Written by Zhang — Paddy & Water Inc. / Data & Analytics

A Technical Deep Dive and Implementation Design Guide for BI, Planning, and AI/ML Capabilities

June 2026

Introduction: The Analytics Democratization That SAC Aims For

SAP Analytics Cloud (SAC) is SAP’s flagship analytics product, unifying “BI, planning, predictive analytics, and AI” on a single cloud platform. It first appeared in 2016 under its former name, SAP BusinessObjects Cloud, and since then SAP has pursued a strategy of progressively consolidating its on-premise analytics tools — BW/4HANA, Crystal Reports, Lumira, Analysis for Office, and others — into SAC.

At the core of SAC’s design philosophy is the realization of “an environment where not only data specialists but also business users and executives can explore, analyze, and plan data themselves.” Supporting this “democratization of analytics” is the set of Augmented Analytics (AI-powered advanced analytics) capabilities detailed later in this article. This article goes into technical implementation detail, covering data model design, connection architecture, AI/ML technical specifications, security design, and the details of Planning.

1. SAC Architecture and Technology Foundation

Foundation as a Multi-Tenant SaaS

SAC is a multi-tenant SaaS that runs on SAP BTP. Users access it solely through a web browser (Chrome or Edge recommended), with no client software installation required. SAC tenants are isolated by region (data center location), allowing customers to select a region that meets data residency requirements (such as GDPR or Japan’s Act on the Protection of Personal Information). Japan regions (jp1, jp10) are available, allowing data residency to be restricted to within Japan.

SAC’s Data Processing Engine

SAC’s analytical processing runs on the “SAP HANA Cloud engine.” SAC’s data models (Analytic Models / Planning Models) are stored in HANA Cloud’s column-oriented in-memory database, achieving high-speed aggregation and multidimensional analysis that leverages HANA’s in-memory processing power. However, users do not need to be aware of HANA Cloud itself; memory and storage capacity are managed through SAC’s administration console (tenant administration).

The Three Major Functional Areas and Licensing Model

SAC is composed of three functional areas — “BI & Visualization, Predictive Planning, and Enterprise Planning” — and licenses are purchased as a combination of these.

• BI Professional: Enables creating stories and dashboards, building data models, and using AI analysis features.

• Business Intelligence (BI): Viewing and interacting with stories. Does not include creation rights.

• Predictive Professional: A higher-tier license that includes the rights to build and manage Smart Predict (machine learning prediction models).

• Planning Professional / Standard: Rights to use planning features (data actions, planning sequences, actual-vs-plan management). Divided between creators and viewers.

The exact combination of license types is determined by the contract with SAP, but the starting point for license selection is defining roles: “who creates stories and who views them” and “who enters plans and who approves them.”

2. Technical Details of Data Connectivity and Model Architecture

Two Major Categories of Connection Methods: Live Connection and Import

SAC’s data connection methods fall broadly into two categories: “Live Connection” and “Import (Acquired Data).” This choice has a significant impact on performance, data freshness, and functional constraints.

Live Connection

With a live connection, every time a query is executed in SAC, a SQL/MDX query is issued to the connected data source, retrieving the latest data in real time. SAC does not retain a copy of the data.

• Supported data sources: SAP BW (BW/4HANA, BW on HANA, BW 7.5 and later), SAP HANA Cloud, SAP HANA on-premise, SAP Datasphere, SAP S/4HANA (CDS view-based), and SAP BPC (partial).

• Technical characteristics: Because queries are executed on the connected database (HANA, etc.), SAC-side memory and compute resources are consumed only for UI rendering. The performance of the connected database directly determines query performance. Data is always current (if the connected database is updated in real time, the SAC screen also reflects those updates).

• Constraints: SAC’s AI features (Smart Insights, Smart Predict, etc.) cannot be used with data models based on a live connection. This is because, with a live connection, the data does not reside within SAC itself, making AI processing impossible. If AI is needed, consider using the import method or replicating data to HANA Cloud. Blending (combining multiple data sources) is also restricted for live connections.

Import Connection (Acquired Data)

An import connection is a method in which data is copied from the data source and stored in SAC’s in-memory database.

• Supported data sources: In addition to HANA, BW, S/4HANA, and Datasphere, more than 80 sources are supported, including OData, REST APIs, SAP Concur, Salesforce, Google BigQuery, Snowflake, Azure Synapse, CSV/Excel, and SAP Business ByDesign.

