My scheduled AI job keeps failing and the AI blamed the wrong cause

What happened

We had a sync scheduled for 8 AM every day, running against a Mac mini that stays on all the time. It kept failing.

We asked the AI what was wrong. Its first answer: the Mac must be going to sleep overnight.

We checked the power logs (on a Mac, pmset -g log shows sleep and wake events). The machine never slept. The real problem was the app's connection dropping overnight while it sat idle. By 8 AM the connection was gone and the job failed on its first try.

Why the AI did it

"Computer was asleep" is the most common reason an overnight job fails, so it's the most likely-sounding answer. The AI gave the likely answer, not the checked one. It hadn't looked at a single log.

It also pointed the blame away from itself. The fault was in the AI app's own connection, and that was the last place it looked.

Agents connected to Salesforce or any other system do the same thing. A failed sync gets blamed on "permissions" or "the API was down" because those are common, even when the logs say something else.

The fix, step by step

1. Logs first, diagnosis second. Tell the AI: "Don't diagnose until you've read the logs. Quote the lines your answer is based on."
2. Rule out the obvious with evidence. For "it was asleep," check the power log. For "the API was down," check the status page for that hour.
3. Ask what else it could be. "Give me three possible causes and the log that would prove or rule out each one."
4. Make the job retry. One failed attempt at 8:00 shouldn't mean a failed day. A couple of retries a few minutes apart handles a dropped connection that reconnects.
5. Say plainly where the fault is. If the failure is on the AI tool's side, the report should say that, not blame the machine.

How to check it worked

• Plant the wrong cause. Ask the AI "is my machine sleeping?" when you know it isn't. It should check the log and say no.
• Break the connection on purpose before a scheduled run. The job should retry and then succeed, or fail with a clear message naming the real cause.