A system nobody feeds stops being a system.
Every automated area leans on a few steps only a person can do, like hours entered or an approval given by name. When those steps stop, nothing breaks loudly. The system keeps running and keeps producing confident output from stale information, which is worse than producing nothing at all.
It is always one of three.
An input does not fail in interesting ways. Something stops being entered, an approval stops happening or a source stops being kept current.
Feeding
Something did not get entered.
Approving
A gate was passed without a named person, or work is backing up behind one.
Refreshing
A connection expired, a source document went stale, a list stopped being maintained.
One or two steps per area. That is all.
It watches for the absence of these steps, not anybody's activity. Every alert names the record that went quiet, never the person.
Write the outcome and the reason onto the pursuit
Weekly
The system stops learning. Scoring freezes at the day it was built
Record what was actually said on the call
After each client conversation
The update gets assembled from an empty file
Enter hours and log scope changes as they happen
Daily
The scope check reports green on bad numbers, which is worse than reporting nothing
Close actuals against the estimate
At project close
The early warning has nothing to warn about
Put finished work where it can be found again
At project close
Everything downstream degrades quietly. This is the most fragile one
Approve at the review point, by name
Per release
The approval point becomes a story you tell clients
Set goals at kickoff and keep the data connections live
At kickoff, then monthly
The report is written confidently about the wrong period
Answer the short pulse, and read the answers
Monthly
You have a survey nobody reads
Three states. Plain words, not colors.
Current
The input arrived when it was expected.
Late
It missed its window.
Gone quiet
Nothing for two full cycles. This is the one that matters, because it means somebody stopped rather than slipped.
These are the only three states. No dashboard to watch, no numbers to read.
An approval you cannot verify is a story you tell clients.
The site already says most of the agency should stop at rung four, and that a person approves anything a client sees. That promise is only real if someone can check it.
An approval point nobody can verify is an approval point nobody is keeping. So every approval writes a row. Who approved, what they approved, when, against which version. An approval with no record is not an approval.
An approval point you cannot verify is a story you tell clients. The Lookout is how it stays true after I hand the work over.
A sheet, a scheduled check and a message.
One row per input. The last time it arrived, when the next one is due, who owns it.
It runs on its own and looks for the rows that are overdue. Nobody has to remember to run it.
One note in the channel your agency already uses. What is current, what is late, what has gone quiet.
When everything is fine it says so in one line, so nobody assumes it quietly died.
The owner of the input hears first. Their manager after two misses. The agency owner once a month, as a summary.
No new login. Nothing to learn. The plainness is the point, not a weakness to hide. A check you have to remember to run is not a check.
Not first. After two or three areas are running.
Build the Lookout once two or three areas are actually running, when there are real inputs worth watching. Before that you would be guessing at which steps matter, and guessing is how you end up watching the wrong thing. It is the layer that keeps everything else honest, so it comes after there is something to keep honest.
