A Complete Understanding of Power BI and AI Predictive Analytics

Natural Language Query (Q&A)
The "Q&A" feature in Power BI allows users to input questions in natural language and instantly display visuals or numerical values that match the content. For example, if you input questions like "What are this month's sales?" or "What are the top 5 products by region?", the corresponding charts or cards are automatically generated.

Q&A uses advanced natural language processing (NLP) and also provides autocomplete and suggestions for example questions while you are typing. Technically, it matches field names and values in the data model with the user's words and returns answers interactively through in-memory processing.

As an additional setting, if you register synonyms for data assets, you can improve recognition accuracy even when asking questions using general terms instead of technical jargon.

Key Influencers Visual
"Key Influencers" is one of the AI visuals provided by Power BI that automatically analyzes factors influencing a certain result from data and visualizes them in order of importance.

For example, if you analyze "What are the factors for low customer satisfaction?", explanatory variables with high impact, such as contract length or product category, are listed. Behind the scenes, Microsoft's machine learning framework ML.NET is used, and as a predictive model, it executes logistic regression (when the analysis target is a category) or linear regression (when the analysis target is a numerical value) to calculate the strength of the factors.

Also, a decision tree algorithm is used to find characteristic groups (segments) within the data, which is utilized in the "Top segments" tab to automatically extract data groups with common characteristics.

Users can obtain impact rankings and segment analysis results based on statistical models just by placing this visual and specifying the "metric to analyze" and "possible factor items" without complex settings.

Decomposition Tree
The decomposition tree is a visual element that analyzes a numerical metric while breaking it down by multiple categories. What is characteristic is that it is equipped with an automatic breakdown feature called "AI split."

Normally, the user selects the item for "what to break down in detail" on the tree, but with AI split, Power BI proposes the "next perspective that has the most impact on the metric" using internal algorithms.

Specifically, it considers all candidate dimensions that can further break down the current node and automatically selects the split that maximizes or minimizes the metric value being analyzed. This automatic selection is displayed with an icon (💡 mark), and the user can drill down into the tree just by clicking it.

For example, when breaking down sales to analyze causes, if you select the "High Value" AI split, Power BI finds the combination of factors that maximizes sales from among region, product category, period, etc., and expands it in the tree. This allows you to discover noteworthy parts of the data without missing them.

Anomaly Detection
This is a feature that automatically detects and marks outliers (data points that deviate significantly from predictions) on a line chart of time-series data. When you enable "Find anomalies" from the line chart analysis menu, the expected range is calculated from the past data of the series, and anomaly markers are attached to points that fall outside that range.

Furthermore, each anomaly has an explanation feature that estimates "why this outlier occurred," investigating correlations with other related data items to infer the cause of the outlier.

Technically, this anomaly detection algorithm uses methods derived from the Azure Cognitive Services Anomaly Detector API (such as the SR-CNN algorithm), achieving advanced anomaly detection and cause analysis with few steps.

For example, when anomaly detection is applied to monthly sales trends, it detects and marks months where sales suddenly plummeted, and automatically checks whether other metrics (such as advertising costs or inventory shortages) showed different values than usual only for that month, making it easy to grasp abnormal events and pursue causes that are often overlooked by humans.

Time Series Forecasting
This is a feature that extrapolates future values on a line chart. The line chart in Power BI Desktop has a "Forecast" section in the analytics pane; when turned on, it draws a forecast line for a specified future period.

The default forecasting model in Power BI is exponential smoothing, which automatically detects periodicity in the data and applies the optimal model. Specifically, it internally switches between the ETS AAA (Holt-Winters method) if it determines there is seasonality, and the ETS ANN model if there is no seasonality.

This allows users to obtain forecast values that account for certain trends and seasonal patterns without needing to be aware of the algorithms. For example, when applying a forecast to monthly sales data, the predicted sales for several months ahead are drawn based on past growth trends and seasonal fluctuations, allowing you to visually confirm future sales prospects.

