AI in data analytics: where it earns its keep, where it does not

The thesis

AI in data analytics is in the same place AI in coding was about two years ago: real productivity gains for the narrow set of tasks where the AI’s training and the user’s intent line up well, and a long tail of demos that do not survive contact with a real environment. The narrow use cases (natural-language query over a governed semantic model, anomaly detection, narrative generation, schema-aware suggestion inside development environments) are mature enough to deploy. The broader claims about autonomous analytics, AI replacing analysts, and AI removing the need for a data team are not. We do not believe they will be inside the next twelve months. The organizations getting value from AI inside analytics are the ones with a strong data foundation and clear governance. The organizations who are not are the ones who tried to use AI to skip the foundation.

What is actually working in production

Several AI features inside analytics platforms are in production at our clients in 2026.

• Natural-language query over governed semantic models. Copilot inside Power BI, the AI assistant inside Tableau, and Qlik’s AI features all support natural-language query over a published semantic model. The quality depends almost entirely on the semantic model. A well-defined model produces useful answers; an unmaintained model produces confident wrong answers.
• Anomaly detection. Time-series anomaly detection is mature. It works well for operational metrics with consistent patterns, less well for sparse or highly seasonal data without tuning.
• Narrative generation. AI-generated executive summaries of dashboards are useful for time-pressed executives who would otherwise skim the visualization. The narrative quality depends on the underlying data being correct; the AI does not catch errors in the source.
• Developer productivity inside the data engineering layer. Schema-aware SQL suggestion, dbt model authoring assistance, and pipeline troubleshooting are real productivity gains for data engineers.

What is overpromised in 2026

Several claims are not yet supported by what we see in production.

• Autonomous analytics. The marketing claim that AI will explore data, identify insights, and recommend actions without analyst involvement is not yet true. The systems that come close are narrow and require careful semantic model curation. The general case is not solved.
• Replacing the analyst. The analysts at our clients are not being replaced; the productive ones are using AI assistants to do more, faster. The non-productive use of AI (asking it to author a dashboard the analyst has not scoped) produces dashboards nobody trusts.
• Skipping the data foundation. AI features that operate over a poor data foundation produce confidently wrong outputs. The organizations getting value from AI are the ones whose data foundation was already good.
• AI as a strategy. “AI strategy” as a standalone document is usually an avoidance behavior. AI workloads are workloads. They belong inside the data strategy, the architecture, and the operating model.

The semantic model is doing most of the work

The pattern across every AI use case that works is the same: a well-maintained semantic model. The natural-language query depends on it. The anomaly detection respects the metric definitions encoded in it. The narrative generation describes the metrics correctly because the metrics are defined unambiguously. The semantic model is the interface between the AI and the data. Organizations that have invested in a clear semantic layer (typically dbt models with documented metrics, Power BI semantic models, Looker’s LookML, or the equivalent in Qlik or Tableau) get useful AI behavior. Organizations that have not get the marketing demos. We treat the semantic model as the prerequisite for any AI feature inside analytics. There is no shortcut.

Governance is the differentiator, not the platform

The risk of AI inside analytics is not that the model is wrong; the risk is that the model is confidently wrong over data the user does not have authority to see, or over metrics that do not mean what the user assumes. The platforms have largely addressed the access control side: Power BI Copilot respects row-level security, Snowflake Cortex respects masking and policy, the analytics platforms respect their published semantic models. The governance work that remains is the human side: making sure the published semantic model is correct, the metric definitions are current, the data quality monitoring is in place, and the AI features are configured against governed data, not raw tables. Organizations that have invested in a modern catalog and a real governance practice are positioned to use AI well. Organizations that have not are not. The differentiator is governance, not platform.

What the next twelve months will and will not change

The realistic forecast for the next twelve months: AI inside analytics will continue to improve at the same kinds of tasks it already does well. Natural-language query will become more reliable as semantic model coverage improves. Narrative generation will become more useful as platforms learn to cite their sources better. Developer productivity inside data engineering will continue to compound. What is less likely to change inside twelve months: the autonomous analytics promise, the replacement-of-analyst promise, and the shortcut-around-the-data-foundation promise. We have been watching these claims for several release cycles. They have moved more slowly than the marketing suggests. The mistake is to plan around the marketing forecast rather than the operational forecast.

Concrete takeaways for the AI conversation

• Invest in the semantic model first. AI features get their answers from the semantic model. Make the semantic model worth using.
• Deploy the narrow use cases now. Natural-language query, anomaly detection, narrative generation, and developer productivity are deployable today.
• Resist the autonomous analytics framing. The technology is not there. Plan around the operational forecast.
• Govern the AI footprint. Row-level security, masking, and policy belong to governance, not to the AI feature. Verify configuration.
• Treat AI as a workload class. Put it inside the data strategy and the architecture. Do not write a parallel AI strategy.