A silent failure gets signed off

Asked for 100 records. Received 80. Reported 96% coverage.

That was an AI pipeline I built and run every day.
Nothing failed. No errors. No alerts.
Just a clean report with a number that looked exactly right.

It was not the only time.

• A watcher reported zero findings every 5 minutes for two days, because what it missed had arrived before it started looking.
• A ranked list put records scoring 95 at the top. They had expired weeks earlier.
• A background job reported "done" while every page it read was an error.

Every one of them came with a green tick ✅.

Each got the same kind of fix. Count what came back against what was asked for.
Judge the output by what it says, not by its status.
Give every record a date that proves it is still alive.

None of that is new. Finance calls it control totals and reconciliation. Across every migration and close I have led, one rule held: no number ships until a second source agrees with it.
The AI failure that should keep us awake is not the wrong answer. A wrong answer gets challenged.
A silent failure gets signed off.

Cleaning the data before AI goes live is the part everyone agrees on. It is also the easy part. Data drifts the day after the project closes, and an automated process keeps running on it, confidently, with nobody looking.

So before anything goes live, two questions:
1. Who owns this data, by name, closely enough to notice when it goes quiet?
2. What checks the output against something other than itself?

AI can help clean your data. It cannot own it.

If your AI went quiet tomorrow, who on your team would notice first?