The AI Rubric Arrives Before the Definition of Good Does

AI fluency has stopped being an optional experiment and started appearing on the performance-review form: at the HR platform Gusto, managers place each worker into one of five AI archetypes during quarterly reviews, according to Business Insider ^[1]. That matters because the same reporting shows the first generation of these measurements already failing, with Amazon and Uber having tried leaderboards to track AI use and since pivoted away ^[4], and Duolingo walking back plans to evaluate AI use in reviews ^[5].

Gusto's ladder runs from "observers," who are not yet using AI on their own, through "integrators," who have made it part of the working day and see better outcomes consistently, to "amplifiers," who find new uses and teach colleagues ^[2]. Scott Helmes, Gusto's chief people officer, frames it as standardisation rather than reward and punishment: "We want to be able to have a coaching conversation," he told Business Insider, adding that the rubric has given the company a consistent way to talk about AI fluency ^[3]. He also concedes the harder part is still ahead. The next step, he says, is to look more directly at the impact workers make with AI [c3a].

That sequencing is the whole story. Grading usage is cheap and immediately available; grading consequence is neither. The reported cautionary case is Meta, which faces a lawsuit from laid-off workers alleging it used AI to rank productivity, including how often employees used AI tools, and then singled out low performers for cuts without accounting for medical and parental leave as reasons for lower usage ^[6]. Meta has denied the allegations ^[7].

The practice is nonetheless spreading. General Assembly surveyed 500 business leaders in the US and UK and found nearly half had incorporated AI into employee reviews, looking at reports of AI tool usage, performance improvements attributable to AI, and anecdotes about improved efficiency ^[8]. On those numbers, roughly 250 of the 500 leaders are already scoring something they cannot fully specify ^[9].

Richard Landers, a professor of industrial-organisational psychology at the University of Minnesota, told Business Insider that some companies have made the false equivalence of rewarding AI use even when it does not produce better work: "You can't really reward AI performance because we don't really know what that even is yet, and it also looks very different for different jobs" ^[10]. He also notes the transfer of risk downward, calling it "a lot of drive to push that responsibility onto frontline workers in a way that's not particularly fair" ^[11]. The context is that firms largely handed out subscriptions and logins and told people to work it out, which the reporting links to workslop, security exposure and sprawling workflows ^[12].

The managers quoted as further along are measuring output, not activity. Stefan Camilleri, vice president of engineering at Typeform, asks engineers how much faster they now do the work, but says speed alone is insufficient: "I'm expecting more ambitious output from fewer people" ^[13]. Shensi Ding, cofounder and CEO of Merge, reviews quality of use, citing an accounting employee who automated payment tracking and lifted team productivity, and reserves the highest marks for those who pass a working use case to colleagues ^[14].

Both of those are judgement calls dressed as metrics. That is defensible in a small company where the reviewer knows the work. It scales badly, and telemetry scales beautifully, which is why the usage leaderboard keeps coming back.