Contents
Key takeawaysWhat the studio costsWhat the templates coverWhen costs startWorked exampleWhat we have seenAccount team linesTerms to get in writingWhat to do nextFAQThe studio and its templates are free with Fusion Cloud. The work the agents do is metered, and three of the four cost drivers are set at design review by people who never see an invoice.
- $0 license. AI Agent Studio, its template library and its orchestration tooling come with your Fusion Cloud subscription at no separate charge.
- Agent work is metered. Every agent action draws on a unit meter once the monthly grant in your subscription is spent.
- Production custom agents are extra. Running custom agents at production scale needs a separate subscription, so promotion to a live pod is a commercial event.
- Design sets the bill. Model tier, grounding breadth and agent chattiness are design choices, and together with promotion they make up the four cost drivers.
- Templates cover more than expected. Across the 35 to 45 Fusion environments we worked in during 2024 and 2025, custom builds were rarer than business cases assumed.
- Model policy first. Recording which model tier each agent runs on, from week one, is the cheapest control you have.
Oracle AI Agent Studio carries a $0 license charge, and that fact causes most of the budget surprises we see with it. Teams read a free tool as free AI. Then agent activity starts drawing on the unit meter, at a rate set mostly by choices made during design and testing.
This guide separates what you get with the studio from what you pay for, shows where the cost comes from, and sets out the controls and contract terms to put in place before the first agent reaches production.
What is Oracle AI Agent Studio and what does it cost?
AI Agent Studio is the environment for configuring, extending and orchestrating AI agents inside Oracle Fusion Applications. It comes with the Fusion Cloud subscription you already hold, at no separate license charge.
Oracle set out the launch detail on the Fusion Insider blog, and the capability material sits on the AI Agents page. To answer the cost question properly, you have to keep three ideas apart:
- The tooling and templates are free. The tools for building, configuring and testing agents carry no extra charge.
- The work the agents do is metered. Agent actions draw on a unit meter once the monthly grant in your subscription is used up.
- Custom agents in production are separately entitled. Running your own agents at production scale needs a separate subscription.
What comes included at no cost?
- The configuration environment. Everything needed to define an agent, wire its tools and chain agents into a sequence.
- The shipped template library. Prebuilt patterns mapped onto processes that already exist in HCM, ERP, SCM and CX.
- Model selection. Choosing which model each agent uses, which is where running cost is decided. The studio also accepts your own credentials for outside providers such as OpenAI through Azure or Gemini through Vertex AI. You pay that provider directly, and Oracle's guide makes you responsible for moving to a supported model version when the provider retires one.
- Testing outside production. The whole first phase of any sensible rollout belongs here.
What is not free?
Consumption and the production entitlement. Every agent action draws against the unit meter once the monthly grant is spent. Custom agents at production scale sit behind a separate subscription, as set out in Oracle's custom agent documentation.
Neither charge is hidden in the contract. Both are invisible from inside a tool that opens without a purchase order, which is why they catch people out. The full unit rates, the monthly grant and pack pricing are covered in our Fusion AI agents licensing reference.
How much will the shipped templates already cover?
More than most business cases assume, so measure the overlap before anyone writes code. In the rollouts we reviewed, the first workshop produces a long list of agent ideas and a strong instinct that most of them need custom work. They rarely do.
| Pillar | Typical shipped pattern | What it displaces | Where money appears |
|---|---|---|---|
| HCM | Employee question handling and onboarding | Tier one service desk tickets | High volume, so the model tier dominates |
| ERP | Invoice narrative and approval summaries | Manual review commentary | Period end spikes drive the peak |
| SCM | Supplier and order status summaries | Status chasing by email | Scheduled runs create a floor |
| CX | Service case drafting and triage | Writing the first draft response | Grounding breadth drives the draw |
Three questions that shrink the custom list
- Is the underlying process standard Fusion? If the steps come out of the box, a shipped pattern probably covers them.
- Is the difference in the data or in the logic? Different data is configuration. Different logic is construction, and only construction is expensive.
- Would a business user notice? If explaining the difference takes more than a sentence, it is not worth maintaining.
Run this scoring in one afternoon with the process owners in the room. It is the cheapest hour in the program, and it usually removes most of the custom pipeline.
Where does configuration end and a custom agent begin?
Oracle's guide draws the line at new tools and topics. Creating a new agent, or changing a delivered one by adding tools or topics, needs the custom agent subscription. Starting from a preconfigured agent team template, uploading your own documents and editing the agent without adding tools does not.
