Case Studies
How Problems Actually Get Solved
Three representative case studies drawn from documented day-to-day practice. Specific amounts and identities are withheld by professional discretion; the method is the point.
The Recurring Cash Shortfall
Business Problem
A branch repeatedly closed its day with small cash shortfalls. Individually negligible, collectively a pattern — and patterns in cash are never noise.
Investigation
Compared POS session totals against physical counts across shifts, mapped which cashiers and hours the variances clustered in, and reviewed CCTV for the matching time windows.
Root Cause
Not theft: a drawer-handover habit. Two cashiers were switching mid-shift without a formal count, so errors migrated between sessions and became untraceable by design.
Solution
Recommended a counted handover — a two-minute checklist at every drawer change — helped the branch put it into practice, and briefed the team on why the step protects the cashiers themselves.
AI Contribution
An LLM-drafted investigation summary turned a day of scattered notes into a clear, structured incident report for management — reviewed line-by-line against source data before filing.
Business Impact
The shortfall pattern stopped at the source, cashier accountability became provable instead of presumed, and the handover checklist entered the branch's daily routine.
Technologies
- Foodics POS
- CCTV review
- Excel variance log
Lessons Learned
Most 'fraud signals' are process gaps. Fix the procedure first; suspicion without evidence corrodes a team faster than a missing banknote.
The Transaction That Didn't Match the Camera
Business Problem
A refund transaction appeared in the POS log at a moment when the counter camera showed no customer present — the kind of mismatch that demands an answer, not an assumption.
Investigation
Pulled the full POS session log around the timestamp, cross-referenced camera footage of the counter and the till, and interviewed the shift staff with the facts — not accusations — on the table.
Root Cause
A legitimate correction of an earlier mis-keyed sale, processed late because the cashier waited for a quiet moment — poor timing discipline, not misconduct.
Solution
Cleared the employee formally in the incident file, then recommended a rule the branch adopted: corrections are processed immediately with a supervisor note, never batched for quiet moments.
AI Contribution
Used an LLM to structure the incident narrative and timeline from raw observation notes, making the exoneration as well-documented as an accusation would have been.
Business Impact
Preserved trust in the monitoring system on both sides: management saw rigor, staff saw fairness. The timing rule closed a real audit-trail gap.
Technologies
- CCTV platform
- Foodics POS logs
- Incident file
Lessons Learned
An investigation that can prove innocence is the same machinery that can prove guilt — and it is worth running properly for exactly that reason.
From Field Notes to Management Report in Minutes
Business Problem
Daily collection rounds produced valuable observations — cashier issues, branch conditions, anomalies — that arrived at management late and unevenly, buried in message threads.
Investigation
Tracked a week of own reporting effort: where the hours went, what management actually read, and which details drove decisions versus which were ignored.
Root Cause
No fixed structure. Every day's report was composed from scratch, so writing consumed control time and inconsistent formats hid the signal.
Solution
Built a reusable prompt template: raw figures and field notes go in, a fixed-structure daily report comes out — same sections, same order, every day — verified against source before sending.
AI Contribution
The core of the case: the LLM handles composition and structure; the human supplies facts and verifies output. Drafting time dropped from the better part of an hour to minutes.
Business Impact
Management now reads a consistent report daily and reacts faster to exceptions; reclaimed time goes into actual branch follow-up rather than formatting.
Technologies
- ChatGPT
- Claude
- Prompt template library
- Excel
Lessons Learned
AI adoption sticks when it removes a hated task and keeps the human visibly in charge of the facts. Start with the report everyone dreads writing.
These are representative composites of recurring, documented work situations — deliberately anonymized, with no confidential figures disclosed.