Asking Your Profitability Model Why: The Conversational Layer for Oracle EPCM
This series started with an inventory of what AI does inside Oracle EPCM. The PCM Agent writes rules. IPM Insights tells you where a margin moved. Generative AI describes the movement in a paragraph. None of them answers the question a CFO asks next: why did it move?
As covered in Part 1, Oracle's September 2026 program for conversational interfaces covers Planning and FreeForm, not EPCM. Teams that want an executive to ask a profitability model a question in plain language have to build that layer themselves. This post covers how I would build it, and the rules that keep it honest.
What "why" means in a profitability model
In a profitability model, "why" has a precise answer. A margin moves because revenue moved, because a driver quantity moved, because a cost pool moved, or because someone changed a rule. A useful answer separates those causes and puts an amount on each one.
An answer like that is only possible if every number comes from the model and the agent's job is limited to finding, comparing and explaining them.
The architecture
The layer has four parts. Each one has a clear job.
The question. A CFO or controller types in plain language. No syntax, no member names.
The agent. A large language model reads the question, decides which points of view and periods to compare, calls the tools and writes the answer. It's the only part that uses AI.
The tools. Thin wrappers around EPCM's REST APIs. Oracle documents Export Data Slice and Execute a Job as supported for EPCM, along with EPCM-specific endpoints such as Calculate Model, Copy Data by Point of View and Generate Model Documentation Report. The first two read balances and drivers. The documentation report gives the agent the rule definitions it needs to explain a cost path.
The model. EPCM stays the system of record. It calculates, it holds the trace and it enforces security.
Five rules that keep it honest
- The model computes, the agent explains. Every number in an answer comes from an export. The language model never estimates, rounds into new totals or fills a gap with a plausible figure.
- Every answer shows its source. Point of view, rule name, driver. If the agent can't cite where a number came from, it says so.
- Read-only by default. The same API that exports data can also calculate, copy and delete points of view. Those calls stay behind a human approval, the same way the PCM Agent waits for someone to press Run Command.
- Teach it the model before you ask it anything. Feed the agent the model documentation and a driver dictionary: what each driver represents, where it comes from and who owns it. That dictionary is the first question of the design check in Part 4.
- Test it like a model. Keep a set of questions with known answers and run them after every model change. If an answer moves and the business didn't, something broke.
What it still can't do
The layer explains the model it reads. It can't tell you the model is wrong. If IT cost is allocated by headcount, the agent will explain the headcount effect clearly and confidently. Part 4 is the prerequisite for Part 5.
It also inherits the model's security through the account it connects with. Give it a service account scoped to what the people asking are allowed to see, not an administrator login.
And it should be easy to replace. Oracle may extend its own program to EPCM. If the driver dictionary, the test questions and the approval rules live outside the agent, moving to a different agent later is a short project.
Where the series leaves us
Five posts, one conclusion. AI in Oracle EPCM is useful and getting better every update. The PCM Agent saves real modeling time, Insights finds movements nobody was looking for and a conversational layer can put the model in front of the people who make decisions. All of it depends on a model designed to be read: drivers with a cause, rates on practical capacity, one stage per rule set, names a stranger can follow.
At Asher we build this layer as part of StrategicFinance.ai™, on models we design for it. If you want to talk about what it would take on yours, write to me at psanmartin@asher.company.
Pedro San Martín is Founder & Principal of Asher & Company, a strategic finance and profitability analytics firm specializing in Enterprise Performance Management, profitability architecture and cost management. He chairs the IMA Profitability & Cost Management SIG.
Sources: Oracle, REST APIs for Oracle Fusion Cloud EPM, chapter "Enterprise Profitability and Cost Management REST APIs"; Oracle, EPM Cloud REST API Compatibility tables; Oracle EPCM documentation on the PCM Agent and IPM Insights, cited in Parts 2 and 3 of this series.
