Finance teams are entering a new stage of AI adoption. The conversation is no longer limited to whether AI can generate a forecast, summarize a variance or identify an anomaly. The real challenge is how finance can adopt AI in a controlled, measurable and repeatable way.

Oracle Cloud EPM Intelligent Performance Management, or IPM, is built around that need. IPM brings AI and machine learning into EPM processes such as planning, forecasting, insights, predictive analysis and narrative reporting. Oracle’s July 2025 Cloud Customer Connect session on IPM – AI in EPM Best Practices, Use Cases and Implementation Considerations outlines how organizations should select use cases, prepare data, run pilots, tune models and move from experimentation to steady-state adoption.

For CFOs, FP&A leaders, finance transformation teams and EPM administrators, the message is clear: AI in EPM delivers the best results when it is treated as a finance operating capability, not a one-time technical feature.

Product Direction Note

Oracle includes a safe harbor statement for forward-looking information. Future functionality, release timing, pricing and product direction may change at Oracle’s discretion and should not be treated as a commitment to deliver specific functionality. For organizations planning AI adoption, this means roadmap items should be used for readiness and strategy, while implementation decisions should be validated against the customer’s own Oracle Cloud EPM environment, subscription and release status.

AI in Oracle Cloud EPM Is Expanding Continuously

Oracle Cloud EPM has built AI capabilities over several years. The innovation timeline includes Predictive Planning, Auto Predict, IPM Insights, Bring Your Own ML, Predictive Cash Forecasting, Gen AI for Management Reporting Narratives, Gen AI in Insights, Insights in Smart View, Dynamic Parent Insights and Predictions, PCM Agent and Advanced Predictions. This progression shows that AI in EPM is not a single feature area. It is becoming a broader layer across forecasting, variance analysis, narrative generation, profitability management, planning and process automation.

For finance leaders, this creates an opportunity to rethink how planning and analysis work:

Forecasts can be compared against AI-generated predictions.
Anomalies can be surfaced automatically.
Variance commentary can be supported by Gen AI.
Parent-level predictions can reduce dependency on incomplete granular data.
Advanced Predictions can use multiple drivers to improve forecast relevance.
AI use cases can be expanded gradually across finance and operational planning.

Licensing and Availability: Start with the Right Entitlement View

Before launching IPM initiatives, organizations should understand which capabilities are available for their application type and subscription. Oracle’s licensing and availability matrix shows that EPM Cloud Enterprise supports Insights, ML Model Import, Predictive Planning, Auto Predict, Advanced Predictions and Gen AI across modules-based, custom and free-form application types. Enterprise PCM and Tax Reporting show Insights and Gen AI availability, with predictive capabilities listed as roadmap for those areas. Financial Consolidation shows roadmap direction for Insights and Gen AI. EPM Cloud Standard and legacy environments have more limited availability. This matters because AI strategy should begin with realistic scope.

A strong readiness assessment should confirm:

Application type
Subscription model
Available IPM capabilities
Existing EPM design
Data availability
Security model
Business process fit
User readiness
AI maturity

Without this clarity, teams may select use cases that are attractive but not currently practical for their environment.

The Four-Phase IPM Implementation Model

Oracle recommends phasing IPM implementation for the best outcomes. The four phases are:

Use Case Selection
Pilot
Refinement
End-User Rollout and Steady-State Care

This is one of the most important best-practice messages. AI adoption should not begin with a broad enterprise rollout. It should start with the right use case, the right pilot group, the right success metrics and the right refinement cycle.

Phase 1: Use Case Selection

The first phase is about choosing the right business problem.

Oracle highlights three starting categories:

Automation of Analysis, supported by Insights and Gen AI in Narrative Reporting
Reliable Forecasting, supported by Predictive Planning, Auto Predict and Advanced Predictions
Task Automation, supported by PCM Agent and Transaction Matching Agent as a short-term roadmap item

Use case selection should consider two factors: ease and impact. A simple use case can help teams learn quickly. Examples include OpEx predictions or basic insights where historical data already exists. A high-impact use case can create stronger FP&A buy-in, such as revenue, COGS, sales volume, cash flow, headcount, inventory or top operating expense areas. The strongest approach is to balance early wins with business relevance. A use case should be simple enough to launch, but meaningful enough to matter.

