Nine Unusual Ways My Clients Use AI with SQL Server

Pinal Dave presents nine real consulting engagements where AI was used not to generate SQL queries, but to read, classify, and extract from existing database artifacts at volume — with human verification as the non-negotiable final step. Every case follows the same shape: the machine reads at scale and proposes; a human verifies and decides. The common thread is that none of these are hard problems — they are large, tedious, low-judgment reading tasks that a competent person could do given three months and no interruptions, a resource that has never existed.

The nine workflows:

1. Business rule extraction from legacy PL/SQL — 800 Oracle packages, 400K lines, 19 years old. AI extracted an inventory of plain-English business rules with line-number citations, surfacing ~60 obsolete or wrong rules the business didn't know about. Key detail: feeding table DDL alongside code improved reasoning quality more than prompt wording.

2. Reverse engineering a vendor's undocumented schema — Five inputs assembled per table (DDL, FK graph, value distributions, vendor's own readable views, Extended Events capture of one screen) produced a 60-page data dictionary for a database the vendor contractually refused to document.

3. Decoding a German ERP from the early 2000s — Two-pass approach: expand abbreviated German column names, then derive domain meaning from context. Compound-noun ambiguity was the failure mode; test data caught what confidence couldn't.

4. M&A due diligence schema mapping — Column-by-column mapping across two company schemas with three required outputs: target column, confidence level, and specific reason. The reason column was what let humans triage 900 columns in a week and catch that CUSTOMER_STATUS meant Active in one company and Archived in the other.

5. Compliance documents turned into SQL checks — A 300-page audit document became a test suite of queries that return zero rows when compliant. Section-number traceability in query comments made it credible to auditors. About a third of requirements weren't testable in SQL at all.

6. Column-level data lineage across 40 procedures — Traced a finance number through views, stored procedures, and SSIS packages. AI read inside dynamic SQL (which sys.sql_expression_dependencies can't see) and found the number was being rounded twice, causing a one-cent reconciliation argument every month for years.

7. SQL Agent job graveyard audit — 240 accumulated jobs across 15 years. AI classified each into four buckets (needed, dead, unclear, broken-but-silent). Found jobs writing to dropped tables, copying to decommissioned shares, and one emailing a report to ex-employees' still-resolving mailboxes. Rule: disable with a note, delete after 90 quiet days.

8. Query plan regression diagnosis — Diffed pre- and post-upgrade execution plans. Correctly identified cardinality estimator changes causing join-order shifts on correlated columns (city + postal code). Wrongly recommended a database-wide legacy setting that would have pessimized hundreds of other queries. "It knew the mechanism. It did not know the blast radius."

9. Dialect drift verification after migration — Instead of asking AI to convert queries, asked it to list every semantic difference between source and target platforms that could affect the query. 180 test cases; 19 failed on first run. NULL ordering, collation, empty-string semantics, integer division, date arithmetic — every category produces wrong answers with no error message.

The article closes by noting that none of these clients bought a product. They got a workflow, a mandatory verification step, and a written note about failure modes. Their own people run it all now.