• Technical characteristics: Data is copied into SAC’s internal HANA Cloud storage. Because query processing runs on SAC’s internal HANA engine, it does not depend on the performance of the source system. A scheduler is configured for periodic loading (daily, hourly, or near-real-time frequency settings are possible).

• Constraints: Data freshness depends on the load schedule (it is not real time). Loading large volumes of data consumes SAC’s memory quota (tenant capacity). If the tenant’s data capacity limit is exceeded, the data load will fail with an error.

Connection Infrastructure: Connecting SAP and On-Premise Systems

Connecting via SAP Cloud Connector (SCC)

Connections to on-premise SAP BW, HANA, and S/4HANA systems go through SAP BTP’s Cloud Connector (SCC). SCC is proxy software installed on the internal corporate network that establishes an encrypted tunnel between SAC and the on-premise database. Because it uses a “reverse tunnel” approach that requires no inbound ports to be opened to the internet, it is characterized by lower network security requirements.

• SCC configuration considerations: Because SCC carries the risk of becoming a single point of failure (SPOF), it is recommended to deploy multiple SCC instances in a master/shadow configuration for high availability (HA). SCC is tied to an SAP BTP subaccount, and both BTP and SAC licenses are required.

Live Connection with SAP Datasphere (Recommended Architecture)

A live connection to SAP Datasphere is the “semantic layer integration recommended by SAP.” In this configuration, Datasphere handles data integration, virtualization, and semantic enrichment, while SAC references Datasphere’s models live for use in BI and planning.

• Datasphere’s role: Datasphere integrates and virtualizes data from multiple sources (HANA Cloud, S/4HANA, Snowflake, etc.) and exposes business-semantic (Business Layer) models to SAC.

• Transparency of row-level security: Row-level security (Data Access Control) defined on the Datasphere side is transparently applied to SAC’s live-connection queries as well. Users automatically see only the data range they are permitted to access within SAC.

Types of SAC Data Models and How to Choose

Analytic Model

The Analytic Model is a data model for SAC’s BI reporting. It defines imported data in a star schema (fact table plus dimension tables).

• Calculated Measures: Define calculation formulas on top of existing measures at the model level, such as “year-over-year sales growth (Current Year / Prior Year – 1).” Calculations run on HANA, delivering excellent performance.

• Restricted Measures: Define, at the model level, aggregated values filtered by specific dimension values (for example, a “Japan sales” measure restricted to Country = Japan).

• Hierarchy: Define parent-child hierarchies (Product Category → Product Family → Product) or level hierarchies (Year → Quarter → Month → Day) in the model to enable drill-down analysis. There are two types of hierarchies: “attribute-based” and “level-based.”

• Row-Level Security (Data Access Control): The model’s Data Access Control (DAC) settings link “user IDs or teams” to “dimension values (such as region or product category)” to restrict the range of data a user can see at the row level.

Planning Model

The Planning Model is a specialized model for xP&A (Enterprise Planning). Unlike the Analytic Model used for BI, it has planning-specific capabilities: “entry of planning data, version management, allocation calculations, and write-back.”

• Version: A planning model maintains an “Actuals version” and a “Plan version” side by side. Actuals are loaded via import or live connection, while plan versions are entered by users. Multiple planning scenarios (Base, Best Case, Worst Case) can be managed in parallel as versions.

• Private Version: Interim data being entered during planning is saved to a “private version” and is not visible to other users. Once entry is finalized, it is promoted to a “public version.”

• Calendar (Time Dimension): The time dimension is defined at a granularity such as fiscal year, calendar year, or weekly, and management of the “current plan period” is automated.

Blending: Combining Multiple Models

SAC provides a “blending (data blending)” feature that combines and displays data from multiple data models (or connections) within a story. It is equivalent to a SQL JOIN, but it can be configured with no code through SAC’s UI.

• Blend key: Specify the join key (a field where dimension values match) between the two models. Example: blending a “sales model” and a “budget model” on “product code and fiscal year-month” to show actual-versus-plan variance.

• Blending constraints: There is a constraint that blending a live-connection model with an import model can only go in one direction (the import model acts as the driver). Real-time blending across different data sources can also affect performance.

3. BI & Visualization: Technical Details

Designing a Story

A Story is SAC’s primary reporting and dashboard format. It is designed by combining three types of page layouts: Responsive Page, Canvas Page (freeform), and Grid Page.