Note that the Power BI forecasting feature is available for line charts with a single series (it does not work with multiple series) and requires the number of data points to be 1000 or less (as sampling occurs with large datasets, reducing accuracy). While basic forecast periods and confidence intervals can be adjusted via the GUI, if you wish to perform more complex forecasting (such as ARIMA models or custom regression forecasting), it is possible to implement them using DAX formulas or R/Python scripts, as described later.

Clustering
While not a feature with an explicit name, cluster analysis can be utilized through several Power BI features. Representative examples include the aforementioned "Top Segments" in Key Influencers or the feature to automatically cluster data points in a scatter plot.

In a Power BI scatter plot, when you select data points and perform "Find Clusters," it internally uses k-means clustering to automatically group similar data points together, and you can obtain the results as a new cluster category column. This makes it easy to perform analyses such as grouping customer data by purchasing patterns or segmenting products by sales and profit characteristics.

Furthermore, clustering results can be used for visualization and filtering, which is useful for understanding characteristics by segment. For example, if you cluster using a scatter plot of sales and profit margins, you can automatically classify items into groups such as "high-sales, high-profit premium products" or "high-sales but low-profit problematic products," which can then be applied to strategy planning.

AI Insights / Text Analytics & Vision
Within the Power Query Editor, which performs data loading and transformation in Power BI, an "AI Insights" feature utilizing Azure's AI services is provided.

Specifically, you can apply text analytics such as "Sentiment Analysis," "Key Phrase Extraction," and "Language Detection," as well as image analysis such as "Tag Images," which are pre-trained models, to columns of data.

For example, if you apply "Sentiment Analysis" to a column of review text, you can add a score indicating whether the sentence is positive or negative as a new column. These are AI models powered by Azure Cognitive Services, but in a Power BI Premium environment, they are available without additional Azure contracts, and the computational processing is executed as an AI workload on the Power BI service.

In marketing and customer support fields that handle large amounts of text data, applications such as automatically visualizing sentiment trends from surveys and reviews, or extracting key topics from text to help with product improvements, can be considered.

The fact that NLP features with high demand in business settings can be used no-code is a major advantage of Power BI's AI features (*these are features that require a Premium license, and there are prerequisites as described later).

Machine Learning Models via AutoML
Power BI has also provided Automated Machine Learning (AutoML) features for advanced users. In Dataflows on the Power BI service,



it was possible to create and train ML models such as binary classification or regression using GUI operations, and perform predictions on new data. For example, it was possible to create a model that predicts "whether or not a purchase will occur" from website visit data and incorporate the scoring results into a report. AutoML is designed to be handled by non-data scientists, and after model training, an evaluation report is automatically generated, showing accuracy and important features. However, this AutoML feature within Power BI (for Dataflow v1) was discontinued in 2023 and has now been integrated into the successor AutoML feature within Microsoft Fabric. Fabric is Microsoft's integrated data analytics platform, and Power BI is positioned as a part of it.

Therefore, moving forward, when building predictive models in Power BI, the shift is toward using the Fabric environment (e.g., Fabric's Data Science notebooks or AutoML features) and importing the results into Power BI reports.

R/Python Integration
In addition to the above, it is also possible to perform custom analysis in Power BI Desktop using R language or Python scripts. You can create machine learning models in R or Python and incorporate the results (such as predicted values or cluster IDs) into your dataset, or display visuals generated by R/Python in your reports.

For example, you can flexibly utilize these by training and predicting with a scikit-learn model using a Python script, and then importing the prediction results into Power BI to create a dashboard.

This allows for the use of advanced algorithms (such as deep learning) and the latest open-source libraries not found in Power BI's standard features, but writing and maintaining scripts requires specialized knowledge.

In general business operations, you can start by handling tasks with standard AI features and incorporate custom analysis via R/Python as needed, allowing for both no-code and code-based analysis.

