21’126 Interviews: How AI Can Take Employee Feedback Back to Its Roots

The Bottleneck Was Never the Data

It helps to remember why employee surveys exist at all. If you want to understand how your people are doing, the best method is to sit down with each of them and talk. Remarkably, an early study did exactly that. Between 1928 and 1930, the researchers behind the famous Hawthorne studies at Western Electric ran 21’126 open, conversational interviews.¹ They had started with closed yes-or-no questions. Those answers proved useless for locating real problems, so the researchers switched to simply listening, whereupon the interviews grew from thirty minutes to as much as two hours.

There was an obvious drawback of this approach: the programme took the research team years. Even today, most organisations could not afford that, and those that could would hardly want to, because feedback ages quickly. Two years on, many of the issues raised at the start of the study are likely to have changed or even disappeared. At any real scale, such in-depth interviews were and still are close to impossible.

So organisations found a solution: quantitative studies. This solved the effort problem, at the cost of losing a substantial amount of information. Nevertheless, this compromise became established and was refined over the following decades. The field developed KPIs, driver models and benchmarks, which became standard tools for managers. Each new technology made the machinery faster, from paper questionnaires to online surveys to real-time dashboards. But the underlying process never changed. Employees check boxes. Managers get dashboards.

The survey, in other words, was always a proxy for a conversation we could not afford to have. This proxy then quietly became the target and the score the deliverable.

Now let’s look at what AI is mostly used for today: faster survey creation, automated analysis, plain-language explanations of the results, summaries of free-text comments. All of this is genuinely useful, and we build some of it into Pulse Feedback ourselves, too. But notice what it amounts to: optimising the compromise. Watching the survey market, we see almost every vendor using AI to do what they already did, only faster. The process of obtaining data and calculating results remains largely untouched. Employees still check boxes. Managers still get dashboards, now with an AI-written summary on top.

We are convinced this will disappoint. Not because the technology falls short, but because the evidence has been telling us for over half a century that the analysis and the resulting report were never the bottleneck. Mann showed in 1957 that the breadth of involvement in working with results, not the quality of the report, drives change.² In fact, the survey feedback method was invented in the 1950s precisely because reports alone were changing nothing. Klein and colleagues found in 1971 that results are received better and seen as more useful when line managers discuss them in team meetings than when they arrive as a written report. The same study adds a warning that supports the point from the other side: badly run meetings, where the leader does not know the data or discourages discussion, produce worse reactions than a written report alone.³ And a 2021 systematic review by Huebner and Zacher confirms it with modern data: what makes surveys matter is the follow-up process, including action planning, and it is exactly this process that organisations most often neglect.⁴ Meanwhile Gallup finds that only 8% of employees strongly agree their employer acts on survey results.⁵

Decades of ever-better analysis did not move that number, because the analysis was never the constraint. It was the alibi. And AI has just stripped the alibi away: the report now writes itself in seconds, so when still nothing changes, only one explanation remains: the conversation never happened.

Which means AI holds two very different promises for employee feedback. The first is to run the old machinery faster, and that is where almost all the energy currently goes. The second is far more interesting: to bring back what the survey was invented to replace, the conversation itself. We will come back to that second promise. But the first one needs a closer look before anything else: what does more data, arriving faster, actually change?

Why More Data Will Not Change an Organisation’s Culture

A faster reporting engine, on its own, does not produce a better organisation. At best, it delivers you to the same dead end sooner. At worst, it speeds you off in the wrong direction.

That last part is not a figure of speech. In one of the most cited meta-analyses in the field, Kluger and DeNisi analysed over 600 effect sizes from feedback interventions and found that feedback improved performance on average, but in more than a third of cases it actually made things worse.⁶ What made the difference was where the feedback directed attention. Feedback that keeps people focused on their task helps. Feedback that makes people feel judged backfires. And a dashboard full of scores is remarkably good at making people feel judged: a score tells a team how it was rated, but not why or what could be done differently.

When feedback does lead to improvement, the driver is rarely the data itself. Research on 360-degree feedback shows that improvement depends on what happens after the results arrive: whether people can make sense of them, believe that change is possible, set goals and get support.⁷ Feedback, in other words, is not information transfer. It is a social process. A report that lands without a conversation is not a neutral event: it can raise anxiety, invite defensiveness, or simply be filed away and forgotten.

