AI Data Analytics: How Teams Use AI Across the Analytics Workflow

Analytics teams are being asked to support more business questions and more kinds of analysis, and AI is increasingly part of how they’re doing it: in a 2026 Snowflake-sponsored Omdia survey involving 2,050 business and technology leaders, 59% of data analytics teams reported using generative AI or agents.

What AI data analytics looks like in practice varies. It helps analysts generate queries, classify and summarize unstructured information, forecast future outcomes and identify patterns worth investigating. Natural-language interfaces give business users a more direct way to explore data, while agents can coordinate several analytical steps within the same investigation.

AI data analytics is the use of machine learning (ML), generative AI and agentic AI throughout the analytics lifecycle, including data preparation, pattern detection, prediction, natural-language querying, insight generation and decision support.

The scope has expanded considerably beyond the ML-supported predictive work long associated with advanced analytics. Language and multimodal models now give teams additional ways to query structured data and analyze documents, text, images, audio and other unstructured information.

For analytics practitioners, this broader scope is changing what the job asks of them. As AI takes on more query generation, forecasting, classification and interpretation, analysts spend less time producing every analytical artifact by hand and more time defining metric calculations, table relationships, model targets, business rules and evaluation criteria. Analysts must also review what AI produces and decide whether the result is fit for its intended use. As John Heisler, Snowflake’s Head of AI for Financial Services, puts it, “The opportunity is not to replace human judgment, but to better allocate it.”

Much of that work happens behind the interface. An AI assistant working from an uploaded extract can analyze the records it has, while enterprise analytics often depends on context maintained elsewhere: current source data, access rules, business definitions, lineage and relationships across data sets. As more of the analytical workflow incorporates AI, those underlying choices have a greater influence on the results users receive.

See how AI-powered BI and interactive analytics are changing the way teams explore data:

Much of the practical value of AI analytics comes from reducing repetitive work around an analysis while giving people more ways to work with the underlying data.

Faster access to routine analysis

Natural-language interfaces can shorten the path from a business question to an answer, especially for recurring requests such as comparing periods, filtering a metric by region or investigating a variance. Instead of routing every question through an analyst to write a query or build a report, users can explore those questions directly and continue with follow-ups as new findings emerge.

More analyst time for higher-value work

When analysts are freed from spending so much time on routine requests, they can dedicate attention to the work that depends heavily on their professional judgment: defining metrics and relationships, investigating ambiguous results, developing analytical models and validating whether an analysis supports the decision it’s meant to inform.

More types of data available for analysis

Gen AI expands the kinds of information that can enter an analytical workflow. Text, documents, images and other unstructured data can be classified, summarized or combined with structured records, giving teams a way to analyze information that previously required separate processing or substantial manual review.

More proactive analysis

AI can also change where an investigation starts. Anomaly detection, forecasting and automated insight generation can surface unusual changes or emerging patterns before someone explicitly asks about them. In cybersecurity, for example, AI can correlate identity, endpoint and cloud activity to flag an unusual activity sequence — behavior that might look unremarkable when each event is analyzed separately.

These benefits depend on the output being trustworthy enough to use. A plausible explanation built on a bad join, a forecast evaluated against the wrong target or an automated insight based on stale data simply produces errors faster. Accuracy, evaluation and clear business definitions are key to the value equation as teams automate more of the workflow.

AI can enter an analytics workflow at several points, from translating a business question into a query, to preparing data, surfacing patterns and communicating the result. Which capabilities teams use depends on the task and how the output will be used.

Natural-language and conversational analytics

Natural-language analytics lets users express analytical questions in ordinary business language, while the system translates those questions into queries or other analytical operations.

For a simple request, that translation is fairly narrow. “Show revenue by month” requires an approved revenue measure and an appropriate time dimension. Business questions get more contextual very quickly. For example, “How are enterprise renewals doing?” leaves several analytical choices unspecified. The system needs to know which accounts qualify as enterprise, what qualifies as a renewal, which measure represents performance and which period provides the intended comparison. An analyst historically resolved those ambiguities before constructing the query — a natural-language system needs the same business context available in a form it can use.

