What AI Actually Does Inside Oracle EPCM in 2026
Does Oracle EPCM have AI now? The honest answer is yes. It does less than the slides suggest, though, and more than most finance teams are using.
This is the first post in a short series on AI inside Oracle Enterprise Profitability and Cost Management. I want to start with a plain inventory: what Oracle ships today, what it does well, and where the gaps are. Everything here comes from Oracle's own documentation and release notes up to the September 2026 update.
Three things you can use today
1. The PCM Agent builds and runs the model on request
The PCM Agent is a generative AI assistant embedded in EPCM. You describe what you want in plain English and it turns the request into system actions. It can create a model, add rule sets, write or edit allocation rules, change members in bulk and run a calculation for a point of view. Since the 26.09 update it also handles trace commands, so you can ask where a cost came from instead of clicking through the trace screens.
Two design choices matter here. First, the agent shows you a preview and waits for confirmation before it changes anything. Second, since 26.08 it lives inside the modeling screens (Models, Waterfall Setup, Mass Edit, Calculation Control) instead of in a separate window. The original standalone agent is being retired over the 26.10 and 26.11 updates.
In practice this shortens the slowest part of an EPCM build. Turning a whiteboard allocation waterfall into executable rules used to take days of careful clicking. Now an administrator can describe the flow and review what the agent proposes.
2. IPM Insights watches the numbers
IPM Insights has been available for EPCM since the 24.08 update. It scans the data for anomalies and variances that deserve attention and surfaces them as insights, instead of waiting for someone to find them in a report. Since 25.04, generative AI can summarize those insights in a short narrative.
On a profitability model this is useful for one job: spotting the customer, product or branch whose margin moved in a way nobody planned. It tells you where to look. It does not tell you why the margin moved, which is the question the CFO will ask next.
3. Narrative summaries on top of the model
The generative summaries are the most visible feature and the least important one. They save time on the first paragraph of a variance commentary. The analysis behind that paragraph still has to come from the model.
What is not there yet
The gaps are as instructive as the features.
No agent chooses your drivers. The PCM Agent writes the rules you ask for. It does not know whether shipment count or weight is the right driver for your pickup activity, and it will happily build a precise model on the wrong one. Driver selection is still a human design decision, and it is still where most profitability models go wrong.
No conversational layer for EPCM in Oracle's agent program. In September 2026 Oracle launched an Implementation Success Program for customers who want conversational interfaces built with ChatGPT, Codex, Claude or their own agents. Its scope covers Planning and FreeForm. EPCM is not in it. Teams that want an executive to ask "why did margin drop in the northern region?" and get an answer grounded in the allocation model have to build that layer themselves, through EPCM's REST API.
No explanation in business terms. Trace shows the path a cost took through the waterfall. Turning that path into "these three customers absorbed most of the new service cost because their contact rate doubled" is still analysis work.
The part nobody puts on the slide
Every AI feature in EPCM sits on top of the model. If the model is clean, the agent is fast and the insights are meaningful. If it isn't, AI makes the mess faster.
Implementers who work with the PCM Agent already say it plainly: the main reason a request fails is that the request was poorly formed. Behind a poorly formed request there is usually a poorly governed model, with inconsistent member names, rule sets nobody documented and drivers that changed meaning between versions. An agent can't infer intent from a structure that has none.
That is why I treat AI readiness in EPCM as a design question first and a technology question second.
Three questions before you switch it on
- Can someone outside finance read your rule names and understand the flow? If not, the agent won't either. Fix naming conventions before you write the first prompt.
- Is every driver documented with its source and its owner? IPM Insights will flag a margin anomaly. Someone has to know whether the cause is the business or a driver feed that broke.
- Who confirms what the agent proposes? Preview and confirmation only protect you if the person approving understands allocation logic. Decide who that is before you go live.
What comes next in this series
The next post goes deeper into the PCM Agent: what it does well, where it gets stuck and how to structure a waterfall so that plain-English requests produce the rules you meant. After that: IPM Insights on a profitability model, why AI fails on a badly designed cost model, and how to build the conversational layer that Oracle's program leaves out.
If you are evaluating AI in EPCM and want a second opinion on your model's readiness, 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, EPM Features with AI (Oracle Fusion Cloud documentation); Oracle EPM 26.08 and 26.09 update notes as summarized by Perficient and fmepm.com; NexInfo, Introducing the Oracle PCM Agent (March 2026).
