven more HR apps at HR Tech on Oct 21: Workforce Orchestration, Candidate Sourcing, Workforce Compensation, Payroll Activity Workspace, and Command Centers for Positions, Recruiting and Payroll.
The technical brief says "Fusion Claw ships with Release 26D," and 26D is the fourth of Oracle's four quarterly updates to its Fusion cloud apps in 2026. Oracle rolls those updates out in three waves, or cohorts, with the first group updated in February, May, August and November and the other two following one and two months later. That points to November for the earliest customers and as late as January for the rest. Major new features also ship switched off and need an opt-in, so someone in each customer will have to decide to turn Claw on.
As of Oct 29, Oracle's list of AI features in its HCM suite tags just three HR features as Fusion Claw in 26D, and only one of them, Workforce Orchestration Command Center, matches an announced app by name. Candidate Sourcing appears as a conventional AI agent rather than Claw, and Workforce Staffing doesn't appear on the list or on Oracle's Fusion Claw capabilities page. That isn't proof anything is missing, but it means buyers should ask for the 26D feature name and readiness document for every Claw app in a proposal, and treat anything without one as roadmap.
Where the documentation does exist, it describes something more cautious than "autonomous." Workforce Orchestration Command Center, according to its 26D readiness notes, "is read-only" and "doesn't perform HCM transactions" such as transfers or terminations. Access to workforce information depends on users' roles and security profiles, so the immediate risk is who can see sensitive data rather than what the system might do with it.
