AI agents on Oracle Fusion and Power BI: a reference architecture
The five layers I use to design AI agents on ERP data: business data, integration, agent core, channels and governance.
Most AI pilots on ERP data stop at the demo. The model answers well in a meeting, but nobody trusts it with real data, real users and real decisions. The difference between a demo and a working agent is not the model: it is the architecture around it.
This is the reference architecture I use to design agents on Oracle Fusion, EBS and Power BI. It has five layers.
1. Business data
The agent works on the company's own data: ERP modules such as finance, procurement, supply chain or HR, BI datasets, and documents such as contracts, procedures and support tickets. If the data is wrong, the agent will be wrong with confidence, so data quality comes first.
2. Integration
Data reaches the agent through secure, read-scoped access: REST APIs, connectors and, above all, a semantic layer. A Power BI semantic model is often the best starting point, because it already turns tables into business concepts such as margin, open invoices or stock coverage.
3. Agent core
The language model plans the task, calls the right tools (query the data, build a table, draft a summary) and checks its answer before returning it. Prompts and context are designed per process, not once for the whole company.
4. Channels
Users meet the agent where they already work: a chat to ask questions in natural language, automatic reports and summaries, or a step inside an existing finance or support workflow. Adoption depends far more on this layer than on the model.
5. Governance
Every agent needs guardrails: role-based security so each user sees only their data, a human validating critical actions, and a log of every answer for audit and continuous improvement. In ERP environments, this layer is what makes the agent acceptable to finance and IT.
Where to start
Start with one business question that people ask every week and that today requires a technical report. Connect the agent to the existing semantic model, keep it read-only, and measure how often the answer is right. Then add the next question.