Integration with Azure Machine Learning
　Azure ML is a platform for full-scale machine learning model development and operation. Power BI can call models built and published in Azure ML as real-time inference services and import the results into reports.　As mentioned above, in addition to using Azure ML models as functions via the AI Insights feature in Power Query, it is also possible in some cases to hit Azure ML endpoints via REST API to obtain results. As for

points of differentiation, if you have already invested in Azure ML and are operating advanced models, you can share the results with end users by connecting them to Power BI.　On the other hand, if you want to complete everything within Power BI or if there are no data scientists in the company, it is better to use AutoML or standard AI features described later. The advantage of using Azure ML is that you can

build models with high flexibility using Python/R, etc., and that you can manage models scalably on Azure.　For example, if you build, train, and deploy a demand forecasting model in Azure ML and visualize the prediction results in Power BI, you can perform advanced operations such as always referencing the latest model results retrained on the Azure side when data is updated.

Azure Cognitive Services (Text Analysis/Image Recognition)
　Azure Cognitive Services is a group of API-based AI features provided by Microsoft. The "Text Analytics" and "Image Tagging" features incorporated into the aforementioned Power Query are made available on Power BI Premium as pre-built models of these Cognitive Services.　Therefore, if a user without Power BI Premium tries to do the same thing, they need to contract for Azure Cognitive Services resources themselves and create a custom script to call the REST API from Power BI.　The AI Insights feature of Power BI saves that trouble, but as for



differentiation, the division of roles is that if "the analysis scale is large and you want to perform dedicated tuning on the Azure side" or "you do not have Premium and want to use it via API billing," you should use Azure Cognitive Services directly, and if "you want to easily analyze text or images without code," you should use AI Insights in a Power BI Premium environment.　Since Cognitive Services itself covers a wide range of areas beyond sentiment analysis, such as speech, translation, and custom AI, if you want to incorporate AI utilization that is more flexible than Power BI reports into applications, you can also consider an approach of building more complex workflows on the Azure side and visualizing/sharing the results separately in Power BI.

Power BI AutoML and Fabric
　As mentioned earlier, the AutoML feature within the Power BI service has been integrated into Fabric. Fabric is an integrated environment that spans data engineering, science, warehouse, and BI, including Power BI.　In the

Data Science experience within Fabric, model development is possible using no-code AutoML or Python notebooks, and the results can be linked to Power BI reports via pipelines.　This is more powerful than the conventional "AutoML within Power BI dataflows" and makes it easier to combine with larger-scale data and advanced preprocessing. As for

differentiation, if the company has introduced a Fabric environment, it is more appropriate to create and manage predictive models there and have Power BI reports focus on displaying the results.　On the other hand, if Fabric has not been introduced and only Power BI Premium is being used, AutoML is not available, so simple predictions are supplemented by Key Influencers or Forecast features, and if necessary, replaced by Azure ML or R/Python. Since Microsoft recommends predictive analysis on Fabric in the future, it is important to develop the environment in accordance with the latest roadmap.

Comparison with other tools
When compared with other clouds (such as GCP's Looker Studio + BigQuery/Vertex AI, or AWS's QuickSight + SageMaker), Power BI excels in GUI-based visualization and dashboarding, and has the advantage of being able to handle everything from big data analysis to BI in an integrated manner when combined with Microsoft's data infrastructure (such as Azure Synapse Analytics).

However, in terms of pure machine learning model development, it is still more flexible to use specialized Python/R environments or Azure ML. In other words, it is important to separate the tasks of "creating predictive models" and "applying model results to business," and to focus Power BI on the latter (the BI portion).

Power BI's AI features are suitable for end users and analysts to quickly gain insights, while Azure ML and similar tools are suitable for data scientists to build precise models. By linking them, it becomes possible to directly apply advanced predictive models built by experts to decision-making in the field.

Marketing budget optimization
In the marketing department, optimizing budget allocation is crucial to maximize the effectiveness of advertisements and campaigns. With Power BI, you can visualize Return on Ad Spend (ROAS) based on past performance data by advertising channel, and by performing predictive analytics, you can quantitatively judge "which channels should have their budgets increased or decreased."

For example, one company analyzed past campaign data using key influencers to identify the factors contributing to conversions. By doing so, they predict the conversion rate of prospective customers and adopt a strategy of prioritizing budget allocation to segments expected to yield high results.

Also, by using regression analysis models to predict the impact of ad clicks and site traffic on final sales, and by drawing a predictive curve of sales per budget, attempts are being made to find the point where ROI is maximized. In the marketing domain, customer behavior prediction and campaign response prediction enable data-driven marketing that increases effectiveness with limited resources.