• Responsive canvas: Components automatically resize and rearrange according to screen size. Support for desktop, tablet, and smartphone can be achieved within a single story. Recommended for mobile-first dashboard design.

• Freeform canvas: Components are placed freely, PowerPoint-style. Suitable for fixed-layout management reports and financial statement formats.

Advanced Interaction Through Scripting

SAC stories have a “Scripting” feature that lets you implement custom logic using a proprietary, JavaScript-like scripting language.

• Button Click event: On a button click, a script controls filter changes, page navigation, execution of data actions, pop-up displays, and more.

• onPageLoaded / onInitialization events: Automatically execute dynamic filter settings, variable initialization, and other tasks when a page loads.

• Application Design (formerly Analytic Application): Advanced interactive applications built with scripting — such as multi-step simulators and custom configurators — can be built on SAC. It offers a business-application experience similar to a Fiori app.

Enterprise Reporting: Pixel-Perfect Reports

Fixed-layout reports such as financial statements and statutory reports are designed using SAC’s “Enterprise Reporting (Enterprise Reports)” feature.

• Report designer: An Excel-like grid-based designer for placing cell references, formulas, tables, and charts.

• Analysis for Office (AfO) integration: BW/HANA reports created in the existing Excel Analysis for Office can also be connected to SAC. Because AfO works only as an Excel add-in, migration to web-browser-based Enterprise Reporting is recommended.

• PDF/Excel export: Enterprise Reports can be exported in PDF and Excel formats. Schedule-based automatic distribution (as an email attachment) can also be configured.

4. Enterprise Planning (xP&A): Technical Details

Data Actions: The Planning Logic Engine

Data Actions are the execution engine that defines planning-processing logic in SAC Planning, such as “calculation, allocation, copy, and conversion.” They are triggered by a button click within a story or by the scheduler.

Types of Data Action Steps

• Copy: Copies data between plan versions, for uses such as “copying 2025 actuals as the initial value for the 2026 plan.” Specify the source and target versions, time axis, and dimension filters for the copy.

• Allocation: Distributes higher-level plan values down to lower levels (top-down allocation). Defines processes such as “allocating the annual sales plan by month and by product.” Specify the allocation driver (prior-year actual ratio, even allocation, manual allocation percentage, etc.).

• Spread: Allocation along the time axis — the process of spreading an annual plan value across months. Choose a spread rule (even distribution, prior-year actual distribution, standard allocation, etc.).

• Advanced Formula: Defines custom calculations using a proprietary formula syntax (SQL-like), rather than ABAP. It supports “IF-THEN, aggregate functions, time functions, and dimension references,” making it possible to implement complex cost allocation, inventory calculation, and foreign exchange conversion calculations in a near no-code manner.

• Multi Action Step: A “nested data action” that calls another data action as a subroutine from within a data action. Complex planning processes can be split across multiple data actions to improve maintainability.

Advanced Formula Syntax

The Advanced Formula in a data action is written in the formula language (Data Action Script). The main syntax elements are shown below.

• RESULTLOOKUP: A function that references another measure, version, or dimension value. Frequently used for time-shifted references such as “same period, prior year.”

• FOREACH: A loop construct that iterates over dimension values. Used for processes such as “applying a different allocation rate for each product.”

• IF / ELSEIF / ELSE: Conditional branching. Used for calculation logic such as “post a bonus if the budget achievement rate is 90% or higher, otherwise post 0.”

• TIMESHIFT: A shifted reference along the time axis. For example, “TIMESHIFT([2026.01], -1)” references data for December 2025.

Planning Sequences

A planning sequence is a “chain of batch processes” that executes multiple data actions in order. A processing chain such as “Step 1: Copy actuals → Step 2: Currency conversion → Step 3: Cost allocation → Step 4: Profit calculation” is defined as a single sequence, and all steps run consecutively with a single button click.

• Parameter passing: When a sequence runs, parameters such as “target version, target year/month, allocation rate” can be passed in dynamically. The same sequence can be reused with different parameters.

• Scheduled execution: Planning sequences can be configured in SAC’s scheduler to run automatically on a recurring basis, such as “on the 1st of every month” or “every Monday at 6:00 AM.” This enables an automated planning flow such as “automatically run actual-versus-plan variance calculations and allocation calculations after actuals are loaded each month.”

Value Driver Trees and Strategic Planning

SAC’s Value Driver Tree is a feature that “visually represents the causal relationships between KPIs as a tree while running simulations in real time.”

