This autumn, when your year-end review lands, there is a decent chance a machine wrote the first draft. In a Harvard Business Review piece on AI and reviews, the examples are the big names: Citi has a tool called Performance Assist that pulls data from across the company to draft evaluations, JPMorgan built an internal assistant, its LLM Suite, that helps write year-end reviews, and Boston Consulting Group's in-house version reportedly cut review-writing time by 40 percent. A manager still signs the thing. The machine did most of the typing.
Here is the gap worth worrying about. In a Betterworks survey of 2,387 HR leaders, managers, and employees, 90 percent of HR leaders said AI has already changed what a high performer looks like, and 88 percent said it changed how performance should be judged. Then only 42 percent said their company had actually updated its goals or review criteria to match. The tool got faster. The rulebook did not move. (Betterworks sells performance software, so read the numbers with that in mind, but the gap is their own finding.)
What that means for you is simple enough. You are being scored, more efficiently than ever, against a standard most of the people running the process think is out of date. The writing got automated. The definition of good did not.
An AI does not know you. It knows your traces: the tickets you closed, the documents you touched, the messages it is allowed to read. Chrysanthos Dellarocas, the Boston University professor behind that Harvard Business Review piece, argues that most companies are just using AI to produce more polished versions of the same flawed reviews. The polish is the danger. A fluent, confident paragraph reads as fair whether or not it is. And what the model cannot see, it cannot credit. The quiet save, the colleague you unblocked, the bad plan you talked someone out of, none of that leaves a clean data trail.
There is a real split under all this. Executives are six times more likely than employees to believe reviews have kept pace with AI-driven work. In the same survey, 92 percent of executives said they are comfortable using AI, against just 51 percent of employees, and fewer than one in six workers said they even understand their own company's AI vision. The people who bought the tool are comfortable using it. The people it scores are mostly in the dark.
Dellarocas's own fix is to point AI at evidence instead of adjectives. "A single consequential episode," he writes, "where an employee challenged a flawed assumption, redirected a failing project, or aligned stakeholders around a difficult trade-off, often reveals more about capability than a page of evaluative language." That is a change a company has to make, and most have not made it. You cannot rewrite the rubric yourself. But you can decide what evidence exists to be read.
So here is the script. Do not wait for review season to learn what the machine can see. Keep your own record across the year: three or four consequential episodes, written the way Dellarocas describes, with the decision, the stakes, and how it turned out. Put them where your manager and the tool both have to look, in the project doc, the summary, the update, not a private file nobody opens. And ask, plainly, what data feeds your evaluation. If the answer is vague, that is your answer.
The company automated the part that was easy to automate. It left you the part that was always yours: making sure the work that mattered is written down somewhere a machine can find it. Which, come to think of it, was decent career advice long before anyone taught a computer to type.
