Targeted alerts: only hear about what leaves the ordinary

What is a targeted alert?

A targeted alert is a message that only goes out when an indicator leaves its usual range, that reaches the person who can act on it, and that carries enough context to decide without opening five tools. The test fits in one sentence: an alert is only sent if it calls for an action.

The opposite is the dashboard you open “when you think of it”, or the inbox that receives forty notifications a day. In both cases the problem is found too late, or drowns in noise. This article explains how to choose what deserves an alert, how to set thresholds, what an AI adds beyond a simple rule, and when you are better off installing nothing.

Why too many alerts are worse than none

An alert that fires constantly ends up ignored, and on the day it is right nobody reads it. The effect is documented mostly in IT and security, where the cost has been measured for years. It carries over to a small business unchanged.

• A 2018 survey by the vendor Imperva of 179 IT professionals reports that 30% had completely ignored some categories of alerts, 4% had switched notifications off, and 56% had ignored an alert because of past false positives.
• A guide on alert fatigue from the vendor Vectra AI cites a 2026 Microsoft and Omdia study putting false positives at 46% of security alerts, and a 2025 SANS survey in which 73% of respondents name false positives as their first challenge.

These figures come from software vendors, in a cybersecurity context, not from small businesses. Read them as an order of magnitude for the mechanism: the more often an alert is wrong, the less it is trusted. The monitoring vendor New Relic gives the same advice: do not alert on everything, but on a small set of indicators tied to what really matters for the business.

What deserves an alert, and what does not

The sorting rule is simple: an alert is justified if one named person can do something concrete within hours, and if doing nothing costs more than being interrupted. Everything else belongs in the morning summary, not in a notification.

These examples are illustrative: your thresholds depend on your business. The test for each row stays the same: who acts, with what context, and what happens if they do not?

Setting the threshold: fixed rule or deviation from habit

There are two ways to decide that a number is “out of the ordinary”, and they complement each other.

A fixed threshold is a value you decide: stock under 50 units, an invoice more than 15 days late, a deal margin under 20%. It is readable, easy to explain and easy to correct. It suits limits that do not move.

A deviation from habit compares a number with its own history: spending twice the six-month average, or a client who usually pays in 30 days and now passes 50. Monitoring tools call this anomaly detection: Apigee’s documentation, for instance, explains that the platform determines conditions with statistical methods and updates them continuously. The notification still has to be configured separately, otherwise the anomaly just sits on a screen nobody watches.

In practice a good setup mixes both: fixed thresholds for hard limits, deviation from habit for what varies by season, client or product. Either way, the threshold is decided with you, not for you.

The six-step flow

The studio’s automation catalogue describes the principle. Here is what it looks like once installed.

1. Pick three to five indicators. Not twenty. Start from what has already cost you money or a client.
2. Connect your existing tools. CRM, accounting, stock or ERP, project management. You keep your tools, we read what you have.
3. Define thresholds with you. Fixed or relative to history, signed off by the person who will have to act.
4. Detect. The AI spots the deal, invoice or indicator that crosses the threshold and checks it is not a case already handled.
5. Attach context. The alert arrives with the likely cause and the useful history: since when, by how much, what happened last time.
6. Notify the right person. In Slack or Teams, to the person who decides and to nobody else. They choose to act or ignore, and that choice is used to adjust the rule.

What an AI adds beyond a simple rule

A rule like “if an invoice is more than fifteen days late, send an email” exists in most software. An AI adds three useful things.

• Readable context. Instead of an invoice number, you read “this client usually pays in 32 days, we are at 61, and a dispute email arrived a week ago”.
• Grouping. Ten alerts with one shared cause become a single message.
• Triage of unclear cases. When a number leaves its range for no obvious reason, the AI flags it as “to check” rather than deciding.

It does not bring judgement. The studio’s rule holds for all its flows: the AI prepares, a person validates. An alert closes nothing, sends nothing to a client and changes no accounting entry.

A worked example, under assumptions

These figures are calculation assumptions, not a result measured at a client. Take a twelve-person company that checks by hand, once a week, its late invoices, its unanswered deals and its spending gaps.

• The check takes about an hour and a half a week, close to 75 hours a year.
• Between two checks a problem can stay invisible for up to seven days. A late invoice found the following Friday has lost five days of follow-up.
• With targeted alerts the routine check disappears, and the time goes to the three or four cases that really leave the ordinary each week.

The time saved matters less than the delay saved: you deal with a problem the day it appears. Redo this calculation with your own numbers, or go through the express diagnostic (French only) to do it in five questions.

Who should receive which alert

Routing matters as much as the threshold. A late invoice goes to whoever chases payments, a stalled deal to the salesperson who owns it, a spending gap to the person who approves purchases. Management only receives what no one else can settle. When an alert has no obvious recipient, that is usually a sign the indicator is a report, not an alert. Write the recipient next to each threshold before anything is connected, and review the list after the first month: people change roles, and an alert sent to the wrong person is as useless as no alert.

When this is not the right choice

• You have almost nothing to watch. With a handful of invoices a month, a weekly glance is enough. Automation does not earn back its cost.
• Your data is not kept up to date. An alert on a half-filled CRM or on accounts entered three weeks late produces false alerts and missed ones. Fix data entry first.
• Nobody knows who should act. An alert without an owner is one more message. Decide who receives what before choosing a threshold.
• The risk is vital or critical. For anything touching people’s safety or a production stoppage, do not delegate monitoring to a flow like this: dedicated systems are needed.

Pitfalls to avoid

• Watching everything at launch. Start with three indicators, observe for two weeks, then add.
• Never correcting a false alert. Every false alert should be used to adjust the rule. Without that feedback, trust erodes within weeks.
• Sending to everyone. An alert sent to ten people is handled by nobody.
• Confusing alert and report. Anything that needs no immediate action belongs in the daily summary, not in an interruption.
• Forgetting to measure. Each month, count the alerts received, those that led to action and those that were false. If fewer than half call for action, tighten the thresholds.

Where to start

Write down the last three times a problem was found too late: a forgotten invoice, a client left without a reply, a spending gap spotted at the close. For each, note what should have triggered a message, who should have received it, and what it should have said. Those are your first three alerts. To put numbers on them and see what connects to your tools, the express diagnostic is a starting point, followed by a free 30-minute call.