Research Brief: Staff rate things well in surveys, then voice concerns to AI

A consulting firm study found 20-41% of sessions paired a favourable survey rating with a concern voiced later to an AI interviewer. Staff stayed guarded.

Research · Luminary Research Brief · 5 October 2026 · 4 min read

A positive survey score does not mean staff have nothing to say. In a field study at a global management consulting firm, Tamme and colleagues found that in 20-41% of sessions, depending on how favourability was defined, an employee gave a favourable rating and then voiced a substantive concern on the same theme later in an AI voice interview. For operators, the point is that your engagement numbers may be flattering, and that even a more open channel will not produce unguarded candour on its own.

What the researchers did

The setting was a global management consulting firm. Its listening process pairs a pre-survey with an adaptive AI voice interview (a conversational agent that adjusts its follow-up questions to what the person says) on the same themes within a single session. That design lets the researchers compare what someone rates with what they then say aloud.

The quantitative part covers 44 first-session interviews, giving 132 matched theme observations, each pairing a survey rating with the interview on that theme. The qualitative part used the Gioia method, a structured way of coding interview material into themes and building a model from them. It drew on 158 protective quotes from 65 eligible sessions. A "protective quote" is one where the employee is managing the risk of what they are saying.

What they found

The headline figure is the 20-41% range of sessions in which a favourable rating co-occurred with a substantive concern voiced later. The range is wide because it depends on the threshold used to count a rating as favourable. The finding holds across thresholds, but its size does not.

The qualitative results matter as much. According to the abstract, disclosure rarely arrived unguarded. Employees softened concerns, deflected accountability, and bounded how far they went. This protective work tracked the perceived legitimacy of the listening structure, meaning how far people believed the process was a proper, safe route for raising issues.

From this the authors build a grounded model of bounded disclosure and derive four propositions for research on voice, channel and listening. Their summary is that silence can persist inside expression: people talk, but still hold back.

What this means for your business

Treat a good score as a prompt, not an all-clear. If your pulse survey says a team is content, the study suggests some of that team may still have concerns they have not put in the number. Follow positive ratings with an open conversation, and read the softened language in free-text comments and interviews as information.

Work on legitimacy before technology. The protective behaviour tracked how legitimate the listening structure seemed. Staff will want to know who sees the output, whether it reaches their manager, and what happens next. Say this plainly before you ask anyone to speak, and then act visibly on something. UK evidence on technology rollouts points the same way: CIPD data from 2020 shows only 35% of employees or their representatives were consulted on introducing or implementing new technology, and that perceptions of job quality were 20% positive without consultation against 70% positive when consulted. That figure concerns job quality, not listening tools, but it suggests that how a tool arrives shapes how it is received.

Do not assume AI is the fix, or the problem. Adoption is rising: the ONS reports that approximately 35% of UK businesses with 10 or more employees use at least one AI technology, up from around 12% in late 2023. Plenty of firms will be tempted to add an AI interviewer to an existing survey. The abstract does not say that AI interviews produce more honest answers than a human one would. It says that people disclose in a guarded way even in this channel. Any pilot should therefore be judged on whether it surfaces concerns your survey missed, not on the novelty of the format.

Run a simple check of your own. If you gather ratings and comments on the same themes, compare them. Count how many people scored a theme favourably yet raised a worry about it. The paper's method is not something a small firm can replicate, but the underlying question is easy to ask of your own data.

Reduce the hierarchical risk. The study concerns disclosure where speaking up carries risk. Options include routing responses to someone outside the reporting line, reporting only in aggregate with a minimum group size, and separating listening from performance review. These are our suggestions, not tested interventions from the paper.

Limits worth knowing

This is a single firm, and a management consulting firm at that, where hierarchy, career progression and professional reputation may shape what people feel safe saying. A 20-person manufacturer or a retailer with shift workers may behave differently. The quantitative sample is small: 44 first-session interviews. The range of 20-41% reflects real uncertainty about what counts as a favourable rating, so quote it as a range and not as a single rate.

The study is also descriptive. As summarised in the abstract, it does not compare the AI channel with a human interviewer or a survey alone, so it cannot say that AI changes disclosure for better or worse. The model of bounded disclosure and its four propositions are theory-building from interview material, and they need testing elsewhere. The paper appears to be a preprint, so it has not necessarily been through peer review. Only the abstract was available for this brief, so details of method and results beyond it are not covered.

Our view: the finding deserves moderate weight. It is a useful warning against reading favourable scores as candour, and it fits what many managers already suspect. It is not yet evidence for any particular tool or channel.

Source: Thilo Tamme, Michael Saatkamp, Alma Bonte, Daniel Weiss, Anton Hantel, Andrej Levin, “Positive Ratings, Hidden Concerns: Employee Voice Disclosure in AI-Mediated Organizational Listening”, arXiv. This paper is a preprint and has not yet been peer reviewed.