I built a mini-sprint and time keeping system that tracks stages of development for bundles of issues with timestamps and explanations so the AI can look at the data and make better estimates because this annoyed me so much.

It often goes exactly the way you mentioned, 3 days estimate for 3 hours in reality. But it often goes the other way too!

It goes the other way usually with some little screw up that turns massive or a real mistake the AI made or something that the AI went sideways on because of the jagged frontier.

I have a command call /bog for that to record "bog time" when we get stuck in mud fighting something stupid like getting the CI push green for two hours after a half hour of coding, or a half day feature turns into a 3 day slog.

That's the biggest problem of planning today. You never know where the slot machine will take you today and you can't project plan it like a normal project because it can either one shot everything twenty times in a row or completely screw everything up on the first slot machine pull.