Module 1 · Card 1 of 8

What text-to-SQL and AI cleaning close

The honest version. Not "AI helps you analyse faster." We'll show you which parts of the daily analyst loop the agent is actually closing — query writing, dedup, null handling — and which parts still require the analyst's judgment about what the data is supposed to mean.

Module 1 · Card 2 of 8

The pattern

An analysis has three layers — getting the data, framing the question, defending the answer. AI is very good at the first. It's confidently wrong about the second when the business context isn't in the prompt. It can't do the third — that's where you're accountable, not the model.

Module 1 · Card 3 of 8

What the evidence shows

On the Spider 2.0 text-to-SQL benchmark (Yale, 2026 Q1), Sonnet 4.6 with schema context resolves 73% of analytical queries correctly — up from 45% one year earlier. The remaining 27% are dominated by tables with overloaded column names, undocumented joins, and silently-deprecated columns the model can't see.

Module 1 · Card 4 of 8

What this means for you

Your competitive moat shifts from speed at writing the query to confidence in whether the query asks the right question. AI is closing the gap on rote SQL. What it can't do is know that `revenue_amount` in the `orders` table excludes returns booked in a different system three years ago — but you do.

Module 1 · Card 5 of 8

How to apply this tomorrow

Before you prompt Hex Magic or Claude on a request: write down the metric definition, the exclusions, and one historical gotcha you know about. The agent generates clean SQL from clean intent. Without those three, it produces queries that compile, run fast, and quietly answer the wrong question.

Module 1 · Card 6 of 8

Common failure mode

The trap: accepting an AI-generated SQL because the row count looks reasonable. A 2026 dbt Labs review of 1,400 LLM-generated queries found 22% returned the wrong-but-plausible result — wrong join cardinality, silent NULL filtering, off-by-one date windows. Stakeholders shipped decisions on them. Smooth output is not correct output.

Module 1 · Card 7 of 8

Cross-reference

This connects to Module 2 (Dashboards + metric definitions as moat) and Module 3 (Insight narrative in an AI era). The query is the cheap layer now. The metric you queried and the story you wrote about it — that's where the moat lives.

Module 1 · Card 8 of 8

End of module

Next: five questions. They take two minutes. Your AI Market Fit Score moves based on how you answer.

What text-to-SQL and AI cleaning actually change about your day

What text-to-SQL and AI cleaning close