High-Value Use Case Areas for AI in EPM

Oracle identifies several business challenges where IPM can support financial and operational improvement. These include demand risk and rising costs, sustainability planning, service-based business models, workforce retention and recruitment, cash forecasting for accounts receivable and payable, project financial management, and sales performance management.

Examples include:

Demand volatility with rising material, labor and overhead costs
ESG planning and reporting challenges
Forecasting service call volumes and predictive maintenance
Planning headcount, attrition, leaves and workforce demand
Improving visibility into working capital drivers
Forecasting project revenue, costs and margins
Setting sales targets, commissions and territory plans

This is important because AI in EPM should not be restricted to finance-only accounts. IPM can also support operational planning where finance needs better predictive visibility.

Data Readiness: The Foundation of IPM Success

AI in EPM depends on usable data. Oracle’s guidance is direct: organizations should ensure the required data exists in the EPM application for the data slices being used, and they should validate this through a Smart View report created with those intersections. Oracle also recommends that data quality should be pre-processed, accurate and consistent. For meaningful insights or predictions, organizations should have at least 2x historical data. For example, if the goal is to predict one year, the organization should have two years of historical data. For better back testing, three years of historical data may be preferable.

Phase 2: Pilot Design

A successful IPM pilot needs the right participants and the right resource plan.

Oracle identifies key participants as:

Project Sponsor
Finance User
EPM Admin
Data Analytics resource, optional

Each role has a distinct responsibility. The project sponsor provides executive support, strategic alignment, stakeholder management and change management. The finance user identifies business use cases, input drivers, KPIs and testing needs. The EPM admin configures IPM jobs, ensures data availability, prepares forms and dashboards, manages additional rules or calculations and supports integration. Data analytics support is optional because EPM handles much of the AI capability behind the scenes, but analytics resources can still help with accuracy measurement, driver selection and model comparison. A pilot should not be treated as a technical demo. It should be a controlled business test with clear ownership, time commitment and success criteria.

Recommended Pilot Commitments

Oracle provides a practical resource plan for an IPM pilot. The EPM Admin may require 5–15 hours per week during the pilot and 2–4 hours per month after go-live. Finance users may require 10–15 hours per week during the pilot and 1–2 hours per month after go-live. Data analytics resources, if used, may require 5–10 hours per week during the pilot and 3–6 hours per week after go-live. This helps organizations set realistic expectations. Even though IPM is designed for finance users and does not require a large data science program, it still needs focused participation during the pilot.

Configuration Best Practices for IPM

Oracle recommends a design approach that considers both BSO and ASO. IPM works on both BSO and ASO, and organizations should enable Hybrid to use IPM features.

Oracle also provides practical guidance:

Predictions are best done in BSO because it helps baseline forecasts with predictions and compare them at the same level as the traditional forecast.
Insights are best done in ASO because ASO usually has more data and is often where actuals and reporting are maintained.
IPM works across different period granularities, including monthly, weekly, daily and custom periods.
Application design should be future-proofed for increased data volume required by IPM.

For implementation teams, this means IPM should influence application architecture. It should not be added as an afterthought.

Break IPM Jobs into Practical Groups

Oracle recommends breaking IPM jobs into multiple jobs because measures, thresholds and model settings can vary.

Jobs can be separated by:

Data slices such as entity, account or product group
Insights versus predictions
Different variations, such as with events and without events

This improves tuning and governance. A revenue prediction job may need different thresholds from an OpEx insight job. A sales volume forecast may need events, while a baseline expense forecast may not. Breaking jobs into logical groups makes the AI setup easier to monitor and easier to improve.

Store Predictions Separately for Traceability

One of the strongest configuration best practices is prediction storage. Oracle recommends creating a new Scenario, Version or Plan Element member to store predictions. This allows predictions to be compared against forecasts. Oracle also recommends adding an adjustment member that can adjust the forecast on top of the prediction so traceability is maintained.