• How it works: For example, a causal tree is defined such as “EBITDA ← Operating Profit ← Revenue ← Product Price × Quantity, Cost ← Fixed Costs + Variable Costs.” When you change the value of a terminal node in the tree (product price, sales volume, variable cost ratio, etc.), the impact is immediately reflected in EBITDA.

• Scenario comparison: Multiple scenarios (optimistic, base, pessimistic) can be compared by changing the tree’s values, allowing what-if analyses — such as “the EBITDA improvement from a 5% price increase” — to be presented visually to executives.

5. Augmented Analytics (AI-Powered Analytics): Technical Details

Overview of SAC’s AI Architecture

SAC’s AI features operate in conjunction with SAP AI Core on SAP BTP. When a user runs an AI feature in SAC, the SAC backend sends data to an AI Core endpoint, AI Core performs the ML processing or LLM inference, and returns the results to SAC. Users consume this simply as “a feature of SAC” without being aware of the underlying processing.

The data used for AI processing is data within SAC’s import data models. Because live-connection models cannot transfer data to AI Core, AI features are supported only for import models.

Smart Insights: Technical Details of Automated Driver Analysis

Smart Insights is an AI feature that automatically performs driver analysis to answer questions such as “why is this value high, low, or different from before.” It is launched simply by right-clicking a data point on a chart and selecting “Explain.”

Technical Processing Flow

Smart Insights’ internal processing runs through the following steps.

• Step 1, Identify the target: Sets the data point selected by the user (for example, January 2026 sales at -15% versus budget) as the analysis target.

• Step 2, Scan related dimensions: For every dimension in the model (region, product, customer, channel, etc.), it statistically calculates “when the target value is broken down by this dimension, which dimension values contribute most to the deviation.”

• Step 3, Apply statistical methods: Contribution analysis uses the coefficient of variation, decomposition of relative variance (variance driver decomposition), and Shapley-value-like contribution calculations.

• Step 4, Generate insights: Natural-language insights are automatically generated, such as “the sales decline in the North America region explains 63% of the total variance” or “the margin decline in Product Family X is the primary driver.”

Smart Insights Configuration Options

• Analysis depth: SAC settings let you control which dimensions are included as targets for driver analysis. Excluding irrelevant dimensions improves both analysis accuracy and processing speed.

• Comparison basis: Baseline periods and baseline values can be set, such as “period-over-period comparison (same period last year),” “plan comparison (budget),” or “moving average comparison.”

• Insight confidence score: Each insight is assigned a “confidence score,” and statistically more plausible explanations are displayed higher in the list.

Smart Discovery: Technical Details of Exploratory Analysis

Smart Discovery is an AI-driven exploratory data analysis (EDA) feature that automatically discovers “what factors are influencing this KPI, and where patterns, outliers, and correlations exist in the data.” The user simply specifies the “measure to analyze (target variable),” and SAC automatically scans all dimensions and measures to discover relationships.

Smart Discovery Processing Algorithms

• Correlation Analysis: Calculates the Pearson correlation coefficient and Spearman rank correlation between the target variable and each explanatory variable (dimension or measure). The strength and direction (positive or negative) of the correlation are shown as a ranking.

• Feature importance via Random Forest: Runs random forest regression for a numeric target variable, or random forest classification for a categorical target variable, to calculate “feature importance” — which explanatory variables are most important for predicting the target variable. SAC visualizes this as “Key Influencers.”

• Clustering (K-Means): Automatically classifies data points into K clusters (groups). The number of clusters, K, is determined automatically using the elbow method. The characteristics of each cluster (centroid values, within-cluster variance) are presented along with an explanation.

• Outlier Detection: Automatically detects and highlights data points that statistically exceed “mean ± n standard deviations” or the interquartile range (IQR) as outliers.

• Time-series pattern analysis: When the model includes a time axis, time-series decomposition into trend (rising, falling, flat), seasonality (weekly, monthly, or annual cyclical variation), and a random component (STL decomposition: Seasonal-Trend decomposition using LOESS) is performed automatically.

Smart Discovery Output Pages

After Smart Discovery runs, the following analysis pages are generated in an automatically created “Discovery” story.

• Overview page: Basic statistics for the target variable (mean, median, maximum, minimum, standard deviation, skewness) and a histogram of the distribution.

• Key Influencers page: A ranked list of feature importance for each explanatory variable — a ranking of “which variables best explain the target variable.”

