What's Free and What's Not in Oracle Fusion AI Agent Studio

You already paid for Oracle Fusion. Now everyone's talking about AI agents. And the natural question is: do I need to open my wallet again?

The short answer is: it depends on what you want to build. Oracle's pricing model for AI Agent Studio is more nuanced than most customers realize — and getting it wrong in either direction is costly. Building too conservatively because you assume everything costs extra means leaving real automation on the table. Building too freely without understanding the licensing boundary means a surprise conversation with your Oracle rep.

This post breaks it down clearly, based on the February 2026 Oracle AI Office FAQ, so you can plan your agent roadmap with confidence

What You Already Own

If you're on Oracle Fusion ERP, HCM, SCM, or Sales, AI Agent Studio is already yours. No new SKU. No separate provisioning request. You navigate to Tools > AI Agent Studio and you're in.

But more than the Studio itself, a substantial amount of functionality is included at no additional cost:

• Every embedded, out-of-the-box agent template Oracle ships — and there are over 1,000 now across the Fusion suite

• Use of Oracle-hosted LLMs: OpenAI GPT-OSS, Meta Llama, and Cohere — unlimited, with no token counting

• Editing prompt text within existing agents

• Configuring or removing tools and data objects from existing agents

• Changing the documents attached to a Document tool

• Adding Human-in-the-Loop approval checkpoints

• Accessing an embedded agent through Microsoft Teams, Slack, or any other channel

That last point surprises a lot of teams. Surfacing an embedded Oracle agent in Teams — which many organizations want to do — doesn't change its licensing status at all.

Where the Line Gets Drawn

Oracle's pricing model defines a category called Custom Agents — and crossing into that territory requires a paid subscription per the Oracle Fusion Cloud Service Price List.

The trigger isn't complexity. It's extension. The moment you reach outside the template's original design and connect it to something new, you've crossed the line.

Specifically, Oracle considers an agent custom when you:

• Create a net-new agent from scratch

• Add a new Business Object not in the original template

• Add an External REST API to any agent — including connections to other Oracle products

• Integrate via MCP (Model Context Protocol) with external systems

• Switch to a non-Oracle LLM (Bring Your Own Model)

• Change the Topic or intent of a delivered agent in a meaningful way

• Enable new output modalities like charts, images, or voice responses

The good news: as of release 25D, the Studio now displays a dialog box and visual icons that warn you when a modification is about to cross the embedded-to-custom boundary. Your team won't step over the line without being notified.

The Token Conversation Nobody Has Until It's Too Late

Tokens are the unit of content that LLMs process — both what goes in and what comes back. This is where many teams get caught off-guard.

The critical rule: If your agents use Oracle-hosted LLMs (GPT-OSS, Llama, or Cohere), tokens are not counted and not charged. You can run these models as heavily as you want.

Token limits only apply when using OpenAI's hosted models (GPT-5 mini, GPT-4.1 mini). Every Fusion subscription includes a base allocation of 200 million tokens per month, which resets monthly and is the same regardless of company size.

If you purchase Custom Agent subscriptions, additional tokens are included:

• Per Authorized User SKU: 2 million tokens/month per user

• Per Employee SKU: 100K tokens/month per employee

• Top-up bundles: Available in 1 billion token blocks, pooled over the subscription term

The practical implication: unless your team specifically wants OpenAI's latest hosted models for their performance characteristics, sticking to OCI-hosted LLMs eliminates token cost concerns entirely.

Two Ways to License Custom Agents

When you're ready to build custom agents, Oracle offers two pricing models:

Per Authorized User — Count the number of users who can access the agent after it's published. You manage this through permission groups in the Studio. If your Purchasing team of 10 people is the only group with access to a procurement agent, you subscribe to 10 Authorized Users for that agent. You can adjust this up or down any time without notifying Oracle.

Per Employee — Counted against your total hosted employee headcount. This carries a significantly lower unit price and makes sense when you're deploying agents broadly — say, an HR policy agent that every employee can use. Oracle recommends subscribing for approximately 5x your employee count when you're planning to roll out multiple agents, to give yourself flexibility.

Both models can coexist. A company might use Per Employee for a benefits Q&A agent available to all staff, and Per Authorized User for a specialized financial close agent available only to accountants.

One important note: if a custom agent team calls multiple agents in the background to complete a task, that entire multi-agent team counts as one agent for licensing purposes. This is a significant cost efficiency for organizations building complex orchestration.

Develop and Test Without a Subscription

Here's a rule worth sharing with your team immediately: you don't need a Custom Agent subscription to build and test in non-production environments.

Oracle explicitly allows development and testing of custom agents in non-production instances without triggering a subscription. The subscription requirement kicks in when you publish the agent to production. This means your team can prototype freely, iterate on agent designs, and validate business logic before committing to a licensing model.

The exception: if you're testing with OpenAI's hosted LLMs, token usage is counted even in development environments.

Start With Five

Oracle's guidance to new customers launching a custom agent program is to plan for at least five agents from the outset — not because five is a magic number, but because customers who plan for only one or two often find themselves scrambling to expand subscriptions as they quickly identify additional use cases. The 5-agent starting point builds in the flexibility to move fast as value becomes clear.

The Bottom Line

Oracle Fusion AI Agent Studio gives you more than most customers realize — for free. The embedded tier is genuinely powerful, and the licensing boundary is clear enough that with proper planning you won't be surprised.

When you're ready to go beyond that boundary, the Custom Agent model is flexible. Two pricing metrics, mix-and-match capability, no surprise bills (Oracle notifies you if you exceed your subscription and gives you the choice to expand or adjust), and a free development lane to prototype before committing.

Plan your agent roadmap against this model early, and the economics become straightforward.