The Amazon Story

Amazon automated a significant portion of its human workforce with AI. Then the AI failed at doing those jobs. Then Amazon blamed the human employees who remained for not catching the AI's errors.

That framing , human oversight failure as a personnel problem rather than a system design problem , is the tell.

11,183 people upvoted coverage of this because engineers recognized the pattern immediately. The jobs that were automated were highly deterministic. Inventory checks. Rule-based approvals. Routing decisions with clear criteria. The AI was not bad at these. The AI failed at the exception cases. And exception cases are where human judgment exists in those workflows for a reason.

The Question

Before automating any task with AI, ask this: is the output deterministic given the input?

If yes , if the same input should always produce the same output, and you can write down the rules , the task is automatable and AI will handle it reliably. Routing invoices. Flagging threshold violations. Sorting tickets by category. These work.

If no , if the task requires reading context, weighing tradeoffs, or applying judgment that cannot be written down in rules , the task is not ready for autonomous AI. It may be ready for AI assistance. It is not ready for AI replacement.

The distinction sounds simple. Almost every failed AI deployment has ignored it.

Why Companies Keep Getting This Wrong

The pressure to automate comes from finance, not from engineering. The person who decides what gets automated is often not the person who understands what the task actually requires.

A workflow that looks deterministic from the outside , "the agent reviews support tickets" , may contain hundreds of judgment calls invisible from the spreadsheet. What counts as a valid complaint? When does a refund exception apply? When should this escalate versus close?

Amazon's AI failed at exception handling. The humans left were blamed for not catching exceptions that the AI was deployed specifically to handle. The system design was wrong from the start.

The Practical Test

Write down the rules for the task. All of them. Every exception you can think of.

If you can write a complete rulebook , one that covers every realistic input , the task is ready for AI. If you keep adding "well, it depends" to the rulebook, the task is not ready yet.

The "it depends" answers are exactly the judgments that will fail quietly, at scale, and come back to someone's desk as a personnel problem.