Power BI AI Insights

Power BI AI Insights are the built-in machine learning features that scan datasets and surface patterns without manual configuration. For example, a sales manager runs Get quick insights on a dataset and sees a card explaining that Women's Accessories revenue dropped 23% in the Southwest - driven by two underperforming stores.

This is Microsoft's approach to augmented analytics: embedding classical ML models directly into the report canvas so that every business intelligence user, regardless of technical skill, gets explanations alongside their charts. Unlike Power BI Copilot, which uses large language models to generate reports and DAX, the classic AI Insights visuals rely on logistic regression, decision trees, and statistical decomposition - deterministic methods that produce consistent results on the same data. The Narrative visual is the exception: its Copilot mode uses a large language model.

What Power BI AI Insights Actually Do

Four features make up the AI Insights toolkit. Each solves a different analytical problem, and each uses a different ML technique under the hood.

Quick Insights scans an entire dataset automatically. You publish a dataset to the Power BI service (Quick Insights is not available in Desktop), select "Get quick insights," and the engine runs a battery of statistical tests - distribution analysis, trend detection, correlation checks, outlier identification - across the model's columns. The output is up to 32 insight cards, each an auto-generated chart with a short description of what the algorithms found noteworthy. Quick Insights works best as a starting point: a way to see what is interesting before you decide what to investigate.

Key Influencers answers the question "what drives this outcome?" You drop a metric into the visual - say, customer churn - and Key Influencers identifies which factors have the strongest statistical relationship. The output might read: "Customers who contact support more than 3 times are 4.2x more likely to churn." Behind the scenes, it runs logistic regression for categorical outcomes and linear regression for numeric ones, and keeps only the factors that pass a significance test.

Decomposition Trees let you drill into a metric along any dimension, with AI suggesting where to look next. If you are exploring revenue by region, an AI split might suggest drilling into product category next because one category holds the highest (or lowest) value of the metric.

The Narrative visual (formerly Smart Narratives) generates plain-language text summaries of charts and report pages. A bar chart comparing quarterly revenue gets an auto-generated paragraph: "Q3 revenue was $4.2M, a 12% decrease from Q2, driven primarily by the Electronics category." The text updates when filters change. The visual has two modes: Custom mode, which generates editable summaries without an LLM, and Copilot mode, which uses a large language model and requires Copilot.

How AI Insights Work Under the Hood

Understanding the statistical methods matters because it sets expectations. AI Insights are not magic - they are well-understood algorithms applied automatically.

Quick Insights runs a library of statistical tests in parallel. It checks for columns with unusual distributions, time-series with trend breaks, pairs of columns with high correlation, and categorical columns where one category dominates or has recently shifted. The top results become insight cards, up to 32 of them. The limitations: it runs only in the Power BI service, only on data imported into Power BI (not DirectQuery, streaming or push datasets), and it does not support row-level security, even in Import mode.

Key Influencers fits a logistic regression model when the target is categorical (e.g., churned vs. retained) and a linear regression when the target is numeric (e.g., satisfaction score). It tests significance with a Wald test (p < 0.05) on a sample of up to 10,000 data points and presents the factors that pass as "X increases the likelihood of Y by Z times." For segmentation, it uses a decision tree to identify groups with the highest concentration of the target outcome. This is classical supervised learning - no neural networks, no LLM. The visual does not support DirectQuery or a live connection to Azure Analysis Services or SQL Server Analysis Services.

Decomposition Trees offer AI splits at each drill-down level. When you choose High value or Low value, the visual considers every available field and picks the one that produces the highest or lowest value of the measure. In Absolute mode it looks for the highest or lowest value outright; in Relative mode it looks for values that stand out compared to the rest of the data in the column. AI splits are not supported on Azure Analysis Services, Power BI Report Server or Publish to web.

The Narrative visual works in two modes. Custom mode analyzes the visuals on a page, picks the most interesting things to summarize (up to four summaries per visual and 16 per page), and writes them as editable text with dynamic values. Copilot mode generates the summary with a large language model and requires a Copilot license, so its output is free-form rather than a fixed set of summaries.

The Narrative Visual and Auto-Generated Text

The Narrative visual deserves its own section because it is the feature most likely to be misunderstood. Users see generated text and assume an AI "understands" the data. In Custom mode, the engine reads the visuals' data, calculates summaries such as growth, highs and lows, and expected ranges, and writes them out as text. In Copilot mode, a large language model writes the summary from the report data.

