Most AI agent demos look good for ten minutes. Then someone asks what happens with a user who has less access, or a question the documents do not answer. A useful Oracle AI Agent Studio demo answers that on screen.
This guide covers how to plan, build and run a demo or proof of concept (PoC) with AI Agent Studio for Fusion Applications. If you are still choosing who should help you, read our guide on choosing an Oracle AI Agent Studio partner first. Feature names below come from Oracle's current AI Agent Studio guide (release 26D). Screens change each quarter, so confirm them in your own instance.
1. Decide what the demo must prove
Write down the questions the audience needs answered:
- Can the agent do a real task in our Fusion process, with our data?
- Does it respect each user's Fusion roles?
- Does it cite its sources, and say "I don't know" when it should?
- Where does a person approve before anything changes?
A demo is for showing. A PoC is for proving, with agreed pass marks. Decide which one you are running. Our PoC plan template helps you set scope, success criteria and dates.
2. Pick one task and the right starting point
Choose one high-volume, low-risk task, such as an invoice hold explanation or an HR policy question with sources.
Then pick how to build it:
| Starting point | When it fits | Watch out for |
|---|---|---|
| Oracle preconfigured workflow or agent | Oracle already covers the task | Check what it reads and does before the demo |
| Duplicate of a seeded workflow | You need your documents, topics or tools | You now own the copy and its retesting |
| Agent template | A template matches a common scenario | Review its tools and approvals before use |
| Custom workflow | No Oracle agent or template fits | Most build, test and support effort |
Oracle marks preconfigured workflows and agentic applications as Seeded. You duplicate one to edit it. Oracle's feature lists by product show which agents exist in your release. Partner-built templates are listed in the Fusion AI Agent Marketplace, and Oracle states it reviews them before listing.
3. Prepare the test environment and access
- Use a Fusion test environment. Never run a PoC in production.
- Grant configuration access. Oracle documents separate access set-up for each product family, such as Financials and HCM.
- Create demo users in different roles. For example, an AP clerk, an AP manager and a user with no Payables access.
- Set workflow security. Workflow settings include a security tab where you choose which roles can use the agent.
- Run the scheduled processes. Monitoring metrics depend on the Aggregate AI Agent Usage and Metrics process. Documents added to a document tool are processed by the Process Agent Documents process before they are published.
4. Prepare data and documents
- Fusion data. Make sure the test records exist: invoices on hold, expense reports, absence balances. Include awkward cases, such as missing fields or conflicting values.
- Documents. Upload current, approved policy documents to a document tool. Remove drafts and superseded versions.
- Mask personal data where test data has come from production.
Our data readiness checklist covers this step.
5. Configure topics, tools and the model
AI Agent Studio builds agents from a few parts:
- Topics set the boundaries and instructions for an area of expertise. Oracle recommends editing topics rather than system prompts, because topics can be reused across agents.
- Tools do the work. Oracle lists business object, document, connector, email, deep link, external REST and MCP tools, among others. Business object tools follow Fusion role-based access, and you choose which objects and fields the agent may use.
- Instructions tell the model what to do and not do. Include "do not make up an answer" style instructions, as Oracle's own examples do.
- Model. Oracle groups models as Basic, Balanced and Frontier, and supports adding your own LLM. You can set the model for the whole workflow or for individual nodes. Available models vary by release and realm, so check the list in your instance.
For a demo, show one read tool, one document tool and at most one action.
6. Add human approval
Any step that sends an email or creates or updates a record needs a person in the loop. Oracle supports this in two places:
- A human approval node in the workflow pauses the run until a reviewer acts. It can use chat, email or a published approval process with levels, approvers and rules.
- Tools such as the document tool have a Require Human Approval option, so a person approves before the tool runs.
In the demo, show both outcomes: approved, and rejected or sent back for changes.
7. Test with the debugger before the audience sees it
The workflow debugger shows each node's inputs, outputs, variables and token metrics. You can set breakpoints, step through nodes and rerun from a node. Save run profiles for each demo scenario so you can replay them without retyping. Before publishing, follow Oracle's checklist: remove breakpoints, unpin outputs, clear overrides and test empty and error results.
