190 checks every night: what I no longer check myself
Every night, 4 diagnostics run about 190 checks and only anomalies come up. What to check first in your company, and the traps to avoid.
Thibault Pons de Vincent
Founder of Oliverlist and VivaSync
At Oliverlist, more than 100 people work every day: employees, contractors and bookers. Close to 7,800 mailboxes send 225,000 emails per working day. At that scale, you cannot check everything yourself. You check what you have time to look at, and you miss the rest.
If you run a company while part of your team works remotely, in other places and time zones, you know the pattern. The question "where are we on this?" goes out in a group chat, answers trickle in, and real problems surface late.
Today, four diagnostics run every night. Together they perform about 190 checks. In the morning I do not read 190 results: I read the anomalies, when there are any. Here is how it works, what I would check first in a mid-sized company, and the traps to avoid.
One rule: only anomalies come up
A check is a closed question put to your data, with an expected answer. Has every delivered order been invoiced? Has a client gone past their payment terms? Has a figure moved more than usual? If the answer is the expected one, nothing happens. If not, the check raises an anomaly.
Every night, my four diagnostics run through about 190 checks of this kind. When everything is normal, I receive nothing. That is the core principle, and the third layer of my method: a check that speaks up when all is well soon stops being read.
On the steering side, 1,334 reports have been published so far. The 84 agents behind my work as CEO are monitored too: they run 200 to 230 times per working day, and 99.4% of those runs succeed. Treat an agent that fails like any other anomaly.
Messages to the team
An anomaly that stays in a report fixes nothing. It has to reach the right person, in a form they can act on. Between the agents and the team, that means about 130 messages per working day.
A good anomaly message has four parts:
- who needs to act, by name;
- what was found, with the figure and the reference;
- why it matters, in one sentence;
- where to act: a direct link to the record, the invoice or the file.
Send it where your team already works, not to an inbox nobody opens. The person replies, fixes or explains. The next night, the check runs again: if the anomaly is gone, the topic is closed; if not, it comes back.
I no longer ask everyone "where are we on this?": the agent asks the right person, with the facts. Along the way, task tracking moved out of Notion and now lives in the same tool as the checks.
What to check first in a mid-sized company
Do not start with 190 checks. Start with about ten, where a mistake costs money or a client. The same families come up again and again.
- Money that should come in. Has every delivery been invoiced? Which invoices are more than 7 days overdue? At Oliverlist, an invoice is generated automatically for every meeting booked: exactly the kind of rule a nightly check verifies with no effort.
- Commitments made to clients. Quotes with no reply, late orders, meetings with no follow-up: what was promised and not done.
- Flows that stop. An import that did not run, a broken sync, a mailbox that no longer receives anything. Silence is a signal: a figure that drops to zero rarely triggers an alert, though it should.
- Deviations from normal. A volume that falls or jumps compared with the same weekdays in previous weeks. Compare a Monday with other Mondays. Weekends differ across the region and working hours change during Ramadan: your baseline has to know that.
- Data quality. Duplicates, incomplete records, inconsistent amounts. Every other check depends on it.
Each time a problem slips through, add the check that would have caught it. That is how the list grows without becoming unreadable: every line has a reason to exist.
The traps
False alerts. Every false alert costs someone time and wears down their trust in the next message. After a few, nobody reads anymore. So measure them and bring them down. On the sorting of prospect replies at Oliverlist, that work cut false alerts by 77% while detecting 19% more hot leads. Balance matters: I would rather have one alert too many than miss a hot reply, but one needless alert every day ends up getting all the others ignored.
Checks nobody reads. A report sent to everyone is read by no one. Every anomaly needs a named recipient and an expected action. Clean up regularly: a check that fires every day no longer flags an anomaly, it describes a state. Fix the cause or change the threshold. And a check that has not fired in six months deserves a question: is it still checking anything?
The checker nobody checks. Whatever does the checking makes mistakes too, so audit it. On prospect reply sorting, an audit of 7,612 replies found about 4 real errors per day. That is few, but it is not zero, and reviewing is how those errors become new rules.
Where to start
List the questions you ask your team every week to find out whether things are on track. Each one is a check waiting to be written. For each, note the expected answer and who should act when it does not come: that is your first check. The first ten are enough to begin.
To find out which ones to automate first in your company, book 30 minutes: you walk me through your week, and I tell you what can run without you. Or start with the online test: 10 questions, your level among four and the 3 tasks to hand to AI first, shown on screen straight away.