Start with the division
Oracle put execution counts on the table for its Q1 FY2027 quarter, and that is rarer than it sounds. Per diginomica, Fusion applications consumed 900 billion tokens during the quarter, embedded AI capabilities were used more than 150 million times, up 42 percent sequentially, and AI agents executed 3.5 million times in production, nearly doubling quarter over quarter. Cloud applications revenue was $4.2 billion, up 10 percent year on year. Co-CEO Mike Sicilia said customers had deployed over 2,300 AI agents in production, up 90 percent quarter over quarter.
Two of those figures sit next to each other and ask to be divided. 3.5 million executions across 2,300 deployed agents comes out to roughly 1,520 runs per agent for the quarter. Call the quarter 90 days and that is about 17 runs per agent per day.
Seventeen a day is a real workload for a narrow task and a rounding error for a broad one. An agent clearing supplier invoice exceptions for one business unit is earning its keep at that rate. An agent scoped to order management across a global Fusion footprint has barely started. The average cannot tell you which, so sit with it before anyone in your building calls 3.5 million mass adoption.
Both numbers doubled, which is the real signal
Agents in production rose 90 percent quarter over quarter. Executions nearly doubled. Those two growth rates landing close together means average work per agent barely moved. Oracle's volume is growing on new agent count rather than on heavier use of the agents already running.
Our read is that this describes an install base still widening, where teams prove out a second and third use case instead of pushing the first into the core of a process. If you are planning your Fusion roadmap on the assumption that peers have agents carrying real transaction volume, the number does not support that yet. Same shape we flagged in Workday's own quarterly agent adoption figures, and it is why a readiness check before you scale past the first use case beats a bigger pilot count.
Most of the AI inside Fusion is still assistive
Set the 150 million embedded AI uses against the 3.5 million agent executions and the ratio comes out around 43 to 1. Agent executions are roughly two percent of the AI activity Oracle reported. Most of what Fusion users touched last quarter sat inside a screen they were already in, answering or drafting while a person stayed in the loop.
The mix is shifting. Embedded use grew 42 percent sequentially and agent executions close to 100 percent, so hold both rates steady for four more quarters and the gap narrows a lot. What your peers run today still argues for spending design effort on the handoff between human and agent work rather than on removing the human.
The 900 billion token figure does not break out by workload, so per-transaction cost math built on it is guesswork. Divide it by the 150 million embedded uses and you get about 6,000 tokens each, reasonable for a short assistive interaction and far too small for a multi-step agent run. Oracle does not say the proportion.
What the counts do not separate
Nothing here splits pilots from production scale. The phrase in production rules out a sandbox. It does not rule out an agent that one team runs twice a week with a human approving every result. The 2,300 figure also arrives without a customer count, and whether that is 2,300 agents across 200 customers or across 1,200 changes what it says about breadth.
The mean hides its own distribution. Our guess is that a small set of high-frequency agents doing document handling and routine approvals account for most of the 3.5 million, while a long tail runs a handful of times a week. If so, the median agent sits far below 1,520 runs a quarter, and the median is the number that describes a normal deployment.
There is a timing question too. If 2,300 counts agents at quarter end while 3.5 million is cumulative, the ones running all quarter did more than 1,520 executions each. Nothing in the reporting settles it.
The NetSuite claim needs a scope attached
diginomica also reports that early NetSuite customers using AI tooling cut implementation timelines from double-digit months to single-digit weeks for production go-lives. That is roughly an order of magnitude, and it is the line most likely to land in somebody's business case.
We would want the scope before this becomes a planning input. Which modules, how much data migration, how many integrations, and what got deferred to a phase two. Timelines compress fastest on projects that were mostly configuration, and the months inside a double-digit ERP schedule usually go to data and to decisions nobody has made yet. Same gap we had with the month-saved claim ServiceNow put out, where the headline arrived without a baseline. It does not change the question that actually decides NetSuite or Fusion.
The figure we would want broken out next call
Median executions per deployed agent. That one number settles whether 3.5 million reflects a broad base doing steady work or a narrow set of high-volume agents carrying an install base still mostly experimenting. Mean divided by count is the figure every vendor offers, because it flatters.
Second is a customer count behind the 2,300, and a definition of what counts as one execution. A run with twelve tool calls counted once reads very differently from the same run counted twelve times, a gap wider than most of the growth rates being celebrated.
Sicilia's line was that the introduction of AI is "an accelerator, not a replacement for packaged applications", and the ratios support him more than the headline execution count does. So carry the 17 runs a day into your next roadmap meeting. Ask what volume your own candidate agents would need to clear it, and whether any process you own has that much repeatable work sitting in it. If the answer is no, you are budgeting for an assistant.