8. Evaluate, then monitor
An evaluation set holds test questions, expected answers and pass or fail thresholds. You can load questions from a CSV file. Metrics include correctness, latency, tokens and AI units. When document tool evaluation is on, they also include groundedness, answer relevance and context relevance. You can compare two runs side by side to spot regressions.
Build the evaluation set from real user questions, including out-of-scope and unanswerable ones. Oracle's guidance is to rerun evaluations after any change to the agent or model. During the PoC, the Monitoring subtab shows sessions, errors, latency and tokens, with a trace view for each session. Oracle also offers AI unit budgets for agents, so agree a budget for the PoC. Our AI test and evaluation plan sets out thresholds and sign-off.
Sample demo script
| Step | What you show | What the audience should see |
|---|---|---|
| 1 | The business problem and today's manual steps | A clear baseline |
| 2 | The workflow in AI Agent Studio | Agent, topics, tools and approval node |
| 3 | User A asks a routine question | Correct answer from Fusion data |
| 4 | Policy question from the document tool | Answer that cites the policy |
| 5 | Out-of-scope or unanswerable question | Agent declines or redirects |
| 6 | User B with less access asks the same question | Agent shows only what User B may see |
| 7 | Agent proposes an update or email | Approval request; nothing changes until approved |
| 8 | Reviewer rejects or asks for a change | Workflow follows the rejection path |
| 9 | Evaluation run results | Pass marks met, with failures explained |
| 10 | Monitoring trace for one session | Tools called, time taken, tokens used |
Our AI demo script template gives you a blank version of this table.
Demo and PoC checklist
- Questions the demo must answer written down
- One task chosen; seeded workflows and templates checked first
- Test environment, access and demo users in different roles ready
- Workflow security set to the right roles
- Test records and approved documents loaded; personal data masked
- Scheduled processes run for documents and metrics
- Human approval on every action that changes data or sends email
- Run profiles saved for each scenario
- Evaluation set built with a business reviewer; thresholds agreed
- AI unit budget agreed for the PoC
- Rehearsal run completed the day before
Common mistakes
Only easy questions. If the demo never shows a refusal or a restricted user, those questions come later, in a worse meeting.
Building custom first. Check Oracle's preconfigured agents and templates before you build.
No approval step. An agent that posts or emails without review will not pass audit.
Demo-day changes. A late prompt tweak can break an answer that worked yesterday. Rerun the evaluation set after every change.
Forgetting migration. If the PoC succeeds, document, email and external REST tools need to be set up again after import into the next environment.
How Egisedge helps
Egisedge Technologies is an Oracle Partner, founded in 2019. We implement Oracle Fusion Cloud ERP, HCM and SCM, and work on AI agents, OCI and Oracle APEX + AI. We have delivered multiple Oracle projects, and our directors bring 20+ and 16+ years of Oracle finance and supply chain experience.
AI Agents for Finance and AI Agents for HCM are products we have in development. Our AI agent demo playbook sets out how we plan Finance and HCM demos. Our teams work from offices in India and cover GCC, UK/Europe, US and Australia/NZ working hours.
Want to plan an Oracle AI Agent Studio demo or PoC? Email info@egisedge.com to talk to our team.
Frequently asked questions
Can we demo AI Agent Studio without building a new agent?
Yes. Oracle ships preconfigured workflows and agents, marked Seeded in AI Agent Studio. You can run one as delivered, or duplicate it and change the copy with your own documents, topics and tools. For a first demo this is usually the quickest and lowest-risk route.
Where should an AI Agent Studio demo run?
In a Fusion test environment, never production. Oracle documents export and import of workflows between instances, so you can build and test in one environment and move a proven workflow later. Document, email and external REST tools need extra set-up after import.
How do we prove the agent's answers are correct?
Use an evaluation set in the Monitoring and Evaluation tab. It holds test questions, expected answers and pass or fail thresholds for metrics such as correctness, latency and tokens, with groundedness and relevance metrics for document tool answers. Run it before the demo and again after every change.
Does the agent need human approval in a demo?
Any step that sends an email, or creates or updates a Fusion record, should have one. AI Agent Studio lets you add a human approval step to a workflow, and tools can require approval before they run. Showing the approval step is often the part of the demo that reassures finance, HR and audit.
Talk to Egisedge about AI
Tell us where you are and we'll reply within one working day with next steps.