The one thing AI feedback tools can't account for

AI has made it easier than ever for managers to write polished, structured performance feedback. A growing roster of platforms now pull data from Slack, project trackers, and goal systems to draft review summaries, flag coaching moments, and even detect bias in language. The technology is efficient, scalable, and increasingly popular. What it doesn’t account for, according to two researchers who study workplace behavior and emotion, is how the employee on the receiving end of that feedback is actually wired to receive it.

“The message is just words on a page,” said Anthony Belluccia, head of science at The Predictive Index in Boston. “What determines whether feedback actually works is what happens inside the person receiving it.”

Belluccia, a behavioral scientist and industrial-organizational psychologist, argues that the current generation of AI performance review tools shares a fundamental blind spot. They’re trained on what employees do and say, not on the behavioral traits that shape how individuals process criticism, respond to conflict, or absorb coaching. A well-crafted sentence, he says, can land as motivating to one employee and feel like an attack to another. AI, as it currently exists, can’t tell the difference.

That blind spot has a measurable cost. Research by Crucial Learning, based on a study of 1,025 people, found that every crucial conversation that’s avoided or mishandled costs an organization approximately $7,500 and more than seven lost working days. Multiply that across a company’s annual review cycle, and the cost of getting feedback wrong extends well beyond hurt feelings. It shows up on the balance sheet.

Why AI feedback tools miss the emotional receiver

Alicia Grandey, a professor of psychology in the industrial-organizational program at Penn State University in University Park, Pennsylvania, and a leading researcher on emotional labor, says the emotional mechanics of feedback are more complex than most managers appreciate.

“Critical or negative feedback can be personally threatening, especially if one cares about their work role,” she said.

When feedback backfires, Grandey explains, it typically unfolds in one of two ways. An employee who blames the supervisor to protect their own ego may become angry, withdraw, or retaliate. One who internalizes the criticism without seeing a way forward may feel shame and disengage entirely. Neither response is uncommon, and neither advances development.

Not every employee comes into a feedback conversation from the same starting point, either. Individual differences in mindset matter enormously. People with a fixed orientation, who believe their performance reflects stable traits they can’t change, tend to hear criticism as a verdict on who they are rather than a signal for growth. Those with a more malleable view of their own abilities are more likely to process the same feedback as useful information.

Workplace culture shapes this too, Grandey notes. In a culture where mistakes are penalized or one bad review ends in termination, employees are more likely to be fearful and reactive when feedback arrives, regardless of how it’s delivered.

The behavioral data current platforms don’t capture

Belluccia’s concern is less with AI-generated feedback itself than with what current tools leave out. In his view, most platforms are optimizing for the message while the person receiving it goes unconsidered. Most analyze communication patterns, goal progress, and historical performance data. What they don’t capture is the behavioral profile that predicts how a specific person will respond when that feedback lands.

That gap matters more than ever. Gallup data from early 2025 found that just 31% of U.S. employees feel engaged at work, a 10-year low. Belluccia connects that directly to people not feeling seen. In an era of AI-generated content, he says, feedback that feels generic or impersonal is making the problem worse.

Belluccia points to a 2025 randomized controlled study of over 20,000 peer reviews at the International Conference on Learning Representations (ICLR) as an analogy for what happens when AI enters any feedback process. Reviewers who received AI assistance wrote longer, more comprehensive reviews. But the AI couldn’t replicate the human judgment that comes from actually knowing the subject matter, the person, or the stakes involved. In a workplace context, he argues, employees can feel that difference.

Research consistently shows a disconnect between why managers think employees leave and why they actually do. Managers tend to assume it’s about pay, while employees more often cite relational reasons: not feeling recognized or seen. Feedback tools built on productivity data, Belluccia says, do nothing to address it.

The hidden cost to the manager

Grandey raises a dimension of the problem that goes beyond the employee experience. When managers use AI to generate feedback, they skip more than the writing process. They skip the emotional work that makes the conversation possible in the first place.

“They have not done the work of reviewing and generating the feedback themselves, which means they may lack depth of knowledge about this employee’s performance and concerns, and have lost an opportunity to wrestle with the difficulty of this feedback and how they feel about it,” she said.

That loss of practice matters. Feedback conversations are a form of emotional labor, and like any skill, they atrophy without use. Grandey’s research suggests the relational cost persists even when managers are transparent about using AI. Employees still perceive them with less trust.

Grandey notes that AI tools tend to get adopted precisely when managers are already stretched thin, supervising more people than they can meaningfully engage with. That dynamic, she argues, is self-reinforcing.

“AI tools then mean that managers can be expected to supervise more employees, making AI even more needed and their emotional skills and relationships even harder to develop,” Grandey said.

HR leaders grappling with a broken culture of feedback already face organizations where managers lack the confidence and skills to have direct conversations, and where employees have stopped trusting that feedback will lead anywhere. Layering AI on top of that dynamic without addressing its root causes may accelerate the problem rather than solve it.

Where AI can genuinely help

Neither Belluccia nor Grandey dismisses AI in performance management entirely. Grandey sees real value in using AI tools to help with tone and structure.

“Using AI or ChatGPT tools to help with the tone of a difficult email could be a helpful starting place,” she said. “This is especially true for people who may struggle with emotional skills.”

The distinction she draws is between AI as a starting point and AI as a replacement. A manager who uses a tool to brainstorm language and then personalizes it based on their knowledge of the employee is doing something meaningfully different from one who sends an AI-generated review with minimal input.

Belluccia frames the opportunity similarly. Better behavioral data, he argues, could help managers prepare for difficult conversations more effectively, not by removing emotional responsibility, but by reducing the uncertainty that makes those conversations so hard to initiate in the first place.

As organizations navigate the risks of widespread AI adoption, the real question is whether organizations are being honest about what these tools can and can’t do. Writing a better performance review is a solvable problem. Understanding how the person on the other side of that conversation is wired to receive it is considerably harder.