← All case studies

Asking the Database Instead of Filing a Ticket

Client: the technology group of a national facilities operator (1,000+ locations). Details anonymized; specifics available on a call.

The problem

Two related pains, one root cause:

1. Ad-hoc data questions became tickets. "How many ETL conversions are scheduled this month?" meant finding the person who knew which system held the answer, waiting for their availability, and getting a number back maybe hours later.
2. Recurring metrics were collected by hand, from everywhere. The counts that fed regular reporting lived across Salesforce reports, Smartsheets, and assorted dashboards — so collecting them meant a person clicking through each system, reading numbers off the screen, and copying them into the report. Five tools, five logins, five chances to grab the wrong number — repeated every reporting cycle, forever.

What I built

A governed AI access layer over the company data warehouse:

• An AI assistant with read-only warehouse access through a controlled tool interface — every query schema-validated before it runs, every call written to an append-only audit log, hard timeouts on everything. Credentials never pass through the model. Plain-English questions come back as numbers with the SQL that produced them, so answers are checkable, not oracular.
• The hand-collection loop, retired. The dashboards, Salesforce reports, and spreadsheets people were reading numbers off of are all fed by the same underlying data — so the metrics now come from direct warehouse queries against those source-of-truth tables. One query replaces the tour through five front-ends, and it returns the same answer every time it's asked.

Results

• Data questions that took hours-to-days now take seconds, self-serve — and the people who used to field them got that time back.
• The click-through-and-copy collection routine is gone — no more reading numbers off five screens into one report.
• Validation caught a real classification error. Before cutover, the new queries were checked line-by-line against the human-produced reports they replaced. One metric disagreed — and the investigation showed the automated answer matched the source field the business actually meant, surfacing a distinction the manual process had been quietly approximating.