A practical structure may include:

Forecast
Prediction
Adjustment
Prediction plus Adjustment

This design helps finance teams’ separate machine-generated output from human judgment. It also creates better transparency when explaining why the final forecast differs from the AI prediction.

Measuring Success: Accuracy and Forecast Value Add

Oracle recommends measuring success using comparative accuracy against actuals and different forecast methods. It also highlights Forecast Value Add, or FVA, which measures improvement over the existing forecast. This is critical. A prediction is not valuable because it looks sophisticated. It is valuable when it improves forecast quality, reduces manual effort, supports better decisions or helps users identify risks earlier.

Success metrics should include:

Accuracy against actuals
Comparison to existing forecast methods
Forecast Value Add
Speed of analysis
User satisfaction
Adoption rate
Reduction in manual review effort
Business confidence in outputs

Back Testing: Proving Prediction Quality

Oracle recommends at least 24 months of historical actuals to support algorithms, especially when seasonal patterns are relevant. Ideally, 36 months of history should be available to support back testing on one year of data. Back testing allows finance teams to generate predictions on historical periods and compare them with actuals. Historical forecasts are optional, but Oracle identifies forecast archival as a best practice because it allows organizations to ask whether predictions are better than the forecast. This matters because trust in AI increases when users can see evidence. Back testing gives finance teams a structured way to evaluate prediction quality before relying on it in active planning cycles.

Dynamic Parent Predictions and Insights

Oracle highlights Dynamic Parent Predictions and Insights as a way to improve flexibility and eliminate workarounds. This is useful when leaf-level data is incomplete, spotty or too detailed for meaningful prediction. It is also useful when users do not want insights at very detailed account levels.

Use cases include:

OpEx forecasting at total operating expense level
Department-level headcount planning
Program-level project forecasting
Revenue planning by brand, product group or territory
Marketing spend planning by spend category

This is highly practical for FP&A teams. Many organizations do not have clean historical data at every detailed level, but they may still have reliable parent-level data that is useful for planning.

Phase 3: Refinement and Model Tuning

After the pilot, teams should refine the model. Oracle identifies tuning areas such as accuracy, performance, advanced options and end-to-end business process impact.

This is where AI adoption becomes serious. Teams need to ask:

Are the predictions accurate enough?
Are the thresholds appropriate?
Are anomalies useful or too noisy?
Are events improving the forecast?
Are users acting on insights?
Are model settings aligned to business needs?
Has the finance process changed because of AI?

The goal is not only to make the model run. The goal is to make the model useful.

Model Settings and Metrics

Oracle identifies several tuning options for Auto Predict and Insights. For Auto Predict, tuning options include events, data screening, data attributes, methods and missing value treatment. Events can represent recurring, one-off or skip events such as holidays, supply disruptions or pandemics. Data screening supports missing value and outlier treatment. Methods allow selection between seasonal and non-seasonal approaches based on the data. For Insights, events can be tagged for anomalies, and metric methods can be selected based on the data being used. This reinforces a key point: AI results improve when model settings are matched to the business context.

Choosing the Right Accuracy and Variance Metrics

Oracle identifies several methods for anomaly, variance and bias analysis. For anomaly detection, methods include Z-score, Modified Z-score and Interquartile Range. Z-score is appropriate when data is roughly normal and clean. Modified Z-score is better when data has outliers or smaller sample sizes. Interquartile Range works well when data is not normally distributed. For variance analysis, methods include Mean Absolute Percentage Error, Mean Absolute Deviation, Root Mean Squared Error, Total Deviation and Total Deviation Percentage. Oracle identifies MAPE as a useful default when users are unsure. For variance bias, Oracle identifies Mean Percentage Error and Relative Percentage Difference, with RPD helping visualize the direction and magnitude of bias over time. This is important because different metrics answer different questions. Accuracy is not one universal number. Finance teams need to select the metric that best reflects the business impact of forecast error.