Culture changes when people sit together, interpret what they are hearing, disagree, prioritise and commit to a next step they genuinely believe in. No model can do that for them. AI makes the report arrive faster and look more polished. It does not make the conversation happen.

Abundant Data, Scarce Attention

If data is becoming abundant, what becomes scarce? Herbert Simon answered this back in 1971: "a wealth of information creates a poverty of attention".⁸ Survey results are no exception. The people who should work with them have no more hours than before, but ever more information competing for those hours. So feedback conversations get pushed aside by whatever is most urgent. Urgent, however, is rarely the same as important. Whether the hours AI frees up flow into the team conversation or quietly disappear into the merely urgent is a decision.

Our hypothesis is straightforward. The organisations that pull ahead over the next decade will not be the ones with the most sophisticated AI listening stack. They will be the ones whose teams meet three conditions: the skill to give and receive feedback honestly, the psychological safety to say what they really think, and the rhythm that makes feedback happen regularly and turns it into action.

The science supports each condition. On skill, Stone and Heen have argued that the real bottleneck in feedback is the receiver, and that receiving feedback well is a trainable skill.⁹

On safety, the evidence runs deepest. Amy Edmondson's work established psychological safety, the shared belief that a team is safe for interpersonal risk-taking, as one of the strongest known predictors of team learning and performance.¹⁰ A later meta-analysis across more than 22’000 employees confirmed the effect and added something hopeful: the strongest drivers of safety – supportive context, leader relations and work design – can all be shaped.¹¹ Google's Project Aristotle, a large internal study of what makes teams effective, reached the same conclusion from a different direction: how a team worked together mattered more than who was on it, and psychological safety sat at the foundation.¹²

And on rhythm: a recent interview study of organisations running pulse surveys found a consistent pattern. Where managers visibly discussed and acted on the results, people kept answering; where they did not, participation dried up.¹³ Stop acting, and people stop answering. Skill can be trained, safety can be built, rhythm can be kept. But all three are relational. None can be manufactured by a model.

That is the heart of it. An AI can write a flawless summary of low psychological safety on a team. But it cannot repair it: safety is rebuilt in how people respond the next time someone risks speaking up. AI can make things much faster, but human action is still what moves the needle.

Five Principles for Feedback in the AI Era

If we accept that premise, five practical principles follow for how to design and run employee feedback.

1. Use AI to remove friction, not to take people out of the loop

Comment clustering, theme detection, translation, drafting possible actions: these are tasks where AI saves real hours and often improves quality. Hand them over without hesitation. But the moment of sense-making, prioritisation and commitment belongs in a room with people in it.

Ethan Mollick makes a similar point about AI at work in general: treat it as a capable co-worker, and keep a human in the loop for the judgement calls.¹⁴ Applied to feedback, AI can support and even structure the team conversation. It cannot substitute for the commitment people make to each other in it. The goal is to spend less time producing the report and far more time acting on it.

2. Make ownership local and tailored to each team

Results have been moving outward as technology improved: from hand-made department reports that arrived weeks late to dashboards every team can open the moment a survey closes. AI can take this much further, from local access to local substance. Instead of the same report with different numbers, each team can get a picture that is genuinely its own: recommendations that know the team's history, what has changed since the last survey, and what has already been tried. Feedback drives change when the people it concerns work with it themselves.

This is the bottom-up principle we advocate, and AI now allows you to put more capability into every team's hands. Of course, not everything a team surfaces can be fixed by that team. Pay, workload and friction between departments need visible ownership at the top. Local ownership for team-level change, central responsibility for systemic themes. The two are complements, not rivals.

3. Invest in the human side at least as much as the AI side

It is tempting to pour budget and attention into the listening technology, because that is where the visible progress is. But feedback skill, the ability to hold a conversation that is both direct and kind, and psychological safety are the half of the loop that actually produces change. They compound slowly, and while training and facilitation can be bought, the accumulated trust cannot. An organisation that automates the easy half and neglects the hard half will end up with more data and less change.

4. Be transparent about what the AI does with people's words

There is a version of AI-powered listening that employees are likely to experience as surveillance. If people suspect that the model reading their comments serves monitoring rather than improvement, they will stop telling the truth, and the loop degrades before any conversation can happen. Honest answers are the one input no technology can produce, and no amount of analytical horsepower compensates for their loss.

