AI Insights: anomaly stream & ask-your-data
AI Insights watches your KPIs against industry benchmarks and your own alert rules, and answers plain-language questions with real numbers — it never invents figures.
AI Insights has two jobs. The anomaly stream scans the live value of every computable KPI and flags anything that breaks a threshold — an industry benchmark (e.g. picking accuracy ≥ 99.5%, dock-to-stock < 24h) or an alert rule you created yourself (e.g. DSO > 45 days). Each flagged event comes with a one-sentence AI reading. The ask box lets you type a question in plain words — "What was our DSO last week?" — and get the real number with its source.
Reading the anomaly stream#
- Each card shows the KPI name, its current value, the threshold or benchmark it broke, how far it deviated (e.g. +10%), and when it was detected — plus the sample size and the calculation definition so you can judge how solid the number is.
- The AI take is one sentence, based only on the numbers shown. It never invents new figures — if the AI service is down, you still see the numbers themselves.
- No anomalies? The page honestly says "No anomalies right now" instead of making up events. KPIs with no data yet are skipped, not treated as anomalies.
- Rule trigger history below the stream lists past firings of your own rules, newest first.
Creating an alert rule#
In the Alert rules section, pick a KPI, choose a condition (greater than / less than / at least / at most / equals) and type a threshold — e.g. DSO, greater than, 45. The next scan will flag it when the live value breaks your rule, and the firing is recorded in the trigger history. Pause a rule any time instead of deleting it if the threshold is seasonal.
Asking questions#
- Type in Chinese or English (the page language is used for the answer). Time words work: "last week" → 7 days, "last month" → 30 days, "this quarter" → 90 days.
- Every answer states its source: which KPI engine value, the sample size, the calculation definition, and when it was computed.
- If the question is vague ("how much did we ship?"), the system asks a clarifying question and suggests the closest KPIs — click one to ask directly. If a KPI has no data yet, it says so honestly instead of guessing.