• Unexpected Values page: Highlights outliers and anomalies. The context in which outliers occur (which combination of dimension values they occur under) is visualized.

• Relationships page: A scatter plot and regression line between the target variable and key explanatory variables.

• Clustering page: A scatter plot of the automatically classified clusters and a comparison table of characteristics by cluster.

Smart Predict: Technical Details of Machine Learning Forecast Models

Smart Predict (SP) is a service for building and managing no-code machine learning prediction models within SAC. Three types of predictive scenarios are offered, each using a different ML algorithm.

Classification Scenario

This model predicts a Yes/No outcome (binary classification) or a multi-class outcome, for questions such as “will this lead convert” or “will this customer churn.”

• Algorithm: Gradient Boosting Machine (GBM) is used as the primary algorithm. GBM is an algorithm that sequentially ensembles weak learners (decision trees) to minimize prediction error, and it delivers best-in-class accuracy on classification and regression problems for tabular data. SAC internally performs automatic hyperparameter optimization (an AutoML-like process).

• Model evaluation metrics: A confusion matrix, AUC-ROC curve, precision, recall, and F1 score are automatically displayed on the model evaluation screen. The “Gini coefficient” is used as SAC’s default evaluation metric (Gini = 2 × AUC − 1).

• Explainability: Each variable’s contribution to the prediction (variable contribution) is calculated and displayed as a SHAP (SHapley Additive exPlanations) value. The basis for an individual prediction — for example, “why this customer was judged to be at high risk of churn” — is explained via its SHAP values.

Regression Scenario

This model predicts a continuous numeric value, such as “how much revenue will be generated next month” or “how many units will the next order be for.”

• Algorithm: GBM regression is the primary algorithm. SAC may also internally perform automatic algorithm selection (AutoML) among Exponential Smoothing, Linear Regression, and Random Forest. The algorithm with the highest predictive accuracy is automatically adopted.

• Model evaluation metrics: RMSE (Root Mean Square Error), MAE (Mean Absolute Error), MAPE (Mean Absolute Percentage Error), and R² (coefficient of determination) are calculated automatically.

Time Series Scenario

This is a specialized scenario for forecasting future values of time-series data (daily, weekly, or monthly sales, demand volume, etc.).

• How algorithm selection works: SAC automatically compares the following time-series algorithms and selects the best-performing model. (1) Exponential Smoothing (ETS): An additive/multiplicative model of three components — trend, seasonality, and residual — including the Holt-Winters method. (2) ARIMA (Auto-Regressive Integrated Moving Average): A combination of autoregression, differencing, and moving average. The parameters of ARIMA(p,d,q) are automatically determined to minimize AIC/BIC. (3) Machine Learning Ensemble: An ensemble forecast that adds time-based features (day of week, week number, month, values from the previous N periods, rolling averages, etc.) to GBM as features.

• Automatic seasonality detection: Automatic FFT (Fast Fourier Transform) analysis of the time-series data automatically identifies seasonal cycles such as weekly, monthly, quarterly, or annual.

• Prediction intervals (confidence intervals): In addition to point forecasts, “80% confidence interval” and “95% confidence interval” bands are displayed to represent the model’s uncertainty.

• Calendar of Events: Registering holidays, promotion dates, and anomalous event dates as “calendar events” allows the model to separately estimate their effects to improve accuracy. For example, demand spikes during Christmas week or Valentine’s Day are treated as normal seasonal variation.

Smart Predict Workflow and Implementation Steps

The following are the implementation steps from building a Smart Predict model through to production application.

• Step 1, Prepare training data: Prepare a training dataset in the import data model that includes the “target variable, explanatory variables, time axis, and identifier.” Checking data quality (missing values, outliers, data types) is important. SAC automatically runs a data quality check and reports any issues.

• Step 2, Create a predictive scenario: Use “Create Predictive Scenario” to select the scenario type (classification, regression, or time series), and configure the target variable, training period, and inclusion/exclusion of explanatory variables.

• Step 3, Run training: The “Train” button executes training of the ML model. Training runs on SAP AI Core infrastructure. After training completes, a “model quality report” (accuracy metrics, variable importance, and the results of the training/validation data split) is generated automatically.

• Step 4, Evaluate and interpret the model: Evaluate accuracy (AUC, MAPE, etc.) in the model quality report. If accuracy is insufficient, try steps such as “addressing insufficient data volume, revisiting feature selection, or changing the training period.” Use variable contribution to confirm “which variables contribute most to the prediction” and ensure interpretability.