Psychological safety now extends to the technology itself: be explicit about who sees what, what the AI does with it and what it will never be used for. And do not expect anonymity settings to carry this weight alone: someone who is genuinely afraid will not trust that anonymous answers cannot be traced back to them, whatever the tool promises. Honesty follows trust, not the absence of names, an argument we have made in more depth before.

5. Use AI to make listening richer, not just faster

The four principles above treat AI as an accelerator of the process we already have. The deeper opportunity is to ask a question that has been off the table for a century: what would we do if we could actually talk with everyone?

For the first time since those 21’126 interviews, we can come close. AI makes conversational listening possible at scale, a direction we are actively exploring at Pulse Feedback: chat-based interviews that ask the follow-up question a good interviewer would ask, going deeper to uncover the reasons behind an answer.

And the response scales? They have always done two jobs: collection and communication. Collection, because numbers were the only way we could gather and analyse feedback from thousands of employees. Communication, because numbers are wonderfully easy to read, compare and track, which is why KPIs and dashboards grew around them.

AI now takes over collection. Communication stays valuable, and tomorrow's figures can be derived from what people actually said instead of scales: how many raised a theme, whether it has grown since the last survey. Scales held their monopoly on how we ask because there was no alternative. That has just changed. What they keep is making results easy to read.

Where This Is Going

The temptation of the AI era is to believe that improving how people work together can be automated. The evidence, spanning seven decades of research, says it cannot. What AI changes is the cost of reaching the human layer of feedback. It makes that layer faster and cheaper to get to. It does not make it less important. If anything, as data turns into a commodity, the human conversation becomes the real differentiator.

The organisations that understand this will use AI for both of its promises: to clear away the friction of the old process, and to listen in ways no questionnaire ever allowed. And then they will spend the time they save where it counts: with their people, deciding together what to do next.

This is where our thinking stands today, and it is genuinely evolving. We would love to get your feedback on it.

References

1. Roethlisberger, F. J., & Dickson, W. J. (1939). Management and the worker. Harvard University Press.
2. Mann, F. C. (1957). Studying and creating change: A means to understanding social organization. In C. M. Arensberg et al. (Eds.), Research in industrial human relations: A critical appraisal (pp. 146–167). Harper & Brothers.
3. Klein, S. M., Kraut, A. I., & Wolfson, A. (1971). Employee reactions to attitude survey feedback: A study of the impact of structure and process. Administrative Science Quarterly, 16(4), 497–514.
4. Huebner, L.-A., & Zacher, H. (2021). Following up on employee surveys: A conceptual framework and systematic review. Frontiers in Psychology, 12, 801073.
5. Pendell, R. (2018). 10 ways to botch employee surveys. Gallup.
6. Kluger, A. N., & DeNisi, A. (1996). The effects of feedback interventions on performance: A historical review, a meta-analysis, and a preliminary feedback intervention theory. Psychological Bulletin, 119(2), 254–284.
7. Smither, J. W., London, M., & Reilly, R. R. (2005). Does performance improve following multisource feedback? A theoretical model, meta-analysis, and review of empirical findings. Personnel Psychology, 58(1), 33–66.
8. Simon, H. A. (1971). Designing organizations for an information-rich world. In M. Greenberger (Ed.), Computers, communications, and the public interest (pp. 37–72). Johns Hopkins Press.
9. Stone, D., & Heen, S. (2014). Thanks for the feedback: The science and art of receiving feedback well. Viking.
10. Edmondson, A. C. (1999). Psychological safety and learning behavior in work teams. Administrative Science Quarterly, 44(2), 350–383.
11. Frazier, M. L., Fainshmidt, S., Klinger, R. L., Pezeshkan, A., & Vracheva, V. (2017). Psychological safety: A meta-analytic review and extension. Personnel Psychology, 70(1), 113–165.
12. Duhigg, C. (2016, February 25). What Google learned from its quest to build the perfect team. The New York Times Magazine.
13. Berthelsen, H., & Muhonen, T. (2026). Taking the pulse: A qualitative study on pulse survey implementation. Frontiers in Organizational Psychology, 3, 1696769.
14. Mollick, E. (2024). Co-intelligence: Living and working with AI. Portfolio.