Improving supply and demand planning through sales forecasting
In retail and manufacturing, forecasting sales of products or services is the key to inventory, production planning, and budget formulation. By building predictive models for monthly sales or demand using Power BI's time-series forecasting features or AutoML, it is possible to respond proactively to future demand fluctuations.

For example, one retail chain created a sales forecast dashboard that considers seasonal factors based on store-by-store sales data, allowing them to grasp periods of surging demand in advance and perform additional ordering or increase staff. As a result, there are cases where they reduced stock-out losses and prevented lost sales opportunities. In the financial industry, it is also used to predict trends in stock prices and interest rates from past market data to inform investment decisions.

While Power BI's predictions are merely statistical estimates, they support decision-making by visually showing future scenarios to management. To increase prediction accuracy, there are cases where advanced predictive models (such as Prophet or deep learning models) are built in Azure ML as needed, and the results are imported into Power BI for visualization.

Customer segmentation and targeting
Predicting future customer behavior and taking appropriate approaches is also an important use case. For example, in subscription businesses, it is possible to perform "churn prediction" and proactively implement campaigns for customer segments with a high risk of cancellation.

In Power BI, you can use key influencers and clustering to extract customer segments and consider countermeasures according to their respective attributes and values. In the marketing example, one could find groups of customers with similar tastes through clustering and recommend products that resonate with those groups.

In fact, one e-commerce company successfully implemented segment-specific strategies by using Power BI to perform RFM (Recency, Frequency, Monetary) analysis to rank customers, proposing loyalty programs to top-tier customers, and distributing coupons to dormant customers to encourage revisits.

Furthermore, data-driven sales activities are also being conducted, such as extracting "common characteristics of high-value customers" using the key influencer's top segment feature and focusing follow-ups on new customers who meet those conditions. While these previously required specialized data mining, Power BI now allows marketers and sales representatives in the field to analyze and act on their own.

Optimization of Staffing (Application to HR and Operations)
Predictive analytics is also useful for planning employee and staff allocation. For example, in a call center, you can forecast the number of inquiries for each time slot and schedule the appropriate number of operators to reduce wait times and improve labor cost efficiency.

By analyzing historical call volume data in Power BI and modeling patterns by day of the week, time, and seasonal factors, you can perform 15-minute interval call forecasting to determine the appropriate number of staff. In fact, Microsoft Dynamics 365 Customer Service provides a solution that predicts the required number of service representatives in 15-minute increments based on short-term demand forecasts, enabling shift planning without shortages or excesses.

Predictive analytics is also used in retail stores to make decisions such as increasing or decreasing cashiers based on foot traffic forecasts or securing temporary staff before busy seasons. Furthermore, in the HR field, there are use cases such as predicting employee turnover to take early action, or predicting the results of talent development to optimize training programs.

These methods capture and anticipate "human movement" through data, improving accuracy compared to plans that rely solely on intuition or experience, and ultimately leading to operational efficiency and improved service levels.

Model Accuracy and Limitations
Predictive analytics is a probabilistic estimate and does not perfectly predict the future. Power BI's built-in forecasting (such as exponential smoothing) is a relatively simple model and cannot handle sudden demand fluctuations, abnormal situations, or structural changes. Furthermore, machine learning models struggle to predict events outside the scope of their training data.

For example, irregular events like COVID-19 cannot be incorporated into existing models. Therefore, it is necessary to adopt a stance of not taking prediction results at face value and combining them with business insights to make decisions.

Even within Power BI, it is important to display not only the predicted values but also confidence intervals and the assumptions used for the prediction (such as seasonal cycles) to convey uncertainty to users. Model accuracy should be verified periodically, and you should consider refreshing models or adding explanatory variables as needed.

Interpretability and Accountability
When using AI in BI tools, the ease of explaining results also becomes a challenge. The Key Influencers visual presents results in a relatively easy-to-understand way, but it can still be difficult to provide a convincing explanation when users ask, "Why is this factor important?" Visualization also becomes difficult when multiple factors are intertwined.

