What happens when a company measures AI adoption by counting how often employees use it? Amazon is finding out.
Workers at Amazon are creating fake tasks just to hit internal AI usage targets, according to a Fast Company report. Managers are pushing teams to demonstrate they're using AI tools, and when employees can't find genuine applications, they're inventing busywork prompts to pad their numbers. Workers described asking AI to summarize documents that didn't need summarizing, or generating drafts with no intention of using them. The goal was not useful output. The goal was a logged interaction.
When the Metric Becomes the Job
This is a classic Goodhart's Law situation: when a measure becomes a target, it stops being a useful measure. Amazon presumably set AI usage quotas to push teams toward tools that could improve their work. Instead, they've created a direct incentive to fake productivity.
AI usage rates are easy to track and easy to game. Did someone open a tool? Did they run a query? These are countable things. Whether the output actually improved their work is much harder to quantify, so it often isn't measured at all. Somewhere down the management chain, a quota gets set based on what can be counted.
The problem is that forcing AI into workflows where it doesn't fit produces worse outcomes than doing the task the normal way. Not every job involves writing, researching, coding, or processing information in volume. For those workers, there may genuinely not be much for an AI assistant to do.
The Real Cost of Compliance Theater
This failure mode is different from employees resisting AI out of unfamiliarity. These workers aren't avoiding the tools - they're using them as a compliance checkbox. That's arguably harder to correct, because activity numbers look healthy while actual adoption stays hollow.
There's a secondary cost, too. When AI use becomes mandatory busywork, employees start associating these tools with overhead rather than help. Reversing that association later - when there are genuinely useful applications - is an uphill battle.
The irony here is specific: Amazon builds and sells AI infrastructure to other businesses, including Amazon Q Developer, its coding assistant for software teams. The gap between Amazon as an AI vendor and Amazon as an employer handling its own internal rollout is hard to ignore.
Genuine adoption doesn't require quotas. A support team summarizing ticket history before customer calls, a marketing team testing copy variations, a developer skipping boilerplate with a coding assistant - these uses spread on their own because they save real time. When organizations have to mandate AI usage and monitor it with targets, that's usually a sign the tools haven't found their natural fit yet. The solution is finding the right use cases, not pressuring employees to invent them.