Most planned agents sit close to that line, so get Oracle's classification of each one in writing before build starts. An outside REST call or your own model credentials are the changes we see questioned most.
Fusion AI Agents guide
Unit costs, the free allowance, model cost triggers and the Agentic Applications platform fee in one download.
Get the white paper →When does AI Agent Studio start to cost money?
At two moments, and opening the tool is neither of them. The first is when agent activity passes the monthly unit grant your subscription carries. The second is when a custom agent is promoted into a live environment and crosses the separate entitlement line.
The four places cost appears anyway
| Driver | What it is | How to see it early | The control |
|---|---|---|---|
| Model tier | Which model the agent reasons on | Run the same task on two tiers and compare the draw | A written model policy, enforced at design review |
| Grounding breadth | How much data the agent reads before answering | Count retrievals per run in the test pod | Limit grounding to the narrowest useful set |
| Agent chattiness | Retries, replanning and multi turn loops | Look at the spread across runs, and ignore the average | Cap steps per run and log the outliers |
| Promotion to production | Moving a custom agent into a live pod | It is a permission, so no purchase order ever flags it | An approval gate on promotion and a named owner per agent |
Three of the four drivers are design decisions, taken by designers and testers whose job is quality. A purchasing control sees the invoice after the fact. The control has to sit at design review, where the model, the data scope and the trigger are chosen.
The test for whether an agent stays cheap
- Does it run on the lowest capable tier? If someone moved it up during testing, that choice stays in place until someone reverses it.
- Does it read the narrowest useful set of data? Every extra source is paid for again on every run.
- Is it started by a person or by a clock? Agents a person triggers scale with value. Scheduled agents scale with the calendar.
How much can design choices change the unit bill?
By a multiple, even when the business outcome is identical. The example below uses illustrative unit rates to show the arithmetic, so replace them with the rates on your own order before you budget.
Say an HR question agent handles 3,000 employee questions a month, averaging 3 steps per run. Assume the lower model tier draws 1 unit per step and the higher tier draws 4.
| Design | Runs per month | Average steps | Units per step | Units per month |
|---|---|---|---|---|
| HR agent, lower tier, person triggered | 3,000 | 3 | 1 | 9,000 |
| HR agent moved to higher tier in testing | 3,000 | 3 | 4 | 36,000 |
| HR agent, higher tier, 10 percent of runs loop to 12 steps | 3,000 | 3.9 | 4 | 46,800 |
| SCM status agent run nightly over 1,500 open orders | 45,000 | 3 | 1 | 135,000 |
The first three rows are the same agent answering the same questions. The jump from 9,000 to 46,800 units, 5.2 times the draw, came from a tier choice in testing and a loop no one capped.
The last row is a different agent and shows the scheduled floor. Checking 1,500 orders every night for 30 nights is 45,000 runs before anyone asks a question.
What have we seen in recent Fusion AI rollouts?
One confusion turns up every time: a tool that costs nothing gets read as AI that costs nothing. That view comes from the 35 to 45 Oracle Fusion environments we worked in during 2024 and 2025, plus the studio rollouts we have reviewed since. Four patterns recurred:
- Teams treated zero license cost as zero running cost, then met consumption in the first week of real use.
- Prototypes that behaved perfectly in a sandbox crossed an entitlement line the moment someone moved them into a live pod.
- The shipped templates covered more of the early use case list than anyone expected, so custom builds were rarer than the business case assumed.
- No one had written down which model tier each agent runs on, which turned every later cost conversation into guesswork.
The last pattern is the cheapest to fix and has the largest effect. A model policy set in week one is worth more than any negotiation you will run in month twelve.
Why we disagree with the advice to start building custom agents
The advice you will hear from partners and from Oracle is that the studio is free, so you should build custom agents to differentiate your processes. We think that gets the economics backwards, because a template you configure and then leave alone is the cheapest asset in the stack.
That template survives quarterly updates with the least attention and stays on the free side of the entitlement line. Custom construction is where consumption climbs, where regression testing lands on your team every quarter, and where the separately subscribed boundary sits.
Build custom only where the process is your advantage. Everywhere else, configure the shipped template and leave it alone.
What will the Oracle account team say, and how should you answer?