Additionally, since segments from automatic clustering are divided based on statistical optimal criteria, there are cases where changing the segmentation method would be more useful for business purposes. It is essential to have a process that combines field knowledge with AI results to provide interpretations in human-understandable language, such as "This segment consists of young customers and has the following purchasing trends."

Power BI also has an automatic commentary generation feature called Smart Narratives, but human checks are still required for use in important decisions.

Scope of Application for Each Feature
Power BI's AI features have limitations on where they can be applied. For example, forecast lines can only be applied to a single line chart series and cannot display forecast comparisons for multiple categories at once. Anomaly detection also cannot be used for data other than time-series data (e.g., outliers in a histogram).

Key Influencers may also fail to capture trends if the amount of data is extremely small or if all explanatory variables are numerical and have non-linear relationships. Furthermore, models that can be built with AutoML are limited to basic classification, regression, and time-series, and cannot handle models like image recognition or text generation.

For analysis methods not covered, you need to supplement with external tools or custom scripts. It is important to understand "what can and cannot be done." While there is a possibility of improvement in future updates, consider the optimal way to use the current feature specifications.

Performance and Scalability
Because AI features involve advanced calculations, increasing data volume affects processing time and report display speed. For example, Key Influencers perform multiple model calculations in the background, so it may take time if the data is in the millions of rows (Power BI samples as appropriate, but this is a trade-off with accuracy). R/Python visuals can also take several tens of seconds to execute, which may compromise dashboard responsiveness.

Countermeasures include using pre-aggregated data for the necessary scope, optimizing the data model into a star-schema for speed, or increasing processing power with Premium capacity. Also, for overly heavy analysis, it is worth considering architectural improvements such as performing batch processing on the Azure side (Synapse Analytics or AI services) instead of Power BI, and only referencing the results in Power BI. In addition, since Power BI reports usually contain multiple visuals, be mindful of load balancing with other visuals so that one AI visual does not become a bottleneck for the entire report.

Data Governance and Privacy
When using personal or confidential data for AI predictive analytics, caution is required in how it is handled. When using Cognitive Services in Power BI, data is analyzed in the cloud (in the case of Premium features, it is executed within the customer's Power BI capacity, but even then, for example, sentiment analysis of text itself can involve employee privacy).

Therefore, you need to check internal policies regarding which data can be used for AI analysis. In some cases, it is necessary to limit analysis to anonymized or aggregated levels and avoid making predictions that could identify individuals. Also, if the prediction results themselves have a significant impact on decision-making, there is a responsibility to explain the basis for them (e.g., credit scoring results).

While Power BI's analysis process is reproducible, you cannot fulfill your accountability if you cannot track the details of the prediction model. In such cases, it is preferable to manage the model in Azure ML to ensure explainability with model cards, and only present the results in Power BI.

Integration with Business Processes
The results of predictive analytics are not an end in themselves, but insights for taking action just displaying predictions in a Power BI report is not enough; value is only realized when they are incorporated into subsequent business processes. For example, in a "demand forecasting to order quantity determination" process, if the forecast is exported to Excel and then manually entered into an ordering system by staff, the benefits of automation are completely lost.

Since Power BI is limited to analysis and visualization, Power Automate and other Microsoft Power Platform products should also be considered to trigger alerts or automate workflows based on predictions.

Furthermore, predictive models need to be reviewed periodically in response to changes in the environment and strategy. A model that worked well at the time of implementation may become obsolete after a few years.PDCA cycle It is important to establish a system for continuous improvement and to update the predictive analytics embedded in business processes.

Identifying Areas of Application
It is also important to keep in mind that not all business tasks are suitable for predictive analytics. For example, in areas where human sensibility and randomness have a strong influence, such as trends or hit products, it is difficult to read the future from past data. In such cases, forcing the use of a predictive model may actually lead to misleading results.

Some companies that have introduced Power BI have noted that "predictive analytics had limited effectiveness for products purchased impulsively." Conversely, predictive analytics is highly effective for businesses with planned purchases or regular demand. Determining whether AI predictive analytics is truly useful for your business area and evaluating the ROI before implementation is also key to success.