• Step 5, Apply the prediction: Apply the trained model to an “apply dataset (new data)” to generate predicted values. The generated predicted values are written back into SAC’s data model.

• Step 6, Automatic model retraining: A “retraining schedule” can be configured to periodically retrain the model automatically on new data. If model accuracy degrades due to a shift in the data distribution (concept drift), this can also be automatically detected and retrained.

Integrating Smart Predict with Planning: Prediction-Driven Planning

The most powerful use case for Smart Predict is its integration with xP&A. Predicted values generated by Smart Predict (future sales, future demand) are automatically written into the planning model as the “initial value (baseline) for the plan.”

• Integration flow: Smart Predict (ML forecasting) → generation of predicted values → data action (write-back into the planning model) → top-down adjustment by planners → approval workflow.

• Combination with driver-based planning: A representative example is designing a data action that predicts “churn probability by customer” with Smart Predict, and then uses that predicted churn rate to automatically calculate the plan value for ARR (annual recurring revenue).

Search to Insight / Just Ask: Technical Details of Natural Language Queries

“Search to Insight” (formerly known as Smart Search) is SAC’s natural language query (NLQ) interface feature. When a user types “show me this year’s sales by product” or “which region had the lowest profit margin last month?” in Japanese (or in one of 11 languages including English and German), SAC parses the query and automatically generates a chart or table.

How the NLQ Engine Works Technically

• Intent Recognition: Identifies the intent of the input text (the purpose of the analysis — display, comparison, ranking, trend, etc.). SAP uses a pretrained Transformer-based NLP model (BERT architecture) to classify intent.

• Named Entity Recognition: A NER model extracts business entities within the query (for example, “sales” → measure, “by product” → dimension, “this year” → time filter). The metadata of SAC’s data model (measure names, dimension names, attribute values) is used as the reference dictionary.

• Query generation: An internal query (MDX/SQL-like) is automatically generated from the recognized intent and entities and issued against the data model.

• Automatic chart selection: The AI automatically selects the optimal chart type based on the nature of the query (trend → line chart, comparison → bar chart, composition → pie chart, etc.).

Just Ask’s Extended Capabilities

“Just Ask” is a conversational AI feature that integrates Search to Insight into SAC’s home page and Digital Boardroom. It supports not only answering a single query but also multi-turn conversations that “dig further into the results of a previous question.”

• Follow-up queries: Continuous, conversational analysis is possible, such as “break the previous result down by quarter” or “which of these has worsened the most year-over-year?”

• Japanese-language support: SAC’s Just Ask supports queries in Japanese. However, recognition accuracy for Japanese measure and dimension names can be lower than for English. Recognition accuracy can be improved by defining the data model’s metadata (display names) in both Japanese and English.

Generative AI Integration: SAC Joule

As SAC’s latest AI capability, “SAP Joule” (Joule’s integration into SAC) has been rolled out. Joule opens as a side panel and provides the following generative AI features within SAC.

Automatic Story Generation (Joule for SAC Stories)

When you instruct Joule in natural language to “create a dashboard analyzing sales trends,” it selects the appropriate measures and dimensions from the connected data model and automatically generates a story that includes charts, filters, and insight text.

• Technical mechanism: User input → Joule (an LLM such as GPT-4o or Claude) calls SAC’s metadata API to obtain the model’s schema (available measures and dimensions) → the LLM designs the story structure → components are automatically generated via SAC’s API (Analytic API).

• Limitations and cautions: Stories generated by Joule are provided as a “first draft.” The LLM does not necessarily fully understand the business meaning of the data, so it is assumed that a human will revise the generated story to fit the business context.

Natural-Language Summarization of Insights (Narrative Insights)

Selecting “Explain This Chart” from the right-click menu on a chart or dashboard causes Joule to analyze the data and automatically generate a natural-language explanation.

• Example: Selecting a line chart of sales trends → Joule automatically generates a summary such as “Q1 2026 sales grew +12.3% year-over-year, driven especially by strong growth of +18.1% in the North America region. Meanwhile, the EMEA region is sluggish at -3.2%, continuing the downward trend that began in Q4.”

• Multilingual output: Supports generating summary text in Japanese, English, German, and other languages. A practical use case is automatically generating Japanese-language explanatory text for management meetings in Japan.

