All cases
Food-service chain · under NDAFour countries · 16 locations with P&L

Owner Intelligence · QA

Owner IntelligenceSelf-auditAI Systems

01

Challenge

The owner's loop had been running for three months: morning digest, per-location anomalies, an advisor. Wording errors were caught after the fact from client screenshots: the group's second brand slipped into the first brand's digest, the advisor issued verdicts instead of facts, and one delivery channel «collapsed» by 93% when what collapsed was order classification in the POS. Quality control had to live inside the loop, not in the client chat.

02

Solution

01
Self-audit across four zones: the location registry against actual POS sales, source freshness and sum reconciliation, recurring advice, and the quality of messages actually sent
02
The model reads sent digests and checks them against nine brand rules: brands do not argue each other's case, location formats are not compared head-on, one deviation means one number, no verdicts
03
A Monday auto-run that stays silent when there is nothing to fix; a one-command manual run on any complaint
04
Source rule: before concluding «a channel fell», reconcile with the POS report for one location; query wrappers now fold the order source themselves
05
Silence rule: no measurable shift, no line in the digest; harsh judgements of people are never shown

03

Deep dive

What the first live audit found

Four critical violations, five for attention, two for information. The digest compared a food-court location with the network average, where most locations are restaurants. One paragraph carried two deviations of the same metric on different bases. The advisor wrote «verdict: failure» where the rules allow only a fact and a question.

All closed within a day in prompts and code: the location format is printed next to every location, the deviation moves into the headline and is not repeated, the record baseline is a four-week median, verdicts are banned at system level. Older messages kept surfacing in the audit for a week as history, not regression.

The second run two days later found a false positive in the check itself: it counted profitability rows as locations. The auditor needs auditing too.

The false channel collapse

The report showed minus 93% on a delivery channel across every location in two cities. The owner checked the POS for one location: the channel had produced normal revenue for the month. The cause sat in the POS: since June that aggregator's orders arrive as a different order type with a source field, and the old query could not see them.

The real picture: minus 40% in July, then stable. POS query wrappers now fetch the source field and fold the channel themselves, and a one-location POS check became mandatory before any «channel fell» conclusion.

04

Results

0
Quality rules
0
Self-audit zones
0 / нед
Auto-run
0
Verdicts in digests