Alerts instead of reports: AI tells you when a number leaves the range
A report requires someone to remember to look. An alert arrives on its own. What has to be defined for a warning to be useful instead of noise, and why most alerts get switched off in two weeks.
A report is a pull system: it waits for someone to remember to open it. An alert is a push system: it arrives even when nobody was watching. Most expensive operational problems happen exactly in the window when nobody was watching — nights, weekends, holiday weeks.
The loss does not come from the bad number. It comes from the days between the number turning bad and someone noticing.
Why almost every alert gets switched off
Anyone who has tried knows the pattern: switch the alert on, it fires too often, the team creates an email filter, and in two weeks nobody reads it. The mistake is rarely technical. It is definitional.
| Mistake | What happens | How to fix it |
|---|---|---|
| Alerting on normal | Daily noise trains the team to ignore | Only alert outside the expected range, not on every change |
| A fixed threshold | Saturday fires every Saturday | Compare with the same weekday, not with an overall average |
| An alert with no owner | Everyone sees it, nobody acts | Every alert has one responsible person |
| An alert with no possible action | Informs and frustrates | Only alert on what can be changed today |
The four alerts that pay for themselves
- 01Stalled source. No sales ingested in the last few hours. It is the only alert that warns about an invisible problem: missing data does not show up in a totals dashboard.
- 02Drop outside the range. Today sales below what is expected for that weekday. Not "below average".
- 03Cost per sale climbing. Flat spend with falling conversion is the signature of a campaign that broke.
- 04Delivery stopping. An ad or email integration that started refusing. Credential failure is silent by nature.
The first and the fourth prevent the most loss and are configured the least, because they are not about performance — they are about the pipeline working. An expired token can cost weeks of badly attributed conversions without any chart looking ugly.
What changes with AI in the loop
A classic alert says a number crossed a threshold. A connected assistant can go one step further: alongside the warning, bring the breakdown that explains it. "Down 22%, concentrated in one channel, one product, starting yesterday afternoon" is a warning that already starts the investigation.
A good alert is not the one that warns fastest. It is the one that arrives with the first question already answered.
How to start without creating noise
- Start with two alerts, not ten. Stalled source and out-of-range drop cover most of the risk.
- Set the range from history, not from ambition. A target and an expected range are different things.
- Write what to do inside the alert itself. A warning with no next step is a notification.
- Review what fired this month. An alert that fired ten times and never produced an action should be switched off.