A librarian, not an oracle: why EPM Agent cites its sources

Ask a general-purpose AI model how to structure an entity hierarchy in Oracle Financial Consolidation and Close, and you'll get a fluent, reasonable answer. Ask it which entities your client agreed to consolidate under the new holding company, and it will still give you a fluent, reasonable answer. That second answer is the problem.

Implementations run on agreements, not general knowledge

An EPM implementation is a long chain of specific decisions: this cost centre rolls up there, this account is reconciled monthly, this form shows these members for this scenario. They are made in workshops, revised in email threads, captured (or not) in design documents, and remembered by whoever was in the room.

The hard part of delivery isn't knowing how Planning works. It's knowing what was decided for this client, by whom, and whether it has since changed. A model that fills the gaps with plausible guesses is worse than no help at all: it moves the errors from the meeting notes into the build.

Select, don't invent

So we built EPM Agent around one rule: the system is a librarian, not an oracle.

• Models select from evidence. When the agent answers, it searches the client's own knowledge base, which holds the meetings, decisions, requirements and correspondence, and builds its answer from what it fetched. The answer carries those sources with it.
• Code validates and transcribes. The agent doesn't free-type its citations. It points to records it actually retrieved, and code expands them. A source that wasn't fetched can't be cited.
• The governed record holds the truth. New facts and decisions that agents propose, from a meeting for example, are reviewed by a person before they become part of the record. A superseded decision is never presented as binding; an open question stays flagged as open.

When the evidence isn't there, the right answer is "I don't have a record of that decision", along with who might know. That answer is far more useful than a confident guess.

The same rule, all the way to Oracle

The principle doesn't stop at answers. When EPM Agent builds a data form or a business rule, it writes a readable specification from the reviewed requirements, and a compiler turns that specification into Oracle's native format. And nothing changes in a production environment until a person approves it.

The result is an agent you can hold to account. Every answer and every artifact can be traced back to the meeting, email or document it came from.

If that's the kind of help your EPM team needs, we'd be glad to show you.