Expect the conversation to push toward earlier and larger commitments. These are the lines we hear most, with replies that keep the decision on your timetable.
| What you will hear | What to say back |
|---|---|
| "The studio is included, so let your teams experiment freely." | "The tooling is included. Show us how test runs and production runs draw on our monthly grant, by model, before we open it to project teams." |
| "Buy a unit pack now so you are covered." | "We will size a pack after we have baselined a full quarter of real use against the grant. Tell us now whether unused units carry over or expire." |
| "Custom agents are where you differentiate." | "We have scored our list against the shipped templates. For the few that remain custom, confirm in writing which ones need the custom agent subscription in production." |
| "The agentic applications subscription covers all of this." | "Show us exactly which agents and unit volumes it covers, and what we pay if we never publish an agentic application." |
The last line matters because the agentic applications subscription is a separate commercial layer with its own trigger. Buying it early to settle an agent question usually means paying for capacity you have not yet designed.
What should you get in writing before agents reach production?
Ask for these terms in the order form or a side letter, before the first custom agent is promoted:
- Agent classification. Oracle's written view of each planned agent as delivered, configured or custom, so the subscription trigger is settled before build starts.
- Consumption reporting. Access to unit draw by agent and by model, so you can match any overage invoice to the agents that caused it.
- Overage notice. An alert when consumption reaches a threshold you set, before any charge above the grant is billed.
- Pack terms. Carryover of unused units across the service period and a price hold on further packs for the full term.
- Model change notice. Advance notice if Oracle reprices a model tier or retires a model your agents run on.
How do you track agent consumption before the invoice arrives?
Keep an agent register from the first workshop. For each agent, record the owner, whether it is a shipped template or custom, the model tier, the trigger (person or schedule), the data sources it reads, and the environment it runs in.
Oracle provides monitoring, testing and debugging views for agent behavior inside the studio, and its March 2026 update added an Agent ROI dashboard for measuring outcomes. Compare what those views show with your register every month.
Also review who holds studio roles, because Oracle's access guide warns that using predefined roles may count toward subscription consumption whether or not you bought the related service.
What to do next
- Week one. Write a model policy that names the default tier for each agent type and who can approve an upgrade.
- Before any build. Score the agent idea list against the shipped templates in one afternoon with the process owners present.
- During design. Default down and promote by exception, so routine template actions run on the lowest capable tier with a named approver for any upgrade.
- Before buying capacity. Baseline real consumption against your monthly grant before agreeing to any pack or bundled block.
- Before go live. Put an approval gate on promotion to production and a named owner on every agent, because that boundary is a permission rather than a purchase.
- Every month. Reconcile unit draw per agent against the register and reverse any tier change that no one approved.
Frequently asked questions
Does Oracle AI Agent Studio carry a license charge?
No. Oracle describes AI Agent Studio as available at no additional cost to Fusion Cloud customers, with the template library and orchestration tooling included. There is no order to sign before you open it, which is why its running costs go unnoticed until the first consumption report.
So what actually costs money in AI Agent Studio?
Two things: unit consumption above the monthly grant, and the custom agent subscription for agents you create or extend with new tools. A model connected with your own provider credentials is billed by that provider directly.
What is the confusion that keeps recurring?
A tool that costs nothing gets read as artificial intelligence that costs nothing. Budget owners assume the free license covers running cost too, so no one watches consumption until the first report lands.
Where do the costs actually appear?
In four places, and none of them is the tool: model tier, grounding breadth, agent chattiness, and promotion of a custom agent to production. The first three show up on the unit meter, the fourth as a subscription line.
Why does a commercial control fail for agent costs?
Because most of the drivers are fixed during design, before procurement is involved. By the time procurement sees the draw, the model and data scope are built into the agent and need a design change to reverse.
Are custom agents usually necessary?
Less often than the business case assumes. Keep custom work for processes that set you apart, and budget for retesting those agents after each quarterly Fusion update.
How do you tell configuration from construction?
Ask whether the difference is in the data or the logic. Uploading your own policy documents to a template agent is configuration. Adding a tool or a topic is construction, and under Oracle's guide it is what brings the custom agent subscription into play.
What keeps an agent cheap?
The lowest capable tier, the narrowest useful grounding, and a person rather than a clock as the trigger. An agent that passes all three stays unremarkable on the meter, even at high volume.
What is the hidden boundary in AI Agent Studio?
Promotion to production. It happens through a permission change with no purchase order attached, so a custom agent that ran for weeks in a test pod can create a subscription requirement the day it goes live.
What is the cheapest fix available?
Writing down which model tier each agent runs on. Without it, every later cost conversation is guesswork. It takes a spreadsheet and an hour, and gives procurement something to check each consumption report against.