Advanced Predictions: Multivariate Forecasting for Finance

Oracle describes Advanced Predictions as multivariate predictions. They enable more powerful predictions by using more sophisticated algorithms and correlating the forecast with multiple provided data points. The capability is designed to be easy to configure through a step-by-step wizard that automatically selects the best algorithms, and it is embedded into EPM for finance users. This is a major advancement over basic univariate forecasting.

A univariate prediction relies mainly on the historical pattern of the target measure. A multivariate prediction can include operational drivers and contextual factors such as marketing promotions, average sales price, industry volumes or GDP rate. For finance, this is important because business performance is rarely driven by history alone. External and internal drivers often explain changes better than trend-based forecasting.

Phase 4: End-User Rollout and Steady-State Care

Oracle’s final phase focuses on change management, end-user education, user experience, success evaluation, model monitoring and continued evolution of business processes. This is where many AI programs succeed or fail. A model may be accurate, but users still need to understand how to interpret it. A forecast may improve, but the planning process must change to use it. An insight may detect anomalies, but business owners must know how to respond.

Steady-state care should include:

Model monitoring
Periodic accuracy review
User feedback sessions
Business process refinement
Documentation
Training refreshers
Ownership model
Change management communication
Expansion roadmap

AI in EPM should evolve with the business.

Overcoming Common AI Adoption Obstacles

Oracle addresses several common objections to AI adoption in EPM. One common concern is that data is not clean or adequate. Oracle notes that organizations need two to three years of historical data, but they do not need very large “big data” volumes to be successful with predictions. Insights do not need additional data beyond what typically exists, and EPM Cloud includes data-cleansing techniques.

Another concern is that teams lack AI knowledge. Oracle states that AI knowledge is helpful but not mandatory, because the product is designed to be implemented and maintained by EPM users with typical administrative skills such as loading data, running rules and building forms and dashboards. Oracle also addresses resistance to change, time constraints, existing AI engines and existing data science teams. The key argument is that IPM is embedded into finance processes and can complement broader AI or data science initiatives rather than replacing them.

Business Impact for CFOs and FP&A Leaders

Oracle Cloud EPM IPM can create value across planning, forecasting and analysis when implemented correctly.

The business impact includes:

Faster Analysis

Insights and Gen AI can reduce time spent manually reviewing variances and commentary.

Better Forecast Confidence

Predictive Planning, Auto Predict and Advanced Predictions help teams compare traditional forecasts with AI-generated predictions.

Stronger Planning Ownership

IPM is designed for finance users, reducing dependency on external AI platforms for every forecasting use case.

Better Decision Support

Multivariate predictions can include business drivers and contextual factors, making forecasts more aligned with real-world conditions.

Improved Governance

Separate prediction storage, adjustment members, back testing and FVA measurement help maintain traceability.

Scalable AI Adoption

A phased approach allows organizations to start small, refine and then scale.

How NexInfo Can Help

NexInfo helps organizations adopt Oracle Cloud EPM IPM with a practical focus on AI readiness, data quality, use case selection, configuration, governance and user adoption.

NexInfo’s delivery model is supported by ISO 9001 for Quality Management and ISO 27001 for Information Security, helping organizations strengthen process quality, delivery discipline and secure enterprise transformation practices. NexInfo has also received the AI-Enabled Workforce Excellence Award at the 1st Annual Long Beach Business AI Summit, reinforcing its focus on practical AI adoption across workforce enablement, enterprise systems and operational transformation.

NexInfo can support organizations with:

Oracle EPM IPM readiness assessment
AI use case selection and prioritization
Data readiness review
Smart View validation of data slices
IPM pilot planning
EPM configuration for BSO, ASO and Hybrid readiness
Insights configuration
Auto Predict setup
Advanced Predictions implementation
Driver selection and model tuning
Prediction storage and adjustment design
Back testing and Forecast Value Add measurement
Dynamic Parent Predictions and Insights planning
User training and change management
Steady-state monitoring and managed support

NexInfo helps finance teams move AI in EPM from experimentation to measurable operating value.