Conversational analytics carries an investigation through a series of questions. If a user asks, “Which regions saw the most change?” and then “What about only customers with multiyear contracts?” the system preserves the relevant metric, population and previous state. Governed semantic context does concrete operational work. Users gain a more flexible interface while the analytics team retains control over the definitions and relationships underneath it.

AI data preparation

AI can reduce some of the repetitive work involved in turning raw data into something analysts can use. For structured data, this includes suggesting transformations, mapping fields across sources, identifying missing or inconsistent values, or generating SQL or code to standardize data before analysis.

Gen AI gives teams the ability to prepare information that isn’t in the form of rows and columns. Teams can extract entities or attributes from documents, classify support tickets or other text, and convert parts of unstructured content into fields that can be analyzed alongside structured data.

The analyst still needs to verify what AI-assisted work produced since incorrect assumptions or omissions can propagate errors into the analysis that follows.

Automated insights

Automated insight systems scan data for changes, patterns or anomalies and flag findings without requiring someone to formulate a query each time. A product analyst monitoring hundreds of metrics, for example, might be alerted that trial-to-paid conversion dropped sharply for Android users after a new app release, giving the team a specific issue to investigate.

AI-generated insights give analysts a faster starting point, but analysts must still determine the cause. This is especially important when AI also generates a narrative about what it found. The system can identify where something changed and suggest possible explanations, but practitioners and business experts still need to determine which conclusions the evidence actually supports.

AI data visualization and AI-powered dashboards

AI data visualizations can shorten the path to useful charts and graphs. Instead of manually selecting fields, chart types and filters for every analysis, a user can ask an AI system to compare support volume by product and have the system generate an appropriate visual. The user can then refine the visual with a follow-up such as “Show me only enterprise customers” or “Break that out by issue category.”

The system must choose a chart or graph that represents the data accurately, use meaningful scales and aggregations, and avoid implying relationships the data doesn’t support, so human review is important.

AI-powered dashboards combine a persistent view of agreed-upon KPIs and trends with AI-assisted exploration. Users might ask a natural-language question about a change, generate a new breakdown or get a summary of notable movements without waiting for someone to build another dashboard view. The dashboard anchors the analysis in shared metrics, while AI gives users more ways to investigate what sits behind them.

Automated reporting

AI-generated reporting can combine calculated metrics with narrative summaries or explanations. A monthly report might pull revenue, margin and pipeline metrics from governed calculations, then use gen AI to draft commentary on the largest changes.

Because reports are often produced repeatedly for the same audience, AI can also reduce the work of rebuilding the narrative each reporting cycle. As the underlying metrics update, the system can identify what changed, generate a first-pass explanation and assemble the results into a recurring report for review and distribution.

AI analytics workloads don’t all interact with data in the same way. A natural-language query needs enough semantic context to translate a business question correctly, for example, while a forecasting model depends more heavily on historical data, feature definitions and evaluation criteria. Gen AI working with documents has a unique set of requirements around retrieval and grounding.

Natural-language analytics requires business context

When a user asks a question such as “Which customer accounts are showing signs of churn?” the system has to translate that request into an executable query. The system needs to know how the organization defines an active account, which behaviors count as churn signals, which tables contain the relevant data, how those tables relate and which filters or time windows apply.

A semantic layer can make those definitions and relationships explicit so natural-language requests reference the same governed analytical logic used elsewhere. Verified queries or other approved examples can provide additional guidance for common questions. Teams can then test whether generated SQL produces the intended result.

Predictive analytics depends on data, targets and evaluation

For predictive analytics, machine learning models learn statistical relationships from historical examples and use them to generate forecasts, classifications or scores. Reliability depends on choices such as which data represents the problem, how the target variable is defined and whether the evaluation metric reflects the consequences of getting a prediction wrong.

