Ask the AI to find anomalies (what is out of order). Return JSON {"data":"...","findAnomalies":"...","rule":"..."}: data — context; findAnomalies - what types of oddities to look for (outliers - amounts much larger than usual, duplicates, sharp spikes, suspiciously round amounts); rule - a rule or threshold for what is considered an anomaly (for example, “the amount is 3 times higher than the median” or “two identical payments on the same day”).
An anomaly is a deviation from the expected, but you have to define what is “expected” yourself. Without a rule, AI will call any large number an anomaly and flood you with false alarms. Mechanics: it compares each line with a typical one, and you decide what is typical. Levers: give a threshold (in times or percentage of the average/median) and name the types of oddities - outliers, duplicates, sharp jumps, omissions. Insider: ask to compare with the MEDIAN, not the average. One huge sum inflates the average and hides the rest of the outliers underneath it; The median is the “middle” payment; one giant does not move it, so outliers against its background are clearly visible. Second technique: separately ask to look for duplicates - match supplier + amount + date. This is a frequent leak (accidentally paid twice), which trends and amounts do not catch. A typical mistake: asking “find errors” - the AI does not know what an error is for you, and will either remain silent or make things up. And remember: AI only SHOWS an anomaly, you still have to check with your eyes.
Unlock access to submit solutions for instant AI review.