Answering Data Questions (Conversational Analytics)

When you ask Joule “what was the main reason sales fell short of the target last month?”, it references and analyzes SAC’s data to provide an answer. This can be thought of as running Smart Insights’ driver analysis conversationally via generative AI.

• Use of RAG (Retrieval-Augmented Generation): For Joule queries, SAC operates on a RAG pattern: “first issue a query against the data model to retrieve the figures → the LLM generates driver analysis and a natural-language answer based on that numeric data.” This does not depend on the LLM’s training data, so answers are always generated based on the user’s most current data.

Anomaly Detection: Details

SAC’s Anomaly Detection feature detects, in real time, “abnormal changes that deviate from the normal pattern” in time-series data, and raises alerts.

• Detection algorithms: A control-chart approach (statistical process control, SPC) based on a time-series exponentially weighted moving average, and multivariate anomaly detection using the Isolation Forest algorithm, are both used.

• Sensitivity setting: Alert sensitivity (how much deviation triggers an alert) can be set using a “Low / Medium / High” slider. Setting it to High reacts even to small fluctuations, while Low triggers alerts only for large deviations.

• Smart alert notifications: When an anomaly is detected, an alert such as “an abnormal decline of -22% versus the previous week in sales has been detected: click here for details” is automatically sent to SAC’s notification center, by email, or to Microsoft Teams (when integration is configured). Clicking the alert link automatically brings the relevant data point into focus within the SAC story.

• Setting the training period: The anomaly detection model learns the “normal range of variation” from data over “the past N periods (for example, 24 months).” If the training period is too short, seasonality is not learned sufficiently and false positives increase.

Constraints and Quality Management of AI Features

The following summarizes the constraints and quality-management points that should be understood when actually putting SAC’s AI features to use.

• Data volume requirements: As a general rule, Smart Predict prediction models require “at least 100 training records, and at least 24 time points for time-series forecasting.” When data volume is small, the model may fail to train, or overfitting may occur, reducing predictive accuracy.

• Dependence on data quality: The accuracy of AI analysis depends directly on data quality. The principle of “garbage in, garbage out” is especially pronounced with AI. Outliers, missing values, and inconsistencies in data definitions must always be cleansed before applying AI features.

• Black-box risk: GBM and random forest are complex nonlinear models. While feature importance and SHAP values provide explainability, there are cases where “an interpretation that humans can intuitively understand” is difficult to obtain. For models used in decision-making contexts such as planning, it is important to evaluate model interpretability not only in terms of accuracy but also from the perspective of whether “executives and staff can understand and trust it.”

• Managing model drift: Once built, a Smart Predict model’s accuracy degrades over time (concept drift). Regular accuracy monitoring (tracking MAPE and AUC trends) and configuring an automatic retraining schedule are essential operational requirements.

6. Digital Boardroom and Executive Dashboard Design

Structure of Digital Boardroom

Digital Boardroom is a specialized SAC view that combines “presentation and interactive management analysis” for board meetings and executive meetings.

• Boardroom layout: The standard is a hierarchical layout with “Overview (company-wide summary)” at the center, surrounded by functional area stories such as “Sales, Finance, Operations, and HR.” During a meeting, participants browse by drilling down through “company-wide overview → sales detail → by region → by representative.”

• Just Ask integration: Querying Just Ask directly from Digital Boardroom makes it possible to answer spontaneous questions in real time, such as “so how are sales in Southeast Asia?”

• Offline mode: In case the internet connection in the meeting room is unstable, Digital Boardroom content can be configured to be cached offline in advance.

Security and Data Access Control

SAC’s security is structured in three layers: “tenant administration,” “content management (folder permissions),” and “data access control (row level).”

• Tenant roles: SAC tenant administrator privileges are defined through roles that correspond to the service tier, such as Admin, BI Content Creator, and Planner Expert.

• Folder permissions: SAC content (stories, data models) is stored in folders, and “Read/Edit/Share” permissions are granted to users/teams on a per-folder basis.

• Data Access Control (row-level security): The model’s DAC settings restrict the rows (dimension values) a user can access. Even when viewing the same story, the range of data displayed differs depending on the user’s permissions. A design that manages DAC rules by mapping them to group attributes in IAS (SAP Identity Authentication Service) improves operability.