Conclusion

Oracle Cloud EPM IPM gives finance teams a structured way to bring AI into planning, forecasting and analysis. But success depends on disciplined adoption. Organizations need the right use case, enough historical data, a focused pilot, strong configuration, meaningful success metrics, proper model tuning and a clear rollout plan. They also need business ownership, change management and steady-state monitoring.

The strongest IPM programs will not try to implement every AI capability at once. They will start with practical use cases, prove value through back testing and Forecast Value Add, then scale into higher-impact areas such as revenue, COGS, workforce, cash flow, project performance and sales planning.

NexInfo helps organizations adopt Oracle Cloud EPM IPM with Oracle EPM expertise, ISO-certified delivery governance, AI-enabled transformation experience and a finance-first implementation approach.

FAQ
What is Oracle Cloud EPM IPM?

Oracle Cloud EPM IPM, or Intelligent Performance Management, is Oracle’s AI and machine learning capability for EPM. It supports features such as Insights, Auto Predict, Predictive Planning, Advanced Predictions, Gen AI in Insights and Gen AI for management reporting narratives.

What are the main phases of IPM implementation?

Oracle recommends four phases: Use Case Selection, Pilot, Refinement, and End-User Rollout with steady-state care.

How should organizations select an IPM use case?

Organizations should consider ease, impact, AI maturity and data availability. Simple use cases such as OpEx predictions can create early wins, while high-impact areas such as revenue, COGS, sales volume, headcount and cash flow can drive stronger business buy-in.

How much historical data is needed for IPM predictions?

Oracle recommends at least 24 months of historical data for meaningful predictions and ideally 36 months for back testing on one year of data.

Does data need to be perfectly clean for IPM?

No. Oracle notes that data does not need to be perfectly clean, and EPM Cloud includes data-cleansing techniques. However, data should still be pre-processed, accurate and consistent for the required data slices.

Who shouldparticipatein an IPM pilot?

Key participants include a project sponsor, finance user, EPM admin and optional data analytics support. The finance user provides business knowledge, while the EPM admin manages configuration and data readiness.

Should IPM predictions be stored separately?

Yes. Oracle recommends creating a new Scenario, Version or Plan Element member to store predictions and adding an adjustment member to maintain traceability between prediction and human adjustment.

WhatisForecast Value Add?

Forecast Value Add, or FVA, measures whether the AI prediction improves over the existing forecast. Oracle identifies FVA as a key success metric for measuring IPM value.

What is back testing in IPM?

Back testing uses historical actuals to generate predictions for past periods, then compares those predictions with actual results or historical forecasts. It helps finance teams evaluate prediction quality before production use.

What are Dynamic Parent Predictions and Insights?

Dynamic Parent Predictions and Insights allow AI to work at parent-level members when detailed leaf-level data is incomplete, spotty or too granular. Use cases include OpEx, headcount, project budgets, revenue by category and marketing spend.

What are Advanced Predictions in Oracle EPM?

Advanced Predictions are multivariate predictions that use multiple drivers and more sophisticated algorithms. They are designed for finance users and configured through a step-by-step wizard inside EPM.

What is the difference between univariate and multivariate prediction?

Univariate prediction relies mainly on historical data for the target measure. Multivariate prediction can include operational drivers and contextual factors such as promotions, average sales price, industry volumes and GDP rate.

What model metrics are used in IPM?

Oracle identifies metrics and methods such as Z-score, Modified Z-score, Interquartile Range, MAPE, MAD, RMSE, Total Deviation, Total Deviation Percentage, Mean Percentage Error and Relative Percentage Difference.

Do finance teams need data scientists to implement IPM?

Oracle states that AI knowledge is helpful but not mandatory. Many EPM customers can self-implement because IPM is designed for EPM users with typical administrative skills.

How can NexInfo help with Oracle EPM IPM?

NexInfo can help with IPM readiness, use case selection, data validation, pilot planning, Insights setup, Auto Predict, Advanced Predictions, model tuning, back testing, Forecast Value Add measurement, user adoption and managed support.

Oracle Cloud EPM IPM Best Practices: Turning AI in Finance into a Governed Business Capability