July 16, 2026
AI Handles Simple People Data Well, But Falls Short on Nuanced Employee Feedback
Managers are increasingly turning to generative AI tools to help oversee their direct reports. A 2025 survey of more than 1,300 US managers found that six in ten used such tools to inform decisions about promotions, raises, and even layoffs and terminations. But how reliable is generative AI when it comes to interpreting employee feedback and surfacing recommendations about people?
Beyond simple analysis, the answer is: not very.
That is the top-line finding from PYX-Voice, a new benchmark developed to assess how well AI can understand employee feedback in all its forms.
Billed as an industry first, the benchmark was created by PYX Labs and backed by employee experience provider Perceptyx. It evaluated each model’s response against criteria developed by industrial-organisational psychologists and organisational behaviour specialists, and the results should raise a flag for HR professionals and managers alike.
Without clear guardrails, policies, and specialised training on AI use in people decisions, the result could be misdirected investment at best, or legal exposure at worst.
Where AI Performs Well, And Where It Struggles
Seven frontier AI models, including those from OpenAI, Google, and Anthropic, were tested against 84 employee listening tasks.
On quantitative tasks with clear, verifiable answers, the models performed well, with scores ranging from 76% to 82%.
But on tasks requiring more complex interpretation, performance fell sharply. When asked to synthesise open-ended employee feedback into an accurate summary, scores dropped as low as 33%.
Asking Which Models Are Best Is The Wrong Question
In the more complex listening scenarios, the strongest performer was Gemini-3.5-flash at 66%, while the weakest was Claude Sonnet 4.6 at 33%.
Perhaps surprisingly, some lighter, faster models matched or outperformed heavier ones on the listening tasks. Claude Opus 4.8, built for complex reasoning, ranked among the lowest performers at 37% for summarising feedback.
But Joseph Freed, Chief Product Officer at Perceptyx and Head of PYX Labs, warns that simply switching models is not an effective response. “The takeaway for HR isn’t ‘use this model, not that one’ — it’s that even the best model passed only two-thirds of these tasks,” he says.
The real takeaway, Freed adds, is that interpretive work cannot yet be handed to any of the tested models without supervision. “Keep a human in the loop at the decision points,” he says.
Synthesis Is AI’s Weakest Capability
Synthesis was a consistent failure point across every model, with scores ranging from 14% to 57% — the lowest of any capability tested. Here, the benchmark asked for a specific, actionable recommendation based on scattered, ambiguous signals.
Synthesis failures typically occurred when models analysed more emotional, incomplete, or context-dependent feedback. The breakdown showed up in three ways: relaying back the input data instead of offering insight, giving generic and ungrounded advice such as “improve communication”, or producing a recommendation that was not feasibly actionable.
Models also struggled to interpret feedback on more nuanced topics such as ‘Change & Innovation’ and ‘Future/Vision’, where responses reflect an individual’s unique experience of the organisation and their personal reaction to change.
When AI Overclaims, The Risk Is Real
Complete fabrication of a number occurred only once across all model tests, while overclaiming occurred roughly twice per model across the 84 tasks.
While infrequent, these errors carry real risk. Freed warns that when they happen, they can push managers towards poor people decisions, leading to misdirected effort in recognition, benefits, tools, and interventions based on the reactions of “a vocal few”.
These fabrications and overclaims are especially dangerous, Freed notes, because they look plausible to anyone skim-reading the output. All figures must therefore be verified by a human before they inform any people or organisational decision.
Where Bias Creeps Into AI Output
The overemphasis on a ‘vocal few’ is a clear signal of bias in AI output. Martyn Redstone, Head of Responsible AI at Warden AI, notes that while overt ideology carries relatively low risk, the current body of LLM research points to compression bias as the greater concern.
“When a model is forced to turn ambiguous, emotionally mixed feedback into one clean takeaway, it tends to privilege the most frequent or most salient pattern, smooth disagreement into a fluent middle, and sometimes shift the emotional framing of the source,” Redstone says.
In these scenarios, minority or dissenting viewpoints can be under-represented, while the majority’s polarity gets amplified. When these tools are heavily relied upon, an inclusive approach to employee listening can easily fall away.
To mitigate the compression bias creep, Redstone says counter-views and source evidence must be “explicitly preserved” in the workflow, alongside human review.
The Manager Context: Overstretched And Under-Trained
HR and people leaders may be relatively alert to the risks of over-reliance on generative AI and act cautiously. The bigger concern is that team managers lean on it too heavily, often unaware of the implications.
This is happening against a backdrop where the ‘middle manager squeeze’ has become especially acute. Gallup data suggests organisations are increasingly flattening the managerial layer, leading to wider spans of control for those left behind.
There is also a longstanding problem: many people fall into management despite lacking human skills and emotional intelligence, because no upward technical track is available to them. AI training for managers is also not yet commonplace, with a 2026 study finding fewer than half of managers (48%) having received it.
Generative AI has arrived within, and fuelled, a reality in which managers are typically overstretched and under-trained. For many, adopting a readily available tool that reduces cognitive load and shaves hours off tasks is an understandable choice. But using the tool without understanding its limitations is where risk enters.
“The danger comes when managers mistake AI’s confidence for good judgement and overinflate their own AI literacy,” says Erica Farmer, AI and future skills specialist and co-founder of Quantum Rise Group.
The solution is not simply more AI training for managers. The skills required to assess and improve AI’s output — judgement, critical thinking, and coaching capability — also need attention, Farmer notes.
A quick win for managers is to work on effective prompting, alongside asking what the summary might have missed, whose voice isn’t being heard, and what assumptions the AI has made. “That human challenge process is where the real value lies,” Farmer adds.
What HR And People Leaders Should Do Next
The onus should not fall on managers alone. Employers need to ensure their people are properly equipped, and given the time, to absorb employee feedback and make people decisions — paying particular attention to managers who have taken on wider spans of control.
Training and development, for both AI use and critical thinking skills, plays a part. So do clear policies and guardrails. But if managers are still left without the time and headspace to assess employee feedback properly, over-reliance on AI will continue to creep in.
“AI can accelerate analysis, but it can’t replace the curiosity and contextual understanding that great people managers bring to employee feedback,” Farmer comcludes.
FAQs
Can AI accurately interpret employee feedback?
It depends on the task. AI models scored well (76–82%) on quantitative tasks with clear, verifiable answers, but performance dropped sharply on interpretive tasks. When asked to synthesise open-ended employee feedback into an accurate summary, scores fell as low as 33%.
What is AI’s biggest weakness when analysing employee feedback?
Synthesis — pulling scattered, ambiguous signals into a single, actionable recommendation — was the weakest capability across every model tested, with scores ranging from just 14% to 57%. Models struggled most with emotional, incomplete, or context-dependent feedback.
Which AI model performed best on employee listening tasks?
Gemini-3.5-flash was the strongest performer in the benchmark’s more complex listening scenarios.
Should HR teams let AI interpret employee feedback without human oversight?
No. Even the best-performing model in the benchmark passed only around two-thirds of the tasks tested, and rare but real instances of fabricated or overclaimed results were found across the models. A human should remain in the loop at decision points, particularly for interpretive judgements.
Your next read: AI Employee Feedback Analysis: Upgrade Your Employee Listening Strategy