7. Case Studies from Other Companies

Case Study 1: Domestic Manufacturer, Company A — Improving Demand Forecast Accuracy with Smart Predict

Background and Challenges

At Company A, a manufacturer producing a wide variety of products in small volumes, monthly demand forecasting relied on staff members’ subjective judgment in Excel, and the forecast error (MAPE) averaged more than 25%. Both disposal costs from excess inventory and lost sales opportunities from stockouts were occurring, making improved inventory planning accuracy a management priority. Because the company carried more than 3,000 SKUs, it also lacked the staffing capacity to manually analyze forecast accuracy for every item.

SAC Smart Predict Adoption and Results

Company A imported data from SAP S/4HANA (inventory and sales actuals) into SAC and built time-series forecasting scenarios with Smart Predict for each item category. With SAC automatically selecting the algorithm (Holt-Winters for items with strong seasonality, ARIMA for non-seasonal items), the average MAPE across all items improved from 25% to 14%. The forecast values were fed into SAP IBP’s planning data via a data action and incorporated into the S&OP cycle. It has been reported that inventory levels were reduced through improved stockout rates and reduced excess inventory.

Case Study 2: Financial Services Company, Company B — A Churn Prediction Model Using Smart Discovery

Background and Challenges

At Company B, a financial services firm, persistently high contract churn rates were an obstacle to revenue planning. Reactive measures taken after a customer had already filed for cancellation (retention offers after the fact) came too late, creating a need for a “proactive customer retention” mechanism that could identify high-churn-risk customers in advance and act preemptively.

SAC Classification Scenario + Smart Discovery Adoption and Results

Company B built a churn prediction model using SAC’s classification scenario. The target variable was set to “whether the contract was cancelled at renewal (Yes/No),” and the explanatory variables were set to “contract tenure, usage frequency, service usage history, number of complaints, date of last login, NPS (customer satisfaction score), and others.” The GBM model achieved an AUC of 0.88, and after implementing proactive measures (dedicated representative outreach, promotional offers) targeting the top 20% of customers by predicted churn probability, the churn rate is reported to have improved. Analysis of the SHAP values revealed that “days elapsed since last login and number of complaints” were the primary drivers of churn, and the product team used this finding to inform UX improvement initiatives.

Case Study 3: Retailer, Company C — Democratizing Analysis for Frontline Staff with Just Ask

Background and Challenges

At Company C, a retailer operating stores nationwide, a process had developed in which store manager-level staff repeatedly requested Excel reports from headquarters’ BI staff in order to check sales and inventory status. Report-preparation effort for BI staff exceeded 200 hours per month, and there were challenges around the immediacy and autonomy of data utilization.

Just Ask (Natural Language Query) Adoption and Results

Company C rolled out SAC’s Just Ask feature to all store managers, establishing an environment where natural-language queries such as “tell me my store’s top 10 selling items this month” or “which product category had the highest disposal amount last week?” could be answered instantly. This was combined with a mobile-optimized Digital Boardroom so that even store managers without BI skills could operate Just Ask from the SAC app on their smartphones. It has been reported that the number of report requests to the headquarters BI team was significantly reduced.

8. SAC’s Technical Constraints and Operational Design

Performance Design Considerations

• Data volume limits for import models: Data volumes exceeding the tenant’s memory capacity limit cannot be imported. Large data volumes should either be referenced via a live connection through Datasphere, or compressed through aggregation (pre-aggregating) of the imported data.

• Number of concurrent users: Concentrated simultaneous access at large organizations (for example, reporting at the start of the month) affects SAC’s response performance. The standard SLA for SAP SAC is 99.9% availability, but managing query complexity during peak periods is also important.

• Query load from live connections: With a live connection, SAC’s queries are issued directly to the connected database (HANA, etc.). As the number of SAC users grows, the number of concurrent queries against the connected database also increases, affecting the load on that database. Query optimization (appropriate aggregation and index design) is required for models targeted by a live connection.

Quarterly Upgrades and Change Management

Because SAC is a SaaS offering, SAP automatically performs regular upgrades. A major release occurs once per quarter, with minor releases occurring monthly.

• Checking the preview period: Before a major release, changes to the UI, new features, and API changes should be verified in advance in a preview environment (test tenant). In particular, changes to the Scripting API can affect custom applications.

• Responding to Scripting API changes: SAC’s scripting language (SAC Script) can have APIs that become deprecated from one version to the next. Reviewing SAP’s “What’s New” notifications before an upgrade and assessing the scope of impact is an essential operational process.

— End —