A line chart showing monthly revenue generates something like: "Revenue peaked at $5.1M in November, then declined 8% in December. The overall trend for the period was upward, with an average monthly growth rate of 2.3%." If you change the filter to a different region, the text recalculates and regenerates.

In Custom mode you can edit and format the generated text and add dynamic values (references to measures or fields with a chosen summarization) that recalculate when data refreshes or filters change.

The limitation is scope. Custom mode summarizes common charts, scatter charts included, but it does not support tables, matrices, R and Python visuals, custom visuals, or map visuals with non-aggregated latitude or longitude. It also does not support a live connection to Azure Analysis Services or SQL Server Analysis Services. For complex reports, you will likely need to write your own narrative text and use the Narrative visual for the straightforward summary sections.

Anomaly Detection in Time Series

Anomaly detection works only on line charts with time series data on the axis. When you enable it, Power BI runs the SR-CNN algorithm over the series (Microsoft developed it for its now retired Anomaly Detector service) and draws an expected range of values. Points that fall outside this range are flagged as anomalies.

A sensitivity setting controls how selective the algorithm is. Low sensitivity flags fewer points - only larger deviations trigger. High sensitivity flags more points, so even a slight deviation is marked. The right setting depends on the use case: financial monitoring typically uses low sensitivity (flag only major issues), while quality control might use high sensitivity to catch small deviations early.

The most useful part is the explanations. When you select a flagged point, Power BI analyzes fields across the data model and returns a natural language explanation plus the associated factors, ranked by explanatory strength. You can restrict the analysis to chosen fields by adding them to Explain by. For example, a revenue spike on March 15 might trace back to the Electronics category, which was far above its expected value while other categories stayed within range. This root-cause analysis runs automatically, saving the analyst from manually filtering through every dimension.

Q&A Natural Language Interface and Its Retirement

Power BI's Q&A feature lets users type questions in plain English - "What were total sales last quarter by region?" - and get an auto-generated chart as the answer. It is important to distinguish this from Copilot. Q&A is not LLM-based. It uses keyword and phrase matching against the data model schema to interpret queries.

When a user types "top 10 customers by revenue in Q3," Q&A parses the query into components: "top 10" maps to a Top N filter, "customers" maps to a table or column named Customer, "revenue" maps to a measure or column, and "Q3" maps to a date filter. The mapping depends entirely on how the data model is structured - column names, table names, and synonyms defined in the linguistic schema.

This is why data model preparation determines Q&A accuracy. If your revenue column is named "amt_usd_v3," Q&A will not match it to "revenue" unless you add a synonym. If "Customer" could refer to three different tables, Q&A will guess - and often guess wrong. Best practices: use business-friendly column names, add descriptions, define synonyms for common terms, and test Q&A with the questions your users actually ask.

Q&A is free and available to all users, making it accessible to a broader audience than Copilot, which requires a paid Fabric F2+ or Power BI Premium P1+ capacity (a Pro or Premium Per User license alone is not enough). For organizations without that capacity, Q&A has been the primary natural language interface to Power BI data. That role is ending: Microsoft retires Q&A and Q&A Setup in February 2027 and points users to Copilot as the replacement. The model preparation described above carries over, because Copilot also depends on a clearly named, well-described semantic model.

Azure ML, AutoML and Cognitive Services in Dataflows

Power BI dataflows (now called Dataflow Gen1) used to offer three ways to bring custom machine learning into reports: invoking a deployed Azure Machine Learning model as a transformation step, training models with AutoML, and enriching data with pre-built Cognitive Services such as sentiment analysis and key phrase extraction. All three are now retired or deprecated in Power BI.

AutoML in Power BI dataflows is deprecated. Cognitive Services and Azure ML features in Power BI were retired in 2025: creating new models with them stopped on August 11, 2025, and refresh support for existing ones ended on September 15, 2025.

The replacement path runs through Microsoft Fabric. Microsoft points teams to AutoML in Fabric for supported AI scenarios and to AI services in Fabric, and Dataflow Gen2 can add AI prompt columns that call Fabric AI functions during data preparation. The practical pattern stays the same - a data science team trains a churn or forecasting model, scored results land as ordinary columns, and the Power BI semantic model reads them - but the training and scoring now happen in Fabric rather than inside a Power BI dataflow.

