IPM Insights on a Profitability Model: Which Alerts Deserve a Look

The first two posts in this series covered what AI does inside Oracle EPCM and how the PCM Agent writes rules. This one looks at the feature that watches the results, IPM Insights.

On paper it is simple. Insights scans the model's data, flags values that moved more than expected and ranks them by size. In practice, a profitability model is a hard place to switch it on with default settings. Allocation models move for reasons that have nothing to do with the business. An alert list full of those reasons teaches executives to stop reading it within a quarter.

What Insights does in EPCM

Oracle's documentation lists three kinds of insight for EPCM:

• Anomaly. A value far from the rest of its history. The threshold is a z-score, typically three standard deviations.
• Period movement variance. A significant change between the current period and a benchmark period, measured as a percentage or an absolute amount.
• Forecast bias. A persistent gap between two historical scenarios, such as forecast and actual.

An administrator defines the data slices to analyze, the thresholds and the High, Medium and Low impact bands. Insights run on demand or as a scheduled job, for example after each actuals load. The dashboard lists them by size of deviation. Generative AI can summarize them in a short narrative, and since April 2026 that narrative can be anchored on the Account dimension to show which child accounts drive a parent insight.

Two constraints shape the design. Users only see insights on slices they have access to. And with multiple currencies, absolute-value impact thresholds aren't available, so impact bands have to be set in percentages.

Why a profitability model generates noise

A planning model changes when someone changes a number. A profitability model also changes when someone changes the logic. Four sources of movement have nothing to do with performance:

1. A driver feed breaks. Last month's ticket counts didn't load, so IT cost lands on whatever products had tickets in the file. Margins swing. Nothing happened in the business.
2. Someone changes a rule. A new occupancy driver shifts cost between branches from one month to the next. The change is intended, and it looks exactly like an anomaly.
3. Periodic allocations. Costs allocated once a quarter or once a year produce a spike in the same period every cycle.
4. Small denominators. A product with few customers shows huge percentage swings on trivial amounts.

Left at default settings, Insights reports all four, ranked by size. In a profitability model the largest deviations are often mechanical, so the top of the list can be the least useful part of it.

Figure 1. Three checks separate model mechanics from a real business change. Only the last branch goes to the business owner.

Where to point it

These are the settings I recommend before the first run.

• Watch the drivers as well as the margins. Run anomaly insights on driver quantities: tickets, transactions, square meters, headcount. When a driver and a margin move together in the same slice, check the feed before you call the business. A driver insight is the cheapest data-quality control the model can have.
• Set slices where someone can act. An insight on "total bank, all products" has no owner. One on a product line in a region has a name attached to it.
• Use the calendar. Oracle lets you register events so expected spikes don't trigger anomalies. Put periodic allocations, rate resets and known reorganizations there.
• Keep materiality in view. Use the impact bands so a 400% swing on a few thousand dollars doesn't outrank a 6% swing on a key account. If you run multiple currencies, control materiality through the slices instead, leaving out the ones too small to matter.
• Run after the calculation. Schedule the job after the allocation run, not after the data load. Insights on pre-allocation data describe the ledger.
• Respect the history. Statistical tests need comparable periods. If the model was restructured last year, the older periods may not mean the same thing, and the first months of alerts deserve extra skepticism.

Make dismissals count

Each insight can be dismissed with a reason and a remark, and its status changes to Closed for everyone. Most teams treat that as housekeeping. I'd treat it as the most useful data the feature produces.

Agree on a short list of reasons, such as driver data issue, rule change, expected event and real business change, and require one every time. After a quarter, the dismissal log tells you which feeds break, which rules keep shifting cost and which thresholds are too tight. Tune the configuration from that log.

The question Insights leaves open

Insights tells you where a margin moved. It doesn't tell you why. The generative summary describes the movement in sentences. Explaining it means tracing the cost through the waterfall and connecting it to a driver, and that is still analysis work.

The rest of the series picks up from there. Part 4 looks at why AI fails on a badly designed cost model. Part 5 covers the conversational layer that would let a CFO ask "why" and get an answer grounded in the model.

If you're about to switch on Insights on a profitability model and want a second look at the configuration, 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, Configuring IPM Insights (EPCM administration guide); Oracle, Analyzing Insights (EPCM administration guide); Oracle, Considerations for IPM Insights; Oracle EPM April 2026 What's New, IPM Insights: Generative AI Narrative Summaries Available for Account Dimension.