Fusion AI is three layers on one meter, and the shape matters more than the rate
Oracle named the meter for Fusion AI: from the 26C release, AI Units are the single currency for every agentic action at one cent each, pooled into one balance across every Fusion pillar, with every customer receiving 20,000 free each month. The commercial model is not a line item but three layers stacked on the subscription you already own, the tooling bundled, the right to run agents at production scale an add on, and the work the agents do metered, and confusing any two of those means overbuying or tripping a charge nobody budgeted.
Prepared by Redress Compliance · August 8, 2026 · Oracle advisory. Based on 35 to 45 Oracle Fusion and ULA engagements advised 2024 to 2025.
Executive summary
The three layers, and what sets each off.
AI Agent Studio and the delivered agents are bundled into the Fusion subscription at no separate fee, arriving with the release; the custom AI agent subscription is a separately priced add on triggered by running custom agents at production scale.
The Fusion Agentic Applications platform fee is an annual charge triggered only by publishing an agentic application to production, though it was framed as mandatory in the deals we reviewed; and the AI Units meter everything above the free grant at one cent per unit.
Notably absent: any per user AI seat, Oracle picking the meter over the seat that Microsoft and Salesforce first chose.
The model choice drives the bill, at 3 to 8 times for the same task.
The unit is a measure of value rather than raw tokens, rated per action in Oracle's actions table with the cost varying by the work and the model behind it.
And agentic pilots that ignored model choice ran 3 to 8 times the AI Unit cost of the same task on the Basic tier, whose general actions price at zero units: the routing decision, which model serves which action, is the largest single cost lever in the entire system, and it is an engineering setting.
Not a contract term.
The free allowance is a trial credit, and the packs carry no volume discount.
The 20,000 free monthly units are worth about $200 a month at Oracle's own price, a trial credit rather than a budget, and additional capacity sells in 100,000 unit packs that carry over during the service period rather than expiring monthly.
Priced at the same one cent as the meter: the published model carries no volume discount, so ask for one.
The billing clock starts at 26C, around July 2026, with non production testing not yet accruing charges, which makes the pre 26C window the free measurement period every sizing should use.
The forecast commitments overshot 20 to 40 percent, and the shape is the thing to write down. Forecast based unit commitments overshot proven consumption by 20 to 40 percent in the deals we reviewed, while buyers who modeled burn by pillar and model kept it predictable.
And the automation paradox needs stating plainly: agents doing the work do not shrink a Hosted Employee count, because that metric counts the workforce rather than the people who sign in.
You can benchmark a rate but not a shape, bundled things get unbundled and meters get repriced, so write down which layer each entitlement sits in today and get it into the ordering document, because that record is what makes the next renewal's reshuffling visible.
The three commercial layers, and their triggers
| Layer | Commercial shape | What triggers it |
|---|---|---|
| AI Agent Studio and delivered agents | Bundled into the Fusion subscription, no fee | Nothing; it arrives with the release |
| Custom AI agent subscription | A separately priced add on | Running custom agents at production scale |
| Fusion Agentic Applications | An annual platform fee with a bundled allowance | Publishing an agentic application to production |
| AI Units | The consumption meter, one cent each | Every metered action above the 20,000 free |
The trigger discipline is the budget defense.
The platform fee was framed as mandatory in deals where only production publishing actually triggers it, which means the pilot and test phases run on the bundled tooling and the free allowance without either add on: the entitlement map, which layer each capability sits in.
Belongs in the ordering document precisely because bundled things get unbundled and add ons get folded into the base at renewals, and the written record is the only thing that makes such a move visible when it happens.
The meter mechanics, and the sizing method
- One pool, every pillar: delivered agents, Studio agents, agentic applications, and embedded generative actions all draw from a single balance, one invoice and one concentrated dependency at once.
- Actions are rated, not equal: a lookup is cheap, a long document generation is not, and a multi step agent run stacks several actions, all defined in Oracle's actions table.
- The model sets the rate: Basic tier general actions at zero units, premium models at multiples, the 3 to 8 times spread on identical tasks.
- Size from the free window: the pre 26C period and the monthly 20,000 units are the measurement instrument, burn modeled by pillar and model before any pack commitment.
- Buy packs against proven burn: the 100,000 unit packs carry over within the service period, and the commitment sized to forecast overshot 20 to 40 percent.
The Fusion AI agents playbook
The three layer map, the burn modeling method, the pack negotiation, and the ordering document language end to end.
