Data & Insights

"How many active users last month per region?" → the model writes the SQL, runs it, returns the table + a chart.

What makes it work in practice:

• Schema description in the prompt — column names, types, foreign keys, sample rows.
• Few-shot examples of question/SQL pairs.
• Allow only SELECT — never let the model issue DDL/DML.
• Show the SQL to the user; let them verify before trusting the chart.

Tools: Snowflake Cortex Analyst, Databricks Genie, BigQuery Gemini, Redshift Q, Hex Magic, Vanna.ai, Defog.

BI tools are baking AI in:

Hybrid is best:

• Classical ML for the detection (Prophet, isolation forest, ESD, statistical control charts) — cheap, fast, explainable.
• LLM for the explanation ("Revenue dropped 18% — likely related to the deploy at 14:32 that affected the checkout API").

Tools: Anomalo, Sifflet, Monte Carlo, Datadog Watchdog, AWS Lookout for Metrics.

Daily / weekly emails: "Here's what changed in the business this week, with charts." Generated from KPI deltas + recent context. Less impressive than text-to-SQL but quietly more used.

Foundation time-series models (TimeGPT, Chronos, Lag-Llama, Moirai) are starting to compete with classical Prophet/ARIMA — zero-shot forecasts with no training. Combine with an LLM for "what if Q4 marketing spend doubles?" scenario narration.

RAG over PDFs, contracts, financial filings, support tickets. Often packaged as "data extraction" or "document AI." Tools: Azure Document Intelligence, AWS Textract + Bedrock, Google Document AI, Unstructured, LlamaParse.

The biggest determinant of AI analytics quality is whether your data has clear, agreed-upon definitions. Cube.dev, dbt Semantic Layer, MetricFlow, Looker LookML formalize "what is an active user" once — then the LLM can use the metric instead of inventing SQL.

• Display the SQL the model generated.
• Link to the source table / dashboard / metric definition.
• Let users edit the SQL and re-run.
• Surface confidence ("I used the orders table; if you meant gross orders, the answer differs").

The model's database role should be read-only. Allow-list specific stored procedures if writes are needed. Never grant DROP, DELETE, or schema-mutation rights to a generative system.

• Apply row-level / column-level security before the LLM sees data.
• Mask emails, phone numbers, IDs unless the user has explicit access.
• Log every NL question + generated SQL for audit.

• "Ask anything" on a 4,000-table warehouse with no semantic layer — guarantees garbage answers.
• Auto-emailing AI summaries to executives without human review.
• LLM-only anomaly detection — too noisy, too expensive.
• Treating AI charts as canonical — they should drive humans to the dashboard, not replace it.