For most teams, the built-in AI Insights visuals described above remain the entry point, because they run inside the report without a separate machine learning pipeline.

Why Data Governance Makes AI Insights Trustworthy

Every AI Insights feature amplifies whatever it finds in the data - including inconsistencies, duplicates, and ambiguous definitions.

Key Influencers runs regression against columns in your data model. If "Revenue" means different things in different tables - gross revenue in one, net revenue after returns in another - the model may produce contradictory explanations. One analysis might say price increases drive revenue up, while another shows they drive revenue down, because the two "Revenue" columns measure different things. Without a data catalog documenting which column is canonical, users have no way to resolve the contradiction.

The Narrative visual generates confident text about whatever data is in the chart. If the underlying dataset has quality issues - stale data, missing rows, incorrect joins - the narrative wraps those errors in professional-sounding language. "Revenue grew 15% in March" sounds authoritative even when the March data is incomplete because an ETL job failed halfway through the load.

The pattern repeats across all four features. Quick Insights might flag a "trend" that is actually a data collection artifact. Decomposition Trees might suggest drilling into a dimension that only has data for half the time period. Anomaly detection might flag a point as anomalous when the anomaly is in the data pipeline, not the business.

The fix is data governance applied before AI Insights runs. A business glossary that defines what "Revenue," "Customer," and "Churn" mean. A data catalog that documents which datasets are governed and quality-checked. Data lineage that traces metrics from source system through transformations to the Power BI dataset. When these foundations are in place, AI Insights produces explanations that are grounded in defined, verified data - not guesses about ambiguous columns.

How Dawiso Supports Power BI AI Insights

Dawiso's data catalog and business glossary provide the metadata layer that AI Insights depends on but does not include natively.

When Key Influencers analyzes "customer churn," the accuracy of the result depends on whether "churn" is consistently defined across the dataset. Dawiso's business glossary documents the canonical definition - for example, "a customer who has not made a purchase in 90 days" - and the Power BI connector ingests the datasets, descriptions and measure expressions, so stewards can review them against glossary definitions.

Dawiso also shows which datasets are documented, owned, and approved through workflow. This gives BI teams a reliable starting point. Instead of running AI Insights against any dataset and hoping the results are meaningful, teams can start from datasets that have a named owner and an approved description.

Through the Model Context Protocol (MCP), AI agents can access Dawiso's catalog programmatically - searching objects, retrieving lineage, previewing objects and data samples, and looking up users and groups. This means AI Insights can be supplemented with automated metadata checks: before trusting a Key Influencers result, an agent can verify that the input columns have documented definitions and a named owner.

Data lineage in Dawiso connects Power BI reports, datasets, data sources and gateways. When the Narrative visual generates a statement like "Revenue grew 15% in Q3," lineage lets stakeholders trace that number back through report, dataset, data source and gateway, and further upstream where other connectors capture the source systems and pipelines.

Conclusion

Power BI AI Insights brings four practical AI features - Quick Insights, Key Influencers, Decomposition Trees, and the Narrative visual - directly into the report canvas. They work without code, need no Copilot capacity (apart from the Narrative visual's Copilot mode), and surface patterns that manual analysis would miss. Anomaly detection extends this with time-series monitoring, while natural language access is moving from Q&A to Copilot. But every one of these features produces results that are only as trustworthy as the data underneath. Organizations that invest in data governance - defining metrics, documenting datasets, tracking lineage - get reliable AI-powered insights. Those that skip governance get polished explanations of unreliable data.

Sources

• Microsoft Learn - Generate data insights on your semantic model automatically
• Microsoft Learn - Key influencers visualizations tutorial
• Microsoft Learn - Create and view decomposition tree visuals
• Microsoft Learn - Create smart narrative summaries (Narrative visual)
• Microsoft Learn - Anomaly detection tutorial
• Microsoft Learn - Copilot for Power BI overview
• Microsoft Learn, Introduction to Q&A in Power BI (Q&A is deprecated in February 2027)
• Microsoft Learn - Migrate from Dataflow Gen1 to Dataflow Gen2
• Microsoft 365 Message Center - MC1124567, retirement of Cognitive Services and Azure ML in Power BI
• Gartner - Gartner Predicts 75% of Analytics Content to Use GenAI for Enhanced Contextual Intelligence by 2027
• Dawiso Help - Power BI connector
• Dawiso Help - Dawiso Model Context Protocol (MCP)