Get the white paper →The shape argument, and the metrics it touches
The pillar's structural warning is the one that outlasts every rate: you can benchmark a rate but not a shape, and Oracle's three layer model can move capabilities between bundled, add on.
And metered at any renewal without reopening a single price, which is why the layer map written into the ordering document is the negotiation's most durable output.
The workforce metrics interact perversely, agents doing the work of employees do not reduce a Hosted Employee count because that metric prices the tracked workforce rather than active users, the arithmetic worked in the Fusion Cloud applications guide.
The capital pressure making Oracle's AI monetization urgent runs in the AI capex and debt report; the cross vendor comparison of meters against seats in the GenAI pricing report; and the repricing wave every AI commitment eventually meets in the AI renewal cliff report.
- Percentile standing for your exact deal size and industry, from real closed transactions
- Scenario simulation before the call: test alternative terms and see the financial impact of each
- A negotiation playbook, talking points, and a two page executive brief on day one
What we saw across Fusion engagements, 2024 to 2025
Across roughly 35 to 45 Oracle Fusion and ULA engagements Fredrik Filipsson advised on in 2024 and 2025, the AI conversation moved from a flat feature toggle to a metered consumption line:
The AI Unit cost of the same task when pilots ignored routing against the Basic tier.
Unit commitments sized on forecasts against proven consumption in the reviewed deals.
The buyer behaviors that worked mirror every consumption meter in this practice: model the burn from the free measurement window before committing, route by task with the Basic tier as the default and premium models earned per action.
Hold the platform fee until production publishing genuinely triggers it, and negotiate the volume discount the published pack pricing conspicuously lacks.
Oracle's own product page and announcement carry no commercial terms, so every figure verifies against the price list in force on the order date, and the layer map in the ordering document is the artifact that survives the renewals where the shape, not the rate, is what moves.
Your first five moves
- Map the three layers and their triggers in the ordering document, the record that makes future reshuffling visible.
- Measure burn in the free window, by pillar and model, before any pack commitment.
- Route by task with Basic as the default, against the 3 to 8 times spread the model choice drives.
- Hold the platform fee until production publishing triggers it, whatever the framing says.
- Ask for the volume discount the pack pricing lacks, and verify every figure against the live price list. The Oracle practice runs the model with you.
Frequently asked questions
How does Oracle price Fusion AI agents?
Through three stacked layers: AI Agent Studio and the delivered agents bundled into the Fusion subscription at no separate fee, a custom AI agent subscription add on triggered by production scale custom agents, an annual Fusion Agentic Applications platform fee triggered by publishing to production.
And AI Units metering every action above the free grant at one cent each, pooled across every pillar.
There is no per user AI seat in the published model.
What are Oracle AI Units?
The single consumption currency for Fusion AI from the 26C release: every metered action, from question answering to document generation, draws units from one pooled balance at one cent per unit, with each action rated in Oracle's actions table by the kind of work and the model behind it.
Every Fusion Cloud customer receives 20,000 free units monthly, worth about $200, before any charge applies.
How much do Fusion AI agents actually cost to run?
The model choice decides more than the rate: agentic pilots that ignored model selection ran 3 to 8 times the AI Unit cost of the same task on the Basic tier, whose general actions price at zero units.
Additional capacity sells in 100,000 unit packs at the same one cent with no published volume discount, carrying over during the service period, and billing starts at 26C with non production testing not yet accruing.
Is the Fusion Agentic Applications fee mandatory?
No: the annual platform fee is triggered by publishing an agentic application to production, though it was framed as mandatory in the deals we reviewed.
Pilots and testing run on the bundled Agent Studio tooling and the free allowance without it, which makes the trigger discipline, paying the fee when production publishing genuinely occurs and not before, a straightforward budget defense.
Do AI agents reduce Oracle Fusion licensing costs?
Not on the workforce metrics: automation does not shrink a Hosted Employee count, because that metric prices the tracked workforce rather than the people who sign in, so the agents doing employees' work add a consumption line without subtracting a subscription one.
The AI economics stand on their own merits, measured in AI Unit burn per outcome, and never as an offset against the seat based lines.
How should buyers prepare for the 26C AI billing start?
Use the free window as the measurement period: model unit burn by pillar and by model before charges accrue, write the three layer entitlement map into the ordering document, size any pack purchase against proven consumption rather than forecasts that overshot 20 to 40 percent.
And negotiate the volume discount the published pack pricing lacks.
Oracle's product pages carry no commercial terms, so every figure verifies against the price list in force on the order date.