Governance is also important in predictive analytics, particularly for data access, lineage and reproducibility, but a semantic metric definition isn’t necessarily the central control for this kind of workload.

Generative AI needs evidence it can ground answers in

Gen AI often relies on the retrieval process to bring relevant source material into the model’s context before it produces an answer. The system first has to identify the right documents or passages, then give the model enough evidence to support the response without introducing claims that aren’t present in the source material.

This means the quality of the result depends on more than the model itself. Teams need to consider whether retrieval surfaced the most relevant evidence, whether important context was omitted and whether the generated answer accurately reflects the sources it was given. Teams can then evaluate both parts of the workflow: whether the system retrieved the right information and whether the model used it correctly.

Governance must follow the data into AI systems

Whatever analytical path is used, AI shouldn’t create a separate set of rules for the same enterprise data. Existing access policies, classifications, lineage and other governance controls still need to apply when data is queried through a conversational interface, used to train or score a model, or passed to an AI-powered application.

The exact controls differ by workload, but the principle is consistent: AI works best when it operates on the same governed data and business context that the organization already relies on for analytics.

AI increasingly participates in routine steps that analysts once handled manually, including drafting queries, suggesting transformations, detecting unusual patterns, recommending visualizations and generating narrative explanations. These AI-assisted workflows fall under the umbrella of augmented analytics — using AI and ML to support tasks across the analytics process.

For an analytics team, the operational question is how much human review each AI-assisted workflow needs. Data preparation needs checks on transformation logic and grain. Generated queries need semantic and SQL validation. Automated insights need statistical context. Generated explanations need a clear relationship to the evidence they summarize.

Augmented analytics reduces the amount of routine production work while keeping analytical judgment attached to the points where an incorrect assumption would materially change the result.

Agentic analytics is useful when an investigation involves several steps, tools, data types or follow-up decisions. An agent can use a model to interpret a request, select from permitted tools, carry information from one step to the next and adjust the workflow based on intermediate results.

A business user investigating a drop in customer renewals, for example, might ask what changed in a particular region. An agent could query renewal and usage data, compare affected accounts with prior periods, retrieve relevant support cases or call notes and then use those results to investigate likely drivers. The final response can bring together structured metrics and evidence from unstructured sources without requiring the user to formulate and execute each step separately.

The agent’s ability to maintain state changes how the investigation unfolds. If the first query shows that renewal declines are concentrated among customers with falling product usage, the agent can carry that finding into the next step rather than starting each request from scratch. It might then examine which features those customers stopped using or whether support activity changed during the same period.

Agentic workflows can also branch according to what the analysis uncovers. A fixed pipeline follows the same sequence every time; an agent might query one data set when it finds an adoption problem, retrieve account notes when it finds a relationship issue, or stop when the available evidence doesn’t support further investigation. That makes agents particularly useful for exploratory work where the next analytical step depends on the answer to the previous one.

Snowflake brings AI analytics closer to the governed enterprise data and business context those workflows depend on. Semantic Views define metrics, dimensions and relationships directly in Snowflake, while Snowflake Horizon Catalog provides governance capabilities such as access policies, classification and lineage. Cortex Agents can then use those semantic definitions to translate natural-language questions into SQL for structured data and combine that analysis with retrieval from unstructured sources or other tools.

Cortex AI Functions bring gen AI processing into SQL, and Snowflake ML supports predictive modeling. Snowflake CoWork provides an interface for multi-step analysis using Cortex Agents and connected tools. The common foundation means teams can apply AI to data without recreating business definitions and governance separately for each analytical experience.

AI is giving analytics teams more ways to make insights accessible, faster. A business user can ask a question without writing SQL, an analyst can start from AI-prepared data or automatically surfaced anomalies, and an agent can continue an investigation across several tools and data sources.

The practical question is how much of each workflow to delegate. Some steps can run with relatively little intervention, while others need defined metrics, constrained tool choices, or human review before the result should influence a decision. As AI takes on more analytical work, designing those boundaries becomes part of designing the analysis itself.