The PCM Agent: From Waterfall Sketch to Rules in Plain English

The first post in this series mapped what AI does inside Oracle EPCM today. This one goes deep on the feature that changes daily work the most, the PCM Agent.

Oracle calls it a "guided natural language tool." That word "guided" matters more than "natural language." The agent is not a chatbot you can talk to like a colleague. It is a command layer that accepts structured English, turns it into modeling actions and waits for you to confirm them. Used that way, it saves real time. Used like a chatbot, it frustrates people quickly.

What it can do

According to Oracle's documentation and tutorial, the PCM Agent handles these actions:

• Build: create a model, a rule set or a rule.
• Edit in bulk: add, replace or delete members across rules, and enable or disable rules.
• Reuse: copy rules within a model or into another model.
• Run: calculate a model for one or several points of view.
• Explain: list rules and run a trace on an allocated amount.

Since the 26.08 update it sits inside the modeling screens (Models, Waterfall Setup, Mass Edit, Calculation Control), and it can create several points of view in one request. Since 26.09 it also takes trace commands from there. The original standalone agent is being retired during the 26.10 and 26.11 updates.

To use it, an administrator has to enable generative AI in the application settings. Oracle also notes that it works in English only and that generative AI is available in OCI environments, though not yet in every OCI region.

What a good request looks like

Here is a request taken from Oracle's own tutorial:

Calculate model '10 Actuals Allocation Process' povs FY23::Jan::Actual::Working run from rule 'Activity Costing Assignments'

Notice what makes it work. The action comes first. Model and rule names sit in single quotes, exactly as they exist in the application. The point of view follows a fixed pattern. This is plain English the way a well-written procedure is plain English: precise, ordered and without ambiguity.

Two features help. Smart Lookup lets you type an artifact type followed by @ and search for the exact name without leaving the screen, so you don't have to remember whether the rule is called "Occupancy Allocation" or "Occupancy Expense Allocation." And every request shows a preview of the parsed command, with the parameters it understood, before anything runs. You can refine and resend until it matches what you meant.

Figure 1. A request that lands has four parts: the action, exact names in quotes, the point of view and the options.

Where it saves time

Four jobs get noticeably faster.

1. Scaffolding a new model. Creating the model, its rule sets and empty rules in the right order used to be a morning of clicking. Now it's a sequence of requests you can review in one sitting.
2. Mass edits. Replacing a retired cost center or a renamed product line across dozens of rules is where manual work breaks down. The agent lists the rules, proposes the replacement and runs it after you confirm.
3. Actuals to plan. Copying a proven set of allocation rules from the actuals model into the plan model is a single request.
4. Answering "where did this cost come from?" A trace request on a customer or product returns the allocation path without navigating the trace screens.

Where it stops

Oracle is candid about the limits, and they are worth repeating.

Rules come out as an outline. The documentation says rules and rule sets created through the agent have an initial outline. Targets, formulas and validation still have to be finished by hand. The agent drafts the rule. It doesn't design it.

Accuracy is your responsibility. Oracle states that generative AI output "may not always be factual, accurate, or appropriate for modeling commands" and that you are responsible for reviewing it. The preview is the control. It only works if someone reads it.

It knows names, not meaning. The agent can find a rule called 'Occupancy Expense Allocations.' It can't tell you whether square meters is the right driver for occupancy in a shared-services center where half the staff works remotely.

Figure 2. The agent removes the clicking. The design decisions stay with the people who own the model.

How to structure a waterfall so requests land

Everything above points to the same conclusion. The agent works best on a model built to be read. These are the conventions I recommend before anyone switches it on.

• Number models and rule sets by stage. Oracle's own sample uses names like '10 Actuals Allocation Process' and '40 Plan Allocation Process.' The number tells a person and the agent where the step sits in the flow.
• One rule set per allocation stage. Resources to activities, activities to cost objects, then reporting adjustments. Mixing stages in one rule set makes every request ambiguous.
• Rule names that say source, driver and target. "IT Cost to Products by Tickets" is a request waiting to happen. "Rule 14b" isn't.
• Readable member aliases. If finance users can't read the dimension members, they won't write good requests against them.
• A short prompt library per model. Ten tested requests for the jobs your team runs every month: the close calculation, the standard trace, the quarterly member update.
• Keep the history. The agent lets you export the full request history. Store it with the model. It becomes a practical change log of who asked for what and when.

Who presses Run Command

The last decision is the one teams skip. Every request ends with a person pressing Run Command. That person should understand allocation logic well enough to spot a wrong target in the preview. In most organizations that is the model owner in finance, not IT and not the newest analyst on the team.

Next in the series

Part 3 looks at IPM Insights on a profitability model. Which anomalies deserve an alert, which are noise, and how do you keep insights from training executives to ignore them?

If your team is about to switch on the PCM Agent and wants a review of the model's structure first, 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, Working with PCM Agent (EPCM tutorial); Oracle, Using Smart Lookup When Creating a User Request (EPCM documentation); Oracle EPM 25.05 readiness note on the PCM Agent; Oracle EPM 26.08 and 26.09 update notes as summarized by Perficient and fmepm.com.