HR teams are asking AI what to flag in a manager's draft review

A finished review has a second audience: the next manager, HR, sometimes a lawyer. Writing help stops being cosmetic at that point. Bora Unlu, co-founder and CEO of the performance management platform Teamflect, wrote that managers mostly bring him a review they have already written and ask what is wrong with it, with questions about tone and exposure, and about how the document will read to anyone who sees it later ^[1][7]. He wrote that when the stakes turn legal or emotional, people want a second set of eyes ^[8].

The two survey figures in his account point in different directions. A 2025 Gartner survey of nearly 3,500 employees found 87% thought algorithms could give fairer feedback than their managers ^[2]. On those numbers, roughly 3,045 respondents said an algorithm could be fairer and about 455 did not ^[14]. Gallup's workforce data put writing and editing at 51%, the most common use of AI at work ^[3]. The biggest buyers went the other way: Citi and JPMorgan both introduced AI to help managers write evaluations, not to grade anyone, according to Unlu ^[4].

The use he calls unexpected sits between the reviewer and the reviewee. Skip-level managers and HR ask AI review assistants what should be flagged and sent back to the reviewer, and Unlu wrote that the concerns can be about a review someone else writes ^[9]. A manager's draft now passes in front of a tool that the layer above them is querying.

Reviews have always been edited, HR has always sent them back, and a tool that tightens a sentence changes nothing about who decides. Unlu's caution is about the writing itself. He wrote that asked to make a passage more professional, the tool smooths the edges, and that this can subtly damage a review ^[11]. "Framing facts and feedback in a professional manner is a core competency for managers, and offloading that skill to AI can lead to it atrophying," he wrote ^[10].

The case for calling AI in performance management low risk is that a human still owns the judgment call, which is close to what Unlu says happens in most cases ^[12]. His account of how employees, managers and administrators use these tools comes from building Teamflect, and he does not say how many teams or cycles he is describing ^[13][15]. The Gartner and Gallup numbers are both cited secondhand inside the post ^[2][3].

So the policy work this quarter is authorship and retention: whether AI-assisted wording is disclosed, and whether the pre-edit draft is kept. The wording in the file is what a promotion or a termination gets argued from. Employees are running the same play from their side of the form, and Unlu wrote that some of them ask for help rating their own performance